{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Gabriele Iocco

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Filtro anti hater per social network

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "import matplotlib.pyplot as plt\n", "import gdown\n", "import seaborn as sns\n", "import pickle\n", "from tensorflow.keras.preprocessing.text import Tokenizer\n", "from tensorflow.keras.preprocessing.sequence import pad_sequences\n", "import re\n", "from keras.utils import pad_sequences\n", "import nltk\n", "from nltk.corpus import stopwords\n", "from nltk.stem import WordNetLemmatizer\n", "import numpy as np\n", "from imblearn.over_sampling import SMOTE\n", "from sklearn.model_selection import train_test_split\n", "\n", "from sklearn.neighbors import NearestNeighbors\n", "import random\n", "from collections import Counter\n", "from tensorflow.keras.models import Sequential\n", "from tensorflow.keras.layers import Input, Embedding, Bidirectional, LSTM, Dense\n", "from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau, ModelCheckpoint\n", "from sklearn.metrics import accuracy_score, f1_score, precision_score, confusion_matrix, classification_report\n", "from sklearn.metrics import roc_auc_score, roc_curve, auc\n", "from sklearn.metrics import hamming_loss\n", "from sklearn.metrics import precision_recall_curve\n", "\n", "from tensorflow.keras.models import load_model, Model\n", "from tensorflow.keras.layers import Dense, Dropout, Input\n", "import tensorflow as tf\n", "\n", "import seaborn as sns\n", "import time\n", "from sklearn.metrics import multilabel_confusion_matrix\n", "from tensorflow.keras.backend import clear_session\n", "from colorama import Fore, Style, init\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import warnings\n", "warnings.filterwarnings(\"ignore\")\n", "import os\n", "os.environ['TF_CPP_MIN_LOG_LEVEL']='3'\n", "import tensorflow as tf\n", "tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from colorama import Fore, Back, Style\n", "\n", "def print_colored(text, color=\"white\", bg_color=None, end=\"\\n\"):\n", " # Dizionario dei colori del testo\n", " color_dict = {\n", " 'red': Fore.RED,\n", " 'blue': Fore.BLUE,\n", " 'white': '\\033[97m', # Bianco puro (ANSI)\n", " 'black': Fore.BLACK\n", " }\n", "\n", " # Dizionario dei colori dello sfondo\n", " bg_color_dict = {\n", " 'black': Back.BLACK,\n", " 'blue': Back.BLUE,\n", " 'white': Back.WHITE\n", " }\n", "\n", " color_code = color_dict.get(color.lower(), '\\033[97m') \n", " bg_color_code = bg_color_dict.get(bg_color.lower(), '') if bg_color else ''\n", " \n", " print(f\"{color_code}{bg_color_code}{text}{Style.RESET_ALL}\", end=end)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analisi del dataset

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "BASE_URL=\"/home/gap/Scrivania/Filtro_anti_hater/\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "df_hater = pd.read_csv(BASE_URL + \"Filter_Toxic_Comments_dataset.csv\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Numero di righe:\n159571\nNumero di colonne:\n8\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Numero di righe:\", \"blue\") \n", "print(df_hater.shape[0])\n", "print_colored(\"Numero di colonne:\", \"blue\") \n", "print(df_hater.shape[1])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[15]:
\n
\n
comment_text     159571\ntoxic            159571\nsevere_toxic     159571\nobscene          159571\nthreat           159571\ninsult           159571\nidentity_hate    159571\nsum_injurious    159571\ndtype: int64
\n
\n
\n
\n
" ] } } ], "source": [ "df_hater.count()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
<class 'pandas.core.frame.DataFrame'>\nRangeIndex: 159571 entries, 0 to 159570\nData columns (total 8 columns):\n #   Column         Non-Null Count   Dtype \n---  ------         --------------   ----- \n 0   comment_text   159571 non-null  object\n 1   toxic          159571 non-null  int64 \n 2   severe_toxic   159571 non-null  int64 \n 3   obscene        159571 non-null  int64 \n 4   threat         159571 non-null  int64 \n 5   insult         159571 non-null  int64 \n 6   identity_hate  159571 non-null  int64 \n 7   sum_injurious  159571 non-null  int64 \ndtypes: int64(7), object(1)\nmemory usage: 9.7+ MB\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_hater.info()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1276568\n1276568\n\nValori mancanti\n0\n
\n
\n
\n
\n
" ] } } ], "source": [ "total_values = df_hater.size\n", "print(total_values)\n", "\n", "non_missing_values = df_hater.count().sum()\n", "print(non_missing_values)\n", "\n", "missing_values = total_values - non_missing_values\n", "print_colored(\"\\nValori mancanti\", \"blue\")\n", "print(missing_values)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Valori mancanti per colonna:\n\n        comment_text  toxic  severe_toxic  obscene  threat  insult  \\\n0              False  False         False    False   False   False   \n1              False  False         False    False   False   False   \n2              False  False         False    False   False   False   \n3              False  False         False    False   False   False   \n4              False  False         False    False   False   False   \n...              ...    ...           ...      ...     ...     ...   \n159566         False  False         False    False   False   False   \n159567         False  False         False    False   False   False   \n159568         False  False         False    False   False   False   \n159569         False  False         False    False   False   False   \n159570         False  False         False    False   False   False   \n\n        identity_hate  sum_injurious  \n0               False          False  \n1               False          False  \n2               False          False  \n3               False          False  \n4               False          False  \n...               ...            ...  \n159566          False          False  \n159567          False          False  \n159568          False          False  \n159569          False          False  \n159570          False          False  \n\n[159571 rows x 8 columns]\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Valori mancanti per colonna:\\n\", \"blue\")\n", "print(df_hater.isna())\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[19]:
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
comment_texttoxicsevere_toxicobscenethreatinsultidentity_hatesum_injurious
0Explanation\\nWhy the edits made under my usern...0000000
1D'aww! He matches this background colour I'm s...0000000
2Hey man, I'm really not trying to edit war. It...0000000
3\"\\nMore\\nI can't make any real suggestions on ...0000000
4You, sir, are my hero. Any chance you remember...0000000
\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_hater.head()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
Out[20]:
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
toxicsevere_toxicobscenethreatinsultidentity_hatesum_injurious
count159571.000000159571.000000159571.000000159571.000000159571.000000159571.000000159571.000000
mean0.0958440.0099960.0529480.0029960.0493640.0088050.219952
std0.2943790.0994770.2239310.0546500.2166270.0934200.748260
min0.0000000.0000000.0000000.0000000.0000000.0000000.000000
25%0.0000000.0000000.0000000.0000000.0000000.0000000.000000
50%0.0000000.0000000.0000000.0000000.0000000.0000000.000000
75%0.0000000.0000000.0000000.0000000.0000000.0000000.000000
max1.0000001.0000001.0000001.0000001.0000001.0000006.000000
\n
\n
\n
\n
\n
" ] } } ], "source": [ "df_hater.describe()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Distribuzione delle classi

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
I commenti appartengono alle seguenti categorie:\n\n['toxic', 'severe_toxic', 'obscene', 'threat', 'insult']\n
\n
\n
\n
\n
" ] } } ], "source": [ "categories = df_hater.columns[1:-2].tolist()\n", "\n", "print_colored(f\"I commenti appartengono alle seguenti categorie:\\n\", \"blue\")\n", "print(categories)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Distribuzione dei commenti per categoria:\n\n               count\ntoxic          15294\nsevere_toxic    1595\nobscene         8449\nthreat           478\ninsult          7877\nidentity_hate   1405\n
\n
\n
\n
\n
" ] } } ], "source": [ "class_distribution = df_hater[categories].sum().to_frame(name='count')\n", "\n", "print_colored(\"Distribuzione dei commenti per categoria:\\n\", \"blue\")\n", "print(class_distribution)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "class_distribution.plot(kind='bar', figsize=(10, 6))\n", "plt.ylabel(\"Numero di Commenti\", fontsize=16)\n", "plt.xlabel(\"Categorie\", fontsize=16)\n", "plt.xticks(fontsize=12, color='#b81414')\n", "plt.title(\"Distribuzione dei commenti\", fontsize=18)\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Distribuzione dei commenti in percentuale:\n\ntoxic            0.958445\nsevere_toxic     0.099956\nobscene          0.529482\nthreat           0.029955\ninsult           0.493636\nidentity_hate    0.088049\ndtype: float64\n
\n
\n
\n
\n
" ] } } ], "source": [ "class_proportion = (class_distribution / df_hater.shape[0])*10\n", "\n", "print_colored(\"Distribuzione dei commenti in percentuale:\\n\", \"blue\")\n", "print(class_proportion)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "class_distribution.plot(kind='bar', figsize=(10, 6))\n", "plt.title(\"Percentuale di distribuzione dei commenti\", fontsize=18)\n", "plt.xlabel(\"Categorie\", fontsize=16)\n", "plt.ylabel(\"Percentuale\", fontsize=16)\n", "plt.xticks(fontsize=12, color='#b81414')\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Numero di commenti non classificati, quindi non tossici:\n143346\nNumero di commenti con una categoria attiva:\n6360\nNumero di commenti con due categorie attive:\n3480\nNumero di commenti con tre categorie attive:\n4209\nNumero di commenti con quattro categorie attive:\n1760\nNumero di commenti con cinque categorie attive:\n385\nNumero di commenti con sei categorie attive:\n31\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo il numero di etichette attive per ogni commento\n", "df_hater['label_count'] = df_hater[categories].sum(axis=1)\n", "\n", "# Distribuzione del numero di etichette attive\n", "multi_label_distribution = df_hater['label_count'].value_counts().sort_index()\n", "\n", "plt.figure(figsize=(10, 6))\n", "multi_label_distribution.plot(kind='bar', )\n", "plt.title(\"Commenti per numero di categorie attive\", fontsize=18)\n", "plt.xlabel(\"Categorie\", fontsize=16)\n", "plt.ylabel(\"Numero di Commenti\", fontsize=16)\n", "plt.xticks(fontsize=12, color='#b81414')\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n", "\n", "print_colored(\"Numero di commenti non classificati, quindi non tossici:\", \"blue\")\n", "print(multi_label_distribution.at[0])\n", "\n", "print_colored(\"Numero di commenti con una categoria attiva:\", \"blue\")\n", "print(multi_label_distribution.at[1])\n", "\n", "print_colored(\"Numero di commenti con due categorie attive:\", \"blue\")\n", "print(multi_label_distribution.at[2])\n", "\n", "print_colored(\"Numero di commenti con tre categorie attive:\", \"blue\")\n", "print(multi_label_distribution.at[3])\n", "\n", "print_colored(\"Numero di commenti con quattro categorie attive:\", \"blue\")\n", "print(multi_label_distribution.at[4])\n", "\n", "print_colored(\"Numero di commenti con cinque categorie attive:\", \"blue\")\n", "print(multi_label_distribution.at[5])\n", "\n", "print_colored(\"Numero di commenti con sei categorie attive:\", \"blue\")\n", "print(multi_label_distribution.at[6])\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Preprocessing dei dati

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "nltk.download('stopwords')\n", "nltk.download('wordnet')\n", "nltk.download('omw-1.4')\n", "\n", "# Inizializzo stopwords e lemmatizer\n", "stop_words = set(stopwords.words('english')) # Stopword per la lingua inglese\n", " # Rimuove parole comuni italiane per ridurre il rumore nei dati\n", "lemmatizer = WordNetLemmatizer() # Riduce le parole alla loro forma base per \n", " # diminuire la dimensionalità del vocabolario.\n", "\n", "def clean_text(text):\n", " # Rimuovo caratteri speciali e punteggiatura\n", " text = re.sub(r'[^a-zA-Z\\s]', '', text)\n", " # Converto tutte le lettere in minuscolo\n", " text = text.lower()\n", "\n", " tokens = text.split()\n", " tokens = [lemmatizer.lemmatize(word) for word in tokens if word not in stop_words]\n", " # Ricostruisco il testo pulito\n", " return ' '.join(tokens)\n", "\n", "def preprocess_data(df_hater, categories):\n", " # Pulizia del testo\n", " df_hater['comment_text'] = df_hater['comment_text'].apply(clean_text)\n", " \n", " # Tokenizzazione\n", " tokenizer = Tokenizer() # Non limitiamo il vocabolario\n", " tokenizer.fit_on_texts(df_hater['comment_text']) # Adatta il tokenizer ai testi\n", " sequences = tokenizer.texts_to_sequences(df_hater['comment_text']) # Trasforma i testi in sequenze numeriche\n", " \n", " # Labels\n", " labels = df_hater[categories].values\n", " \n", " return sequences, labels, tokenizer\n", "\n", "# Preprocessing\n", "sequences, labels, tokenizer = preprocess_data(df_hater, categories)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Salvo il tokenizer\n", "with open('tokenizer.pickle', 'wb') as handle:\n", " pickle.dump(tokenizer, handle, protocol=pickle.HIGHEST_PROTOCOL)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dimensione del vocabolario:\n212562\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Verifica del vocabolario\n", "word_index = tokenizer.word_index # Dizionario parola -> indice\n", "vocab_size = len(word_index)\n", "\n", "print_colored(\"Dimensione del vocabolario:\", \"blue\")\n", "print(vocab_size)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Lunghezza minima:\n0\nLunghezza massima:\n1250\nLunghezza media delle sequenze:\n34.67\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Lunghezza delle sequenze (senza padding)\n", "sequence_lengths = [len(seq) for seq in tokenizer.texts_to_sequences(df_hater['comment_text'])]\n", "\n", "# Statistiche sulla lunghezza\n", "min_length = min(sequence_lengths)\n", "max_length = max(sequence_lengths)\n", "avg_length = sum(sequence_lengths) / len(sequence_lengths)\n", "\n", "\n", "print_colored(\"Lunghezza minima:\", \"blue\")\n", "print(min_length)\n", "print_colored(\"Lunghezza massima:\", \"blue\")\n", "print(max_length)\n", "print_colored(\"Lunghezza media delle sequenze:\", \"blue\")\n", "print(f\"{avg_length:.2f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "plt.figure(figsize=(10, 6))\n", "plt.hist(sequence_lengths, bins=10, edgecolor='black')\n", "plt.title(\"Distribuzione della Lunghezza delle Sequenze\", fontsize=18)\n", "plt.xlabel(\"Lunghezza delle Sequenze\", fontsize=16, color='black')\n", "plt.ylabel(\"Frequenza\", fontsize=16, color='black')\n", "plt.xticks(fontsize=12, color='#b81414')\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Numero di commenti con lunghezza compresa tra 0 e 20 token:\n86418\nNumero di commenti con lunghezza compresa tra 21 e 50 token:\n44601\nNumero di commenti con lunghezza compresa tra 51 e 100 token:\n18214\nNumero di commenti con lunghezza compresa tra 101 e 200 token:\n7137\nNumero di commenti con lunghezza compresa tra 201 e 300 token:\n1620\nNumero di commenti con lunghezza compresa tra 301 e 400 token:\n1021\nNumero di commenti con lunghezza compresa tra 401 e 600 token:\n500\nNumero di commenti con lunghezza compresa tra 601 e 1250 token:\n60\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Tokenizzazione dei commenti (eseguita una sola volta)\n", "tokenized_sequences = tokenizer.texts_to_sequences(df_hater['comment_text']) # Tokenizzazione dei commenti\n", "sequence_lengths = [len(seq) for seq in tokenized_sequences] # Calcola le lunghezze delle sequenze\n", "\n", "# Intervalli di lunghezza\n", "outlier_thresholds = [\n", " (0, 20),\n", " (21, 50),\n", " (51, 100),\n", " (101, 200),\n", " (201, 300),\n", " (301, 400),\n", " (401, 600),\n", " (601, 1250)\n", "]\n", "\n", "# Commenti per ciascun intervallo\n", "counts = []\n", "for low, high in outlier_thresholds:\n", " count = sum(1 for length in sequence_lengths if low <= length <= high)\n", " counts.append((low, high, count))\n", "\n", "for low, high, count in counts:\n", " print_colored(f\"Numero di commenti con lunghezza compresa tra {low} e {high} token:\", \"blue\")\n", " print(count)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "interval_labels = [f\"{low}-{high}\" for low, high, _ in counts] # Etichette degli intervalli\n", "frequencies = [count for _, _, count in counts]\n", "\n", "plt.figure(figsize=(10, 6))\n", "plt.bar(interval_labels, frequencies, color=\"#1f77b4\", edgecolor='black')\n", "plt.title(\"Distribuzione dei Token\", fontsize=18)\n", "plt.xlabel(\"Lunghezza delle Sequenze (Token)\", fontsize=16, color='black')\n", "plt.ylabel(\"Numero di Commenti\", fontsize=16, color='black')\n", "plt.xticks(fontsize=12, color='#b81414', rotation=45)\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

I commenti estremamente lunghi sono poco frequenti nel dataset rispetto alle sequenze corte. Per questo motivo, li considererò come outlier

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
[[  457    57    60 ...     0     0     0]\n [85242   940  1175 ...     0     0     0]\n [  318   313    17 ...     0     0     0]\n [  122    22   238 ...     0     0     0]\n [ 1442  2369   765 ...     0     0     0]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Applico il padding\n", "max_len = 400 # Lunghezza massima scelta\n", "padded_sequences = pad_sequences(tokenized_sequences, maxlen=max_len, padding='post', truncating='post')\n", "\n", "print(padded_sequences[:5])\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Distribuzione per toxic:\ntoxic\n0    144277\n1     15294\nName: count, dtype: int64\n\nDistribuzione per severe_toxic:\nsevere_toxic\n0    157976\n1      1595\nName: count, dtype: int64\n\nDistribuzione per obscene:\nobscene\n0    151122\n1      8449\nName: count, dtype: int64\n\nDistribuzione per threat:\nthreat\n0    159093\n1       478\nName: count, dtype: int64\n\nDistribuzione per insult:\ninsult\n0    151694\n1      7877\nName: count, dtype: int64\n\nDistribuzione per identity_hate:\nidentity_hate\n0    158166\n1      1405\nName: count, dtype: int64\n\n
\n
\n
\n
\n
" ] } } ], "source": [ "for category in categories:\n", " counts = df_hater[category].value_counts()\n", " print_colored(f\"Distribuzione per {category}:\", \"blue\")\n", " print(counts)\n", " print()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "# 0 e 1 per ogni feature\n", "feature_counts = {feature: df_hater[feature].value_counts() for feature in categories}\n", "\n", "features = list(feature_counts.keys())\n", "counts_0 = [feature_counts[feature].get(0, 0) for feature in features]\n", "counts_1 = [feature_counts[feature].get(1, 0) for feature in features]\n", "\n", "# Intervallo per le barre\n", "x = np.arange(len(features))\n", "width = 0.4\n", "\n", "plt.figure(figsize=(12, 6))\n", "plt.bar(x - width/2, counts_0, width=width, label='Classe 0', color='#1f77b4')\n", "plt.bar(x + width/2, counts_1, width=width, label='Classe 1', color='orange', alpha=0.7)\n", "\n", "plt.title(\"Distribuzione di 0 e 1 per ogni feature\", fontsize=18)\n", "plt.xlabel(\"Categorie\", fontsize=16, color=\"black\")\n", "plt.ylabel(\"Numero di commenti\", fontsize=16, color=\"black\")\n", "plt.xticks(x, features, fontsize=14, rotation=45, ha='right', color=\"#b81414\")\n", "plt.yticks(fontsize=14, color=\"#b81414\")\n", "plt.legend(loc='center left', bbox_to_anchor=(1, 0.5), fontsize=12)\n", "plt.grid(axis='y', linestyle='--', alpha=0.5, color=\"#1f77b4\")\n", "\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analizzando i dati, si nota che le categorie sono poco sbilanciate tra loro, ma all'interno di ciascuna categoria le classi mostrano un forte squilibrio. Addestrare il modello con questi dati potrebbe portare a una distorsione verso la classe di maggioranza, penalizzando la performance sulla classe minoritaria. Per mitigare questo problema, applicherò il sovracampionamento. Invece di duplicare semplicemente i dati esistenti, utilizzerò tecniche come SMOTE per generare nuovi campioni sintetici che rappresentino meglio la distribuzione della classe minoritaria, riducendo il rischio di overfitting e bias.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Divisione del dataset

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

70% per il training set, 15% per il set di validazione e 15% per il set di test applicando stratify per garantire che la suddivisione dei dati preservi la proporzione delle classi ed evitare che una classe sia sottorappresentata o assente. Per rappresentare la distribuzione reale.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dimensione Training Set:\n111699\nDimensione Validation Set:\n23936\nDimensione Test Set:\n23936\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Estraggo i valori dalle colonne\n", "labels_df = df_hater[categories].values\n", "\n", "# Creo una etichetta che rappresenta il numero totale di classi positive per ogni campione\n", "# Questo è necessario per stratificare i dati preservando la distribuzione multi-classe\n", "stratify_labels_df = labels_df.sum(axis=1)\n", "\n", "# Verifico che il numero di campioni nei dati corrisponda al numero di etichette\n", "assert len(padded_sequences) == len(labels_df), \n", "\n", "X_train_mlsmote, X_temp_mlsmote, y_train_mlsmote, y_temp_mlsmote = train_test_split(\n", " padded_sequences, labels_df, test_size=0.3, random_state=42, stratify=stratify_labels_df\n", ")\n", "\n", "stratify_temp_labels_df = y_temp_mlsmote.sum(axis=1)\n", "\n", "X_val_mlsmote, X_test_mlsmote, y_val_mlsmote, y_test_mlsmote = train_test_split(\n", " X_temp_mlsmote, y_temp_mlsmote, test_size=0.5, random_state=42, stratify=stratify_temp_labels_df\n", ")\n", "\n", "print_colored(f\"Dimensione Training Set:\", \"blue\")\n", "print(len(X_train_mlsmote))\n", "print_colored(f\"Dimensione Validation Set:\", \"blue\")\n", "print(len(X_val_mlsmote))\n", "print_colored(f\"Dimensione Test Set:\", \"blue\")\n", "print(len(X_test_mlsmote))\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Implemento il bilanciamento delle classi con SMOTE per ciascuna colonna

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Bilanciando la classe:\ntoxic\nDistribuzione bilanciata per toxic: Counter({0: 100996, 1: 100996})\n\nBilanciando la classe:\nsevere_toxic\nDistribuzione bilanciata per severe_toxic: Counter({0: 110589, 1: 110589})\n\nBilanciando la classe:\nobscene\nDistribuzione bilanciata per obscene: Counter({0: 105787, 1: 105787})\n\nBilanciando la classe:\nthreat\nDistribuzione bilanciata per threat: Counter({0: 111374, 1: 111374})\n\nBilanciando la classe:\ninsult\nDistribuzione bilanciata per insult: Counter({0: 106170, 1: 106170})\n\nBilanciando la classe:\nidentity_hate\nDistribuzione bilanciata per identity_hate: Counter({0: 110711, 1: 110711})\n\nBilanciamento completato per tutte le classi.\n
\n
\n
\n
\n
" ] } } ], "source": [ "if not isinstance(X_train_mlsmote, pd.DataFrame):\n", " X_train_mlsmote = pd.DataFrame(X_train_mlsmote, columns=[f'feature_{i}' for i in range(X_train_mlsmote.shape[1])])\n", "if not isinstance(y_train_mlsmote, pd.DataFrame):\n", " y_train_mlsmote = pd.DataFrame(y_train_mlsmote, columns=categories)\n", "\n", "balanced_datasets = {}\n", "\n", "for category in categories:\n", " print_colored(f\"Bilanciando la classe:\", \"blue\")\n", " print_colored(category, \"red\")\n", " \n", " smote = SMOTE(random_state=42)\n", " \n", " # Applico SMOTE per la specifica colonna\n", " X_resampled, y_resampled = smote.fit_resample(X_train_mlsmote, y_train_mlsmote[category])\n", " \n", " # Salvo il risultato per la colonna bilanciata\n", " balanced_datasets[category] = {\n", " \"X_resampled\": X_resampled,\n", " \"y_resampled\": y_resampled\n", " }\n", " \n", "\n", " print(f\"Distribuzione bilanciata per {category}: {Counter(y_resampled)}\")\n", " print()\n", "\n", "# Ogni chiave contiene \"X_resampled\" e \"y_resampled\".\n", "print_colored(\"Bilanciamento completato per tutte le classi.\", \"blue\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dimensione X_train_balanced:\n(201992, 400)\nDimensione y_train_balanced:\n(222748, 6)\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Converto dict_keys in una lista per l'accesso tramite indice\n", "categories_list = list(categories) # Se categories è dict_keys\n", "\n", "# Combino i dati riequilibrati per il training\n", "X_train_balanced = balanced_datasets[categories_list[0]][\"X_resampled\"]\n", "y_train_balanced = pd.concat(\n", " [pd.DataFrame(balanced_datasets[cat][\"y_resampled\"], columns=[cat]) for cat in categories_list],\n", " axis=1\n", ")\n", "\n", "print_colored(f\"Dimensione X_train_balanced:\", \"blue\")\n", "print(X_train_balanced.shape)\n", "print_colored(f\"Dimensione y_train_balanced:\", \"blue\")\n", "print(y_train_balanced.shape)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
toxic\nDimensioni X_resampled: (201992, 400)\nDimensioni y_resampled: (201992,)\n\nsevere_toxic\nDimensioni X_resampled: (221178, 400)\nDimensioni y_resampled: (221178,)\n\nobscene\nDimensioni X_resampled: (211574, 400)\nDimensioni y_resampled: (211574,)\n\nthreat\nDimensioni X_resampled: (222748, 400)\nDimensioni y_resampled: (222748,)\n\ninsult\nDimensioni X_resampled: (212340, 400)\nDimensioni y_resampled: (212340,)\n\nidentity_hate\nDimensioni X_resampled: (221422, 400)\nDimensioni y_resampled: (221422,)\n\n
\n
\n
\n
\n
" ] } } ], "source": [ "for category in categories:\n", " print_colored(category, \"blue\")\n", " print(f\"Dimensioni X_resampled: {balanced_datasets[category]['X_resampled'].shape}\")\n", " print(f\"Dimensioni y_resampled: {balanced_datasets[category]['y_resampled'].shape}\")\n", " print()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Calcolo il numero minimo di campioni per far corrispondere le dimensioni di X_train_balanced e y_train_balanced allo stesso valore

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Numero minimo di campioni:\n201992\n
\n
\n
\n
\n
" ] } } ], "source": [ "min_samples = min(len(balanced_datasets[cat][\"X_resampled\"]) for cat in categories)\n", "print_colored(f\"Numero minimo di campioni:\", \"blue\")\n", "print(min_samples)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Dimensione X_train_balanced: (201992, 400)\nDimensione y_train_balanced: (201992, 6)\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Riduco ogni dataset bilanciato alla dimensione minima\n", "X_train_balanced = balanced_datasets[categories_list[0]][\"X_resampled\"][:min_samples]\n", "\n", "# Combino le etichette per tutte le categorie\n", "y_train_balanced = pd.concat(\n", " [\n", " pd.DataFrame(balanced_datasets[cat][\"y_resampled\"][:min_samples], columns=[cat])\n", " for cat in categories\n", " ],\n", " axis=1\n", ")\n", "\n", "print(f\"Dimensione X_train_balanced: {X_train_balanced.shape}\")\n", "print(f\"Dimensione y_train_balanced: {y_train_balanced.shape}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Il bilanciamento è avvenuto a livello di singola classe, trattandola come un problema binario. Questo è ideale per problemi multi-label in cui ogni classe è indipendente.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "categories = balanced_datasets.keys()\n", "counts_per_category = {\n", " category: Counter(balanced_datasets[category][\"y_resampled\"])\n", " for category in categories\n", "}\n", "\n", "counts_0 = [counts_per_category[cat][0] for cat in categories]\n", "counts_1 = [counts_per_category[cat][1] for cat in categories]\n", "\n", "x = np.arange(len(categories))\n", "width = 0.4\n", "\n", "plt.figure(figsize=(12, 6))\n", "plt.bar(x - width / 2, counts_0, width, label=\"Classe 0\", alpha=0.7)\n", "plt.bar(x + width / 2, counts_1, width, label=\"Classe 1\", alpha=0.7)\n", "\n", "plt.xlabel(\"Categorie\", fontsize=16, color='black')\n", "plt.ylabel(\"Conteggio\", fontsize=16, color='black')\n", "plt.title(\"Bilanciamento delle classi dopo SMOTE\", fontsize=18)\n", "plt.xticks(x, categories, rotation=45, ha='right', color='#b81414')\n", "plt.yticks(fontsize=12, color='#b81414')\n", "plt.legend(loc='center left', bbox_to_anchor=(1, 0.5), fontsize=12)\n", "plt.grid(axis=\"y\", linestyle=\"--\", alpha=0.5, color=\"#1f77b4\")\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Vocab size padding:\n212563\n
\n
\n
\n
\n
" ] } } ], "source": [ "vocab_size_padding = len(tokenizer.word_index) + 1 # Aggiungi 1 per il token di padding\n", "print_colored(f\"Vocab size padding:\", \"blue\")\n", "print(vocab_size_padding)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "clear_session()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Creazione del primo modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Model: \"sequential\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ embedding_layer (Embedding)     │ (None, 400, 128)       │    27,208,064 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ bidirectional_lstm              │ (None, 128)            │        98,816 │\n│ (Bidirectional)                 │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ output_layer (Dense)            │ (None, 6)              │           774 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 27,307,654 (104.17 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 27,307,654 (104.17 MB)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
" ] } } ], "source": [ "model = Sequential()\n", "\n", "# Ho dovuto aggiungere i valori in modo esplicito per non farmi restituire 'unbuilt'\n", "model.add(Input(shape=(400,), name='input_layer'))\n", "\n", "model.add(Embedding(\n", " input_dim=212563, # vocab_size_padding\n", " output_dim=128,\n", " name='embedding_layer'\n", "))\n", "\n", "model.add(Bidirectional(a\n", " LSTM(64, activation='tanh'),\n", " name='bidirectional_lstm'\n", "))\n", "\n", "model.add(Dense(6, activation='sigmoid', name='output_layer')) # num_labels = 6\n", "\n", "# Compilazione\n", "model.compile(\n", " optimizer='adam',\n", " loss='binary_crossentropy',\n", " metrics=['accuracy']\n", ")\n", "\n", "model.summary()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "early_stopping = EarlyStopping(\n", " monitor='val_loss', # metrica da monitorare\n", " patience=2, # numero di epoche da aspettare prima di fermarsi\n", " restore_best_weights=True, # ripristina i migliori pesi\n", " mode='min', # minimizzare la loss\n", " min_delta=0.001, # cambiamento minimo da considerare come miglioramento\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "reduce_lr = ReduceLROnPlateau(\n", " monitor='val_loss',\n", " factor=0.2,\n", " patience=1, # riduce il learning rate dopo 1 epoca senza miglioramenti\n", " min_lr=1e-6,\n", " verbose=1\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model_checkpoint_path = \"anti_hater_model_stratify_400_128_6__1.keras\"\n", "\n", "# Callback per salvare il miglior modello basato sulla metrica monitorata\n", "model_checkpoint = ModelCheckpoint(\n", " filepath=model_checkpoint_path, # Percorso per salvare il modello\n", " monitor='val_loss', # Metrica da monitorare\n", " save_best_only=True, # Salva solo il modello migliore\n", " save_weights_only=False, # Salva l'intero modello (inclusa l'architettura)\n", " mode='min', # minimizzare la val_loss\n", " verbose=1 \n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/5\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 142ms/step - accuracy: 0.8102 - loss: 0.4710\nEpoch 1: val_loss improved from inf to 0.29014, saving model to anti_hater_model_stratify_400_128_6__1.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 916s 145ms/step - accuracy: 0.8102 - loss: 0.4709 - val_accuracy: 0.8091 - val_loss: 0.2901 - learning_rate: 0.0010\nEpoch 2/5\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 142ms/step - accuracy: 0.7150 - loss: 0.2032\nEpoch 2: val_loss improved from 0.29014 to 0.26987, saving model to anti_hater_model_stratify_400_128_6__1.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 914s 145ms/step - accuracy: 0.7150 - loss: 0.2032 - val_accuracy: 0.8569 - val_loss: 0.2699 - learning_rate: 0.0010\nEpoch 3/5\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 142ms/step - accuracy: 0.6883 - loss: 0.1091\nEpoch 3: ReduceLROnPlateau reducing learning rate to 0.00020000000949949026.\n\nEpoch 3: val_loss did not improve from 0.26987\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 913s 145ms/step - accuracy: 0.6883 - loss: 0.1091 - val_accuracy: 0.8546 - val_loss: 0.3012 - learning_rate: 0.0010\nEpoch 4/5\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 142ms/step - accuracy: 0.6895 - loss: 0.0640\nEpoch 4: ReduceLROnPlateau reducing learning rate to 4.0000001899898055e-05.\n\nEpoch 4: val_loss did not improve from 0.26987\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 915s 145ms/step - accuracy: 0.6895 - loss: 0.0640 - val_accuracy: 0.9006 - val_loss: 0.2988 - learning_rate: 2.0000e-04\nEpoch 4: early stopping\nRestoring model weights from the end of the best epoch: 2.\nIl miglior modello è stato salvato in: anti_hater_model_stratify_400_128_6__1.keras\n
\n
\n
\n
\n
" ] } } ], "source": [ "history = model.fit(\n", " X_train_balanced,\n", " y_train_balanced,\n", " validation_data=(X_val_mlsmote, y_val_mlsmote),# Usa i dati di validazione originali l set di validazione serve per monitorare il \n", " # progresso del modello durante il training\n", " epochs=5,\n", " batch_size=32,\n", " callbacks=[early_stopping, reduce_lr, model_checkpoint], \n", " verbose=1\n", ")\n", "\n", "print(f\"Il miglior modello è stato salvato in: {model_checkpoint_path}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del primo modello sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nValutazione sul Test Set:\n748/748 ━━━━━━━━━━━━━━━━━━━━ 20s 27ms/step\n\nMetriche del Test Set:\n        Category  Accuracy  F1-Score  Precision\n0          toxic  0.869694  0.527209   0.403855\n1   severe_toxic  0.897518  0.063383   0.034992\n2        obscene  0.887701  0.423176   0.291113\n3         threat  0.902323  0.007640   0.003953\n4         insult  0.879428  0.358096   0.241960\n5  identity_hate  0.896892  0.027581   0.015002\n\nInferenza su alcuni esempi del Test Set:\nCommento #1:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.12 0.   0.01 0.   0.01 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #2:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #3:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #4:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #5:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #6:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #7:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.01 0.01 0.01 0.01 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #8:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.13 0.13 0.14 0.13 0.09 0.12])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #9:\n - Predetto: [1 1 1 0 1 0] (probabilità: [1.   0.64 0.99 0.26 0.95 0.42])\n - Vero: [1 1 1 0 1 0]\n--------------------------------------------------\nCommento #10:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.02 0.01 0.01 0.02 0.02 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #11:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.01 0.01 0.01 0.01 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #12:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #13:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.3  0.14 0.19 0.14 0.18 0.15])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #14:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #15:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.04 0.   0.01 0.   0.01 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(\"\\nValutazione sul Test Set:\")\n", "y_test_pred = model.predict(X_test_mlsmote) # Predizioni sul test set\n", "y_test_pred_binary = (y_test_pred > 0.5).astype(int) # Converto probabilità in etichette binarie\n", "\n", "test_metrics_df = pd.DataFrame()\n", "test_metrics_df['Category'] = categories\n", "\n", "test_accuracies = []\n", "test_f1_scores = []\n", "test_precisions = []\n", "\n", "for i in range(len(categories)):\n", " test_acc = accuracy_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_f1 = f1_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_prec = precision_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " \n", " test_accuracies.append(test_acc)\n", " test_f1_scores.append(test_f1)\n", " test_precisions.append(test_prec)\n", "\n", "test_metrics_df['Accuracy'] = test_accuracies\n", "test_metrics_df['F1-Score'] = test_f1_scores\n", "test_metrics_df['Precision'] = test_precisions\n", "\n", "print(\"\\nMetriche del Test Set:\")\n", "print(test_metrics_df)\n", "\n", "print(\"\\nInferenza su alcuni esempi del Test Set:\")\n", "\n", "num_examples = 15\n", "for idx in range(num_examples):\n", " comment = X_test_mlsmote[idx]\n", " true_labels = y_test_mlsmote[idx]\n", " predicted_probs = y_test_pred[idx]\n", " predicted_labels = y_test_pred_binary[idx]\n", " \n", " print(f\"Commento #{idx + 1}:\")\n", " print(f\" - Predetto: {predicted_labels} (probabilità: {predicted_probs.round(2)})\")\n", " print(f\" - Vero: {true_labels}\")\n", " print(\"-\" * 50)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Precisione Globale (Global Accuracy):\n0.8844\n
\n
\n
\n
\n
" ] } } ], "source": [ "# Calcolo la Hamming Loss tra le etichette vere (ytest) e le etichette predette (ypred)\n", "# la somma degli errori tra ytest e il ypred diviso il numero totale di etichette\n", "global_hamming_loss = hamming_loss(y_test_mlsmote, y_test_pred_binary)\n", "\n", "# Inverto la Hamming Loss per ottenere la Global Accuracy\n", "global_accuracy_1 = 1 - global_hamming_loss\n", "\n", "print_colored(f\"Precisione Globale (Global Accuracy):\", \"blue\")\n", "print(f\"{global_accuracy_1:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per toxic:\n[[19064  2581]\n [  621  1670]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per severe_toxic:\n[[21287  2402]\n [  159    88]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per obscene:\n[[20164  2499]\n [  307   966]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per threat:\n[[21497  2360]\n [   71     8]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per insult:\n[[20119  2648]\n [  366   803]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per identity_hate:\n[[21317  2414]\n [  173    32]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "conf_matrices = multilabel_confusion_matrix(y_test_mlsmote, y_test_pred_binary)\n", "\n", "# Controllo dimensioni\n", "if conf_matrices.shape[0] != len(categories):\n", " print(\"Errore: Il numero di confusion matrix non corrisponde al numero di categorie.\")\n", "else:\n", " for i, category in enumerate(categories):\n", " plt.figure(figsize=(6, 4))\n", " \n", " conf_matrix = conf_matrices[i]\n", " \n", " labels = np.array([['TN', 'FP'], ['FN', 'TP']])\n", " annotated_matrix = np.empty_like(conf_matrix, dtype=object)\n", " for row in range(conf_matrix.shape[0]):\n", " for col in range(conf_matrix.shape[1]):\n", " annotated_matrix[row, col] = f\"{labels[row, col]}: {conf_matrix[row, col]}\"\n", " \n", " sns.heatmap(conf_matrix, annot=annotated_matrix, fmt='', cmap='Blues', cbar=False)\n", " plt.title(f\"Confusion Matrix per {category}\")\n", " plt.xlabel(\"Predizioni\")\n", " plt.ylabel(\"Valori Reali\")\n", " plt.show()\n", " \n", " print(f\"Confusion Matrix per {category}:\")\n", " print(conf_matrix)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " fpr, tpr, _ = roc_curve(y_test_mlsmote[:, i], y_test_pred[:, i])\n", " roc_auc = auc(fpr, tpr)\n", " \n", " plt.figure(figsize=(6, 4))\n", " plt.plot(fpr, tpr, label=f'AUC = {roc_auc:.2f}')\n", " plt.plot([0, 1], [0, 1], 'k--', label='Random Guessing')\n", " plt.title(f\"ROC Curve per {category}\")\n", " plt.xlabel('False Positive Rate')\n", " plt.ylabel('True Positive Rate')\n", " plt.legend(loc='lower right')\n", " plt.grid(alpha=0.3)\n", " plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nPrecision-Recall per toxic (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.10        1.00        \n0.00      0.11        1.00        \n0.01      0.12        0.99        \n0.01      0.14        0.98        \n0.03      0.17        0.96        \n0.07      0.20        0.93        \n0.16      0.25        0.87        \n0.37      0.34        0.78        \n0.75      0.53        0.61        \n1.00      1.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per severe_toxic (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.02        1.00        \n0.01      0.02        1.00        \n0.02      0.02        0.99        \n0.06      0.03        0.96        \n0.18      0.04        0.82        \n0.47      0.04        0.38        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per obscene (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.05        1.00        \n0.00      0.06        1.00        \n0.00      0.07        0.99        \n0.01      0.08        0.98        \n0.01      0.09        0.98        \n0.03      0.11        0.96        \n0.09      0.15        0.92        \n0.26      0.20        0.85        \n0.63      0.33        0.69        \n1.00      1.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per threat (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.00        0.99        \n0.01      0.01        0.96        \n0.02      0.01        0.89        \n0.06      0.01        0.73        \n0.17      0.01        0.43        \n0.45      0.00        0.13        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per insult (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.05        1.00        \n0.00      0.05        1.00        \n0.00      0.06        0.99        \n0.01      0.07        0.99        \n0.01      0.09        0.98        \n0.04      0.11        0.96        \n0.10      0.13        0.91        \n0.27      0.18        0.81        \n0.62      0.27        0.62        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per identity_hate (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        0.99        \n0.01      0.02        0.98        \n0.02      0.02        0.91        \n0.07      0.02        0.82        \n0.19      0.02        0.57        \n0.47      0.01        0.18        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " precision, recall, thresholds = precision_recall_curve(y_test_mlsmote[:, i], y_test_pred[:, i])\n", " \n", " print(f\"\\nPrecision-Recall per {category} (valori campionati):\")\n", " print(f\"{'Threshold':<10}{'Precision':<12}{'Recall':<12}\")\n", " print(\"-\" * 34)\n", " \n", " sampled_indices = np.linspace(0, len(thresholds) - 1, 10, dtype=int) # Campiona 10 valori\n", " for idx in sampled_indices:\n", " print(f\"{thresholds[idx]:<10.2f}{precision[idx]:<12.2f}{recall[idx]:<12.2f}\")\n", " \n", " plt.figure(figsize=(6, 4))\n", " plt.plot(recall, precision, label='Precision-Recall Curve')\n", " plt.title(f\"Precision-Recall Curve per {category}\")\n", " plt.xlabel('Recall')\n", " plt.ylabel('Precision')\n", " plt.legend(loc='upper right')\n", " plt.grid(alpha=0.3)\n", " plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

I valori di val_accuracy, Auc Roc e Global Accuracy non sono bassi ma alcune categorie presentano valori bassissimi di F1-score e Precision. Il modello tende a classificare molti commenti innocui come tossici.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Creazione di un secondo modello LSTM usando come base il primo modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Sommario del Modello:\n
\n
\n
\n
\n
\n
\n
Model: \"sequential\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ embedding_layer (Embedding)     │ (None, 400, 128)       │    27,208,064 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ bidirectional_lstm              │ (None, 128)            │        98,816 │\n│ (Bidirectional)                 │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ output_layer (Dense)            │ (None, 6)              │           774 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 81,922,964 (312.51 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 27,307,654 (104.17 MB)\n
\n
\n
\n
\n
\n
\n
 Optimizer params: 54,615,310 (208.34 MB)\n
\n
\n
\n
\n
\n
\n
Model: \"functional_34\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ input_layer_4 (InputLayer)      │ (None, 400)            │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ embedding_layer (Embedding)     │ (None, 400, 128)       │    27,208,064 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ bidirectional_lstm              │ (None, 128)            │        98,816 │\n│ (Bidirectional)                 │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dense_1 (Dense)             │ (None, 64)             │         8,256 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dropout (Dropout)           │ (None, 64)             │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dense_2 (Dense)             │ (None, 32)             │         2,080 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_output_layer (Dense)        │ (None, 6)              │           198 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 27,317,414 (104.21 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 10,534 (41.15 KB)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 27,306,880 (104.17 MB)\n
\n
\n
\n
\n
" ] } } ], "source": [ "def load_model_safely(path):\n", " try:\n", " model = load_model(path)\n", " return model\n", " except Exception as e:\n", " print(f\"Errore nel caricamento del modello: {e}\")\n", " return None\n", "\n", "model_checkpoint_path = \"/home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_6__1.keras\"\n", "\n", "first_model = load_model_safely(model_checkpoint_path)\n", "\n", "if first_model is None:\n", " print(\"Impossibile caricare il modello. Controllare il file del modello.\")\n", "else:\n", "\n", " # Preservo i pesi già addestrati per evitare che vengano riaddestrati\n", " for layer in first_model.layers:\n", " layer.trainable = False # Congelo tutti i layer esistenti\n", " \n", " print(\"Sommario del Modello:\")\n", " first_model.summary()\n", "\n", " # Estraggo la forma dell'input, rimuovendo la dimensione del batch\n", " if isinstance(first_model.input_shape, tuple):\n", " input_shape = first_model.input_shape[1:] # Rimuove la dimensione del batch\n", " else:\n", " input_shape = first_model.input_shape[0][1:] # input multipli\n", " \n", " # Creo un nuovo strato di input con la forma estratta\n", " inputs = Input(shape=input_shape)\n", " \n", " # Clono i layer del modello originale (escludendo l'ultimo layer di output)\n", " x = inputs\n", " for layer in first_model.layers[:-1]:\n", " x = layer(x)\n", " \n", " # Nuovi layer\n", " x = Dense(64, activation='relu', name='new_dense_1')(x)\n", " \n", " # Layer di Dropout per ridurre l'overfitting\n", " x = Dropout(0.5, name='new_dropout')(x)\n", " \n", " x = Dense(32, activation='relu', name='new_dense_2')(x)\n", " \n", " # Layer di output con 6 neuroni e attivazione sigmoid per classificazione multi-label\n", " output_layer = Dense(6, activation='sigmoid', name='new_output_layer')(x)\n", " \n", " second_model = Model(inputs=inputs, outputs=output_layer)\n", " \n", " second_model.compile(\n", " optimizer='adam',\n", " loss='binary_crossentropy', # Funzione di perdita per problemi multi-label\n", " metrics=['accuracy'] # Metrica per valutare le prestazioni\n", " )\n", " \n", " second_model.summary() \n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "early_stopping = EarlyStopping(\n", " monitor='val_loss', # metrica da monitorare\n", " patience=3, # numero di epoche da aspettare prima di fermarsi\n", " restore_best_weights=True, # ripristina i migliori pesi\n", " mode='min', # minimizzare la loss\n", " min_delta=0.001, # cambiamento minimo da considerare come miglioramento\n", " verbose=1 \n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "reduce_lr = ReduceLROnPlateau(\n", " monitor='val_loss',\n", " factor=0.2,\n", " patience=1, \n", " min_lr=1e-6,\n", " verbose=1\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model_checkpoint_path = \"/home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_64_32_dropout0.5_6__2.keras\"\n", "\n", "# Callback per salvare il miglior modello basato sulla metrica monitorata\n", "model_checkpoint = ModelCheckpoint(\n", " filepath=model_checkpoint_path, # Percorso per salvare il modello\n", " monitor='val_loss', # Metrica da monitorare\n", " save_best_only=True, # Salvo solo il modello migliore\n", " save_weights_only=False, # Salvo l'intero modello (inclusa l'architettura)\n", " mode='min', # minimizzare la val_loss\n", " verbose=1 \n", ")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Fit del secondo modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/5\n6312/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 40ms/step - accuracy: 0.6845 - loss: 0.1304\nEpoch 1: val_loss improved from inf to 0.30043, saving model to /home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_64_32_dropout0.5_6__2.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 273s 43ms/step - accuracy: 0.6845 - loss: 0.1304 - val_accuracy: 0.9256 - val_loss: 0.3004 - learning_rate: 0.0010\nEpoch 2/5\n6312/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.7937 - loss: 0.1024\nEpoch 2: val_loss improved from 0.30043 to 0.29876, saving model to /home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_64_32_dropout0.5_6__2.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 266s 42ms/step - accuracy: 0.7937 - loss: 0.1024 - val_accuracy: 0.9316 - val_loss: 0.2988 - learning_rate: 0.0010\nEpoch 3/5\n6312/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.7844 - loss: 0.1014\nEpoch 3: ReduceLROnPlateau reducing learning rate to 0.00020000000949949026.\n\nEpoch 3: val_loss did not improve from 0.29876\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 266s 42ms/step - accuracy: 0.7844 - loss: 0.1014 - val_accuracy: 0.9365 - val_loss: 0.3012 - learning_rate: 0.0010\nEpoch 4/5\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.8049 - loss: 0.0999\nEpoch 4: ReduceLROnPlateau reducing learning rate to 4.0000001899898055e-05.\n\nEpoch 4: val_loss did not improve from 0.29876\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 267s 42ms/step - accuracy: 0.8049 - loss: 0.0999 - val_accuracy: 0.9095 - val_loss: 0.3016 - learning_rate: 2.0000e-04\nEpoch 5/5\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.8330 - loss: 0.0977\nEpoch 5: val_loss improved from 0.29876 to 0.29588, saving model to /home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_64_32_dropout0.5_6__2.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 266s 42ms/step - accuracy: 0.8330 - loss: 0.0977 - val_accuracy: 0.9022 - val_loss: 0.2959 - learning_rate: 4.0000e-05\nRestoring model weights from the end of the best epoch: 5.\nIl miglior modello è stato salvato in: /home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_64_32_dropout0.5_6__2.keras\n
\n
\n
\n
\n
" ] } } ], "source": [ "history = second_model.fit(\n", " X_train_balanced,\n", " y_train_balanced,\n", " validation_data=(X_val_mlsmote, y_val_mlsmote), # dati di validazione originali\n", " epochs=5,\n", " batch_size=32,\n", " callbacks=[early_stopping, reduce_lr, model_checkpoint], \n", " verbose=1\n", ")\n", "\n", "# Messaggio finale\n", "print(f\"Il miglior modello è stato salvato in: {model_checkpoint_path}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del secondo modello sul Test Set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nValutazione sul Test Set:\n748/748 ━━━━━━━━━━━━━━━━━━━━ 20s 27ms/step\n\nMetriche del Test Set:\n        Category  Accuracy  F1-Score  Precision\n0          toxic  0.869527  0.524151   0.402622\n1   severe_toxic  0.890709  0.069701   0.038207\n2        obscene  0.884442  0.407709   0.280247\n3         threat  0.895931  0.007965   0.004112\n4         insult  0.876379  0.346655   0.233631\n5  identity_hate  0.891544  0.022590   0.012240\n\nInferenza su alcuni esempi del Test Set:\nCommento #1:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.24 0.   0.01 0.   0.02 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #2:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #3:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #4:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #5:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #6:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #7:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.02 0.01 0.01 0.01 0.01 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #8:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.1  0.1  0.09 0.09 0.09 0.09])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #9:\n - Predetto: [1 1 1 0 1 0] (probabilità: [1.   0.65 0.99 0.33 0.93 0.47])\n - Vero: [1 1 1 0 1 0]\n--------------------------------------------------\nCommento #10:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.03 0.02 0.02 0.02 0.02 0.02])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #11:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #12:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #13:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.18 0.08 0.11 0.09 0.13 0.09])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #14:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #15:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.04 0.   0.01 0.   0.01 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(\"\\nValutazione sul Test Set:\")\n", "y_test_pred = second_model.predict(X_test_mlsmote) \n", "y_test_pred_binary = (y_test_pred > 0.5).astype(int)\n", "\n", "test_metrics_df_2 = pd.DataFrame()\n", "test_metrics_df_2['Category'] = categories\n", "\n", "test_accuracies = []\n", "test_f1_scores = []\n", "test_precisions = []\n", "\n", "for i in range(len(categories)):\n", " test_acc = accuracy_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_f1 = f1_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_prec = precision_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " \n", " test_accuracies.append(test_acc)\n", " test_f1_scores.append(test_f1)\n", " test_precisions.append(test_prec)\n", "\n", "test_metrics_df_2['Accuracy'] = test_accuracies\n", "test_metrics_df_2['F1-Score'] = test_f1_scores\n", "test_metrics_df_2['Precision'] = test_precisions\n", "\n", "print(\"\\nMetriche del Test Set:\")\n", "print(test_metrics_df_2)\n", "\n", "\n", "print(\"\\nInferenza su alcuni esempi del Test Set:\")\n", "\n", "num_examples = 15\n", "for idx in range(num_examples):\n", " comment = X_test_mlsmote[idx]\n", " true_labels = y_test_mlsmote[idx]\n", " predicted_probs = y_test_pred[idx]\n", " predicted_labels = y_test_pred_binary[idx]\n", " \n", " print(f\"Commento #{idx + 1}:\")\n", " print(f\" - Predetto: {predicted_labels} (probabilità: {predicted_probs.round(2)})\")\n", " print(f\" - Vero: {true_labels}\")\n", " print(\"-\" * 50)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Precisione Globale (Global Accuracy):\n0.8889\n
\n
\n
\n
\n
" ] } } ], "source": [ "global_hamming_loss = hamming_loss(y_test_mlsmote, y_test_pred_binary)\n", "\n", "global_accuracy_2 = 1 - global_hamming_loss\n", "\n", "print_colored(f\"Precisione Globale (Global Accuracy):\", \"blue\")\n", "print(f\"{global_accuracy_2:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per toxic:\n[[19093  2552]\n [  571  1720]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per severe_toxic:\n[[21222  2467]\n [  149    98]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per obscene:\n[[20218  2445]\n [  321   952]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per threat:\n[[21435  2422]\n [   69    10]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per insult:\n[[20192  2575]\n [  384   785]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per identity_hate:\n[[21310  2421]\n [  175    30]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "conf_matrices = multilabel_confusion_matrix(y_test_mlsmote, y_test_pred_binary)\n", "\n", "if conf_matrices.shape[0] != len(categories):\n", " print(\"Errore: Il numero di confusion matrix non corrisponde al numero di categorie.\")\n", "else:\n", " for i, category in enumerate(categories):\n", " plt.figure(figsize=(6, 4))\n", " \n", " conf_matrix = conf_matrices[i]\n", " \n", " labels = np.array([['TN', 'FP'], ['FN', 'TP']])\n", " annotated_matrix = np.empty_like(conf_matrix, dtype=object)\n", " for row in range(conf_matrix.shape[0]):\n", " for col in range(conf_matrix.shape[1]):\n", " annotated_matrix[row, col] = f\"{labels[row, col]}: {conf_matrix[row, col]}\"\n", " \n", " sns.heatmap(conf_matrix, annot=annotated_matrix, fmt='', cmap='Blues', cbar=False)\n", " plt.title(f\"Confusion Matrix per {category}\")\n", " plt.xlabel(\"Predizioni\")\n", " plt.ylabel(\"Valori Reali\")\n", " plt.show()\n", " \n", " print(f\"Confusion Matrix per {category}:\")\n", " print(conf_matrix)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "from sklearn.metrics import roc_auc_score, roc_curve, auc\n", "\n", "# Calcolo AUC per ogni categoria\n", "for i, category in enumerate(categories):\n", " fpr, tpr, _ = roc_curve(y_test_mlsmote[:, i], y_test_pred[:, i])\n", " roc_auc = auc(fpr, tpr)\n", " \n", " plt.figure(figsize=(6, 4))\n", " plt.plot(fpr, tpr, label=f'AUC = {roc_auc:.2f}')\n", " plt.plot([0, 1], [0, 1], 'k--', label='Random Guessing')\n", " plt.title(f\"ROC Curve per {category}\")\n", " plt.xlabel('False Positive Rate')\n", " plt.ylabel('True Positive Rate')\n", " plt.legend(loc='lower right')\n", " plt.grid(alpha=0.3)\n", " plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nPrecision-Recall per toxic (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.10        1.00        \n0.00      0.11        1.00        \n0.00      0.12        0.99        \n0.00      0.14        0.98        \n0.01      0.17        0.97        \n0.03      0.20        0.94        \n0.10      0.26        0.90        \n0.32      0.35        0.82        \n0.79      0.51        0.60        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per severe_toxic (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.02        1.00        \n0.00      0.02        1.00        \n0.00      0.02        1.00        \n0.02      0.03        0.98        \n0.10      0.04        0.92        \n0.47      0.04        0.42        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per obscene (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.05        1.00        \n0.00      0.06        1.00        \n0.00      0.07        0.99        \n0.00      0.08        0.99        \n0.00      0.09        0.98        \n0.01      0.12        0.97        \n0.04      0.15        0.96        \n0.18      0.21        0.90        \n0.68      0.32        0.67        \n1.00      1.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per threat (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.01        1.00        \n0.00      0.01        0.97        \n0.02      0.01        0.84        \n0.09      0.01        0.54        \n0.42      0.00        0.14        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per insult (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.05        1.00        \n0.00      0.05        1.00        \n0.00      0.06        1.00        \n0.00      0.07        0.99        \n0.00      0.09        0.98        \n0.01      0.11        0.97        \n0.04      0.14        0.95        \n0.20      0.19        0.87        \n0.66      0.25        0.56        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per identity_hate (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.02        1.00        \n0.00      0.02        0.96        \n0.02      0.02        0.89        \n0.11      0.02        0.64        \n0.43      0.01        0.19        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " precision, recall, thresholds = precision_recall_curve(y_test_mlsmote[:, i], y_test_pred[:, i])\n", " \n", " print(f\"\\nPrecision-Recall per {category} (valori campionati):\")\n", " print(f\"{'Threshold':<10}{'Precision':<12}{'Recall':<12}\")\n", " print(\"-\" * 34)\n", " \n", " # Indici campionati per valori distribuiti\n", " sampled_indices = np.linspace(0, len(thresholds) - 1, 10, dtype=int) # Campiona 10 valori\n", " for idx in sampled_indices:\n", " print(f\"{thresholds[idx]:<10.2f}{precision[idx]:<12.2f}{recall[idx]:<12.2f}\")\n", " \n", " plt.figure(figsize=(6, 4))\n", " plt.plot(recall, precision, label='Precision-Recall Curve')\n", " plt.title(f\"Precision-Recall Curve per {category}\")\n", " plt.xlabel('Recall')\n", " plt.ylabel('Precision')\n", " plt.legend(loc='upper right')\n", " plt.grid(alpha=0.3)\n", " plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

I valori di val_accuracy, Auc Roc e Global Accuracy non sono bassi ma alcune categorie presentano valori bassissimi di F1-score e Precision. Il modello tende a classificare molti commenti innocui come tossici.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Creazione di un terzo modello LSTM usando come base il primo modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Sommario del Modello:\n
\n
\n
\n
\n
\n
\n
Model: \"sequential\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ embedding_layer (Embedding)     │ (None, 400, 128)       │    27,208,064 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ bidirectional_lstm              │ (None, 128)            │        98,816 │\n│ (Bidirectional)                 │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ output_layer (Dense)            │ (None, 6)              │           774 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 81,922,964 (312.51 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 27,307,654 (104.17 MB)\n
\n
\n
\n
\n
\n
\n
 Optimizer params: 54,615,310 (208.34 MB)\n
\n
\n
\n
\n
\n
\n
Model: \"functional_52\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ input_layer_7 (InputLayer)      │ (None, 400)            │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ embedding_layer (Embedding)     │ (None, 400, 128)       │    27,208,064 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ bidirectional_lstm              │ (None, 128)            │        98,816 │\n│ (Bidirectional)                 │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dense_1 (Dense)             │ (None, 256)            │        33,024 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dropout (Dropout)           │ (None, 256)            │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dense_2 (Dense)             │ (None, 128)            │        32,896 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dropout_1 (Dropout)         │ (None, 128)            │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dense_3 (Dense)             │ (None, 64)             │         8,256 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dropout_2 (Dropout)         │ (None, 64)             │             0 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_dense_4 (Dense)             │ (None, 32)             │         2,080 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ new_output_layer (Dense)        │ (None, 6)              │           198 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 27,383,334 (104.46 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 76,454 (298.65 KB)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 27,306,880 (104.17 MB)\n
\n
\n
\n
\n
" ] } } ], "source": [ "def load_model_safely(path):\n", " try:\n", " model = load_model(path)\n", " return model\n", " except Exception as e:\n", " print(f\"Errore nel caricamento del modello: {e}\")\n", " return None\n", "\n", "model_checkpoint_path = \"/home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_6__1.keras\"\n", "\n", "first_model = load_model_safely(model_checkpoint_path)\n", "\n", "if first_model is None:\n", " print(\"Impossibile caricare il modello. Controllare il file del modello.\")\n", "else:\n", "\n", " # Preservo i pesi già addestrati per evitare che vengano riaddestrati\n", " for layer in first_model.layers:\n", " layer.trainable = False # Congelo tutti i layer esistenti\n", " \n", " print(\"Sommario del Modello:\")\n", " first_model.summary()\n", "\n", " # Estraggo la forma dell'input, rimuovendo la dimensione del batch\n", " if isinstance(first_model.input_shape, tuple):\n", " input_shape = first_model.input_shape[1:] # Rimuovo la dimensione del batch\n", " else:\n", " input_shape = first_model.input_shape[0][1:] # input multipli\n", " \n", " # Creo un nuovo strato di input con la forma estratta\n", " inputs = Input(shape=input_shape)\n", " \n", " # Clono i layer del modello originale (escludendo l'ultimo layer di output)\n", " x = inputs\n", " for layer in first_model.layers[:-1]:\n", " x = layer(x)\n", " \n", " # ReLU aggiunge non-linearità al modello accelerando la convergenza\n", " x = Dense(256, activation='relu', name='new_dense_1')(x)\n", " \n", " # Layer di Dropout per ridurre l'overfitting\n", " x = Dropout(0.5, name='new_dropout')(x)\n", "\n", " x = Dense(128, activation='relu', name='new_dense_2')(x)\n", "\n", " x = Dropout(0.5, name='new_dropout_1')(x)\n", "\n", " x = Dense(64, activation='relu', name='new_dense_3')(x)\n", " \n", " x = Dropout(0.5, name='new_dropout_2')(x)\n", " \n", " x = Dense(32, activation='relu', name='new_dense_4')(x)\n", " \n", " # Layer di output con 6 neuroni e attivazione sigmoid per classificazione multi-label\n", " output_layer = Dense(6, activation='sigmoid', name='new_output_layer')(x)\n", " \n", " thirth_model = Model(inputs=inputs, outputs=output_layer)\n", " \n", "\n", " thirth_model.compile(\n", " optimizer='adam', \n", " loss='binary_crossentropy', # Funzione di perdita per problemi multi-label\n", " metrics=['accuracy'] \n", " )\n", "\n", " thirth_model.summary() \n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "early_stopping = EarlyStopping(\n", " monitor='val_loss', # metrica da monitorare\n", " patience=3, # numero di epoche da aspettare prima di fermarsi\n", " restore_best_weights=True, # ripristina i migliori pesi\n", " mode='min', # minimizzare la loss\n", " min_delta=0.001, # cambiamento minimo da considerare come miglioramento\n", " verbose=1 \n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "reduce_lr = ReduceLROnPlateau(\n", " monitor='val_loss',\n", " factor=0.2,\n", " patience=2, \n", " min_lr=1e-6,\n", " verbose=1\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model_checkpoint_path = \"anti_hater_model_stratify_400_256_128_64_32_dropout0.5_6__3.keras\"\n", "\n", "model_checkpoint = ModelCheckpoint(\n", " filepath=model_checkpoint_path, # Percorso per salvare il modello\n", " monitor='val_loss', # Metrica da monitorare\n", " save_best_only=True, # Salvo solo il modello migliore\n", " save_weights_only=False, # Salvo l'intero modello (inclusa l'architettura)\n", " mode='min', # minimizzare la val_loss\n", " verbose=1 \n", ")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Fit del terzo modello

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/10\n6312/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.5496 - loss: 0.1317\nEpoch 1: val_loss improved from inf to 0.36338, saving model to anti_hater_model_stratify_400_256_128_64_32_dropout0.5_6__3.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 263s 42ms/step - accuracy: 0.5496 - loss: 0.1317 - val_accuracy: 0.9337 - val_loss: 0.3634 - learning_rate: 0.0010\nEpoch 2/10\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.7276 - loss: 0.1039\nEpoch 2: val_loss improved from 0.36338 to 0.29900, saving model to anti_hater_model_stratify_400_256_128_64_32_dropout0.5_6__3.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 264s 42ms/step - accuracy: 0.7276 - loss: 0.1039 - val_accuracy: 0.9352 - val_loss: 0.2990 - learning_rate: 0.0010\nEpoch 3/10\n6312/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.7599 - loss: 0.1054\nEpoch 3: val_loss did not improve from 0.29900\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 263s 42ms/step - accuracy: 0.7599 - loss: 0.1054 - val_accuracy: 0.5988 - val_loss: 0.3403 - learning_rate: 0.0010\nEpoch 4/10\n6312/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.8286 - loss: 0.1035\nEpoch 4: ReduceLROnPlateau reducing learning rate to 0.00020000000949949026.\n\nEpoch 4: val_loss did not improve from 0.29900\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 262s 41ms/step - accuracy: 0.8286 - loss: 0.1035 - val_accuracy: 0.9694 - val_loss: 0.3085 - learning_rate: 0.0010\nEpoch 5/10\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 39ms/step - accuracy: 0.9264 - loss: 0.1027\nEpoch 5: val_loss did not improve from 0.29900\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 262s 41ms/step - accuracy: 0.9264 - loss: 0.1027 - val_accuracy: 0.9387 - val_loss: 0.3232 - learning_rate: 2.0000e-04\nEpoch 5: early stopping\nRestoring model weights from the end of the best epoch: 2.\nIl miglior modello è stato salvato in:\nanti_hater_model_stratify_400_256_128_64_32_dropout0.5_6__3.keras\n
\n
\n
\n
\n
" ] } } ], "source": [ "history = thirth_model.fit(\n", " X_train_balanced,\n", " y_train_balanced,\n", " validation_data=(X_val_mlsmote, y_val_mlsmote),\n", " epochs=10,\n", " batch_size=32,\n", " callbacks=[early_stopping, reduce_lr, model_checkpoint], \n", " verbose=1\n", ")\n", "\n", "print_colored(f\"Il miglior modello è stato salvato in:\", \"blue\")\n", "print(model_checkpoint_path)\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del terzo modello sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nValutazione sul Test Set:\n748/748 ━━━━━━━━━━━━━━━━━━━━ 20s 26ms/step\n\nMetriche del Test Set:\n        Category  Accuracy  F1-Score  Precision\n0          toxic  0.863929  0.514533   0.390675\n1   severe_toxic  0.886990  0.025225   0.013845\n2        obscene  0.879763  0.406108   0.275399\n3         threat  0.892421  0.006942   0.003580\n4         insult  0.871950  0.334997   0.224419\n5  identity_hate  0.887909  0.016135   0.008723\n\nInferenza su alcuni esempi del Test Set:\nCommento #1:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.24 0.01 0.05 0.01 0.04 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #2:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.04 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #3:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #4:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.01 0.   0.01 0.   0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #5:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #6:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.03 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #7:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.04 0.03 0.02 0.03 0.02 0.03])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #8:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.23 0.18 0.16 0.18 0.17 0.19])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #9:\n - Predetto: [1 0 1 0 1 0] (probabilità: [1.   0.45 0.96 0.28 0.85 0.39])\n - Vero: [1 1 1 0 1 0]\n--------------------------------------------------\nCommento #10:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.07 0.06 0.04 0.08 0.03 0.06])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #11:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.01 0.   0.01 0.   0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #12:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #13:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.3  0.15 0.19 0.15 0.2  0.17])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #14:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #15:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.1  0.01 0.03 0.02 0.03 0.02])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(\"\\nValutazione sul Test Set:\")\n", "y_test_pred = thirth_model.predict(X_test_mlsmote) \n", "y_test_pred_binary = (y_test_pred > 0.5).astype(int) \n", "\n", "test_metrics_df_3 = pd.DataFrame()\n", "test_metrics_df_3['Category'] = categories\n", "\n", "test_accuracies = []\n", "test_f1_scores = []\n", "test_precisions = []\n", "\n", "for i in range(len(categories)):\n", " test_acc = accuracy_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_f1 = f1_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_prec = precision_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " \n", " test_accuracies.append(test_acc)\n", " test_f1_scores.append(test_f1)\n", " test_precisions.append(test_prec)\n", "\n", "test_metrics_df_3['Accuracy'] = test_accuracies\n", "test_metrics_df_3['F1-Score'] = test_f1_scores\n", "test_metrics_df_3['Precision'] = test_precisions\n", "\n", "print(\"\\nMetriche del Test Set:\")\n", "print(test_metrics_df_3)\n", "\n", "\n", "print(\"\\nInferenza su alcuni esempi del Test Set:\")\n", "\n", "num_examples = 15\n", "for idx in range(num_examples):\n", " comment = X_test_mlsmote[idx]\n", " true_labels = y_test_mlsmote[idx]\n", " predicted_probs = y_test_pred[idx]\n", " predicted_labels = y_test_pred_binary[idx]\n", " \n", " print(f\"Commento #{idx + 1}:\")\n", " print(f\" - Predetto: {predicted_labels} (probabilità: {predicted_probs.round(2)})\")\n", " print(f\" - Vero: {true_labels}\")\n", " print(\"-\" * 50)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Precisione Globale (Global Accuracy):\n0.8805\n
\n
\n
\n
\n
" ] } } ], "source": [ "global_hamming_loss = hamming_loss(y_test_mlsmote, y_test_pred_binary)\n", "\n", "global_accuracy_3 = 1 - global_hamming_loss\n", "\n", "print_colored(f\"Precisione Globale (Global Accuracy):\", \"blue\")\n", "print(f\"{global_accuracy_3:.4f}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Confronto tra i modelli

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Categoria toxic\n   Primo     Modello:   Accuracy: 0.8697        F1-Score: 0.5272        Precision: 0.4039       Global Accuracy: 0.8844    \n   Secondo   Modello:   Accuracy: 0.8695        F1-Score: 0.5242        Precision: 0.4026       Global Accuracy: 0.8844    \n   Terzo     Modello:   Accuracy: 0.8639        F1-Score: 0.5145        Precision: 0.3907       Global Accuracy: 0.8805    \n\nCategoria severe_toxic\n   Primo     Modello:   Accuracy: 0.8975        F1-Score: 0.0634        Precision: 0.0350       Global Accuracy: 0.8844    \n   Secondo   Modello:   Accuracy: 0.8907        F1-Score: 0.0697        Precision: 0.0382       Global Accuracy: 0.8844    \n   Terzo     Modello:   Accuracy: 0.8870        F1-Score: 0.0252        Precision: 0.0138       Global Accuracy: 0.8805    \n\nCategoria obscene\n   Primo     Modello:   Accuracy: 0.8877        F1-Score: 0.4232        Precision: 0.2911       Global Accuracy: 0.8844    \n   Secondo   Modello:   Accuracy: 0.8844        F1-Score: 0.4077        Precision: 0.2802       Global Accuracy: 0.8844    \n   Terzo     Modello:   Accuracy: 0.8798        F1-Score: 0.4061        Precision: 0.2754       Global Accuracy: 0.8805    \n\nCategoria threat\n   Primo     Modello:   Accuracy: 0.9023        F1-Score: 0.0076        Precision: 0.0040       Global Accuracy: 0.8844    \n   Secondo   Modello:   Accuracy: 0.8959        F1-Score: 0.0080        Precision: 0.0041       Global Accuracy: 0.8844    \n   Terzo     Modello:   Accuracy: 0.8924        F1-Score: 0.0069        Precision: 0.0036       Global Accuracy: 0.8805    \n\nCategoria insult\n   Primo     Modello:   Accuracy: 0.8794        F1-Score: 0.3581        Precision: 0.2420       Global Accuracy: 0.8844    \n   Secondo   Modello:   Accuracy: 0.8764        F1-Score: 0.3467        Precision: 0.2336       Global Accuracy: 0.8844    \n   Terzo     Modello:   Accuracy: 0.8720        F1-Score: 0.3350        Precision: 0.2244       Global Accuracy: 0.8805    \n\nCategoria identity_hate\n   Primo     Modello:   Accuracy: 0.8969        F1-Score: 0.0276        Precision: 0.0150       Global Accuracy: 0.8844    \n   Secondo   Modello:   Accuracy: 0.8915        F1-Score: 0.0226        Precision: 0.0122       Global Accuracy: 0.8844    \n   Terzo     Modello:   Accuracy: 0.8879        F1-Score: 0.0161        Precision: 0.0087       Global Accuracy: 0.8805    \n\n
\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " print_colored(f\"Categoria {category}\\n\".ljust(16), \"blue\", end=\"\")\n", " \n", " # Primo Modello\n", " print_colored(f\" Primo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy = test_metrics_df['Accuracy'][i]\n", " f1_score = test_metrics_df['F1-Score'][i]\n", " precision = test_metrics_df['Precision'][i]\n", " print_colored(f\" Accuracy: {accuracy:<10.4f} \", bg_color=\"blue\" if accuracy == max(accuracy, test_metrics_df_2['Accuracy'][i], test_metrics_df_3['Accuracy'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score:<10.4f} \", bg_color=\"blue\" if f1_score == max(f1_score, test_metrics_df_2['F1-Score'][i], test_metrics_df_3['F1-Score'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision:<10.4f}\", bg_color=\"blue\" if precision == max(precision, test_metrics_df_2['Precision'][i], test_metrics_df_3['Precision'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_1:<10.4f}\", \"black\")\n", "\n", " # Secondo Modello\n", " print_colored(f\" Secondo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_2 = test_metrics_df_2['Accuracy'][i]\n", " f1_score_2 = test_metrics_df_2['F1-Score'][i]\n", " precision_2 = test_metrics_df_2['Precision'][i]\n", " print_colored(f\" Accuracy: {accuracy_2:<10.4f} \", bg_color=\"blue\" if accuracy_2 == max(accuracy, accuracy_2, test_metrics_df_3['Accuracy'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score_2:<10.4f} \", bg_color=\"blue\" if f1_score_2 == max(f1_score, f1_score_2, test_metrics_df_3['F1-Score'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision_2:<10.4f}\", bg_color=\"blue\" if precision_2 == max(precision, precision_2, test_metrics_df_3['Precision'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_1:<10.4f}\", \"black\")\n", "\n", " # Terzo Modello\n", " print_colored(f\" Terzo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_3 = test_metrics_df_3['Accuracy'][i]\n", " f1_score_3 = test_metrics_df_3['F1-Score'][i]\n", " precision_3 = test_metrics_df_3['Precision'][i]\n", " print_colored(f\" Accuracy: {accuracy_3:<10.4f} \", bg_color=\"blue\" if accuracy_3 == max(accuracy, accuracy_2, accuracy_3) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score_3:<10.4f} \", bg_color=\"blue\" if f1_score_3 == max(f1_score, f1_score_2, f1_score_3) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision_3:<10.4f}\", bg_color=\"blue\" if precision_3 == max(precision, precision_2, precision_3) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_3:<10.4f}\\n\", \"black\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

In blu le metriche con il valore più alto
Ho provato a colorare di bianco il testo su background blu ma è poco leggibile in quanto non è bianco puro ma color panna.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "# Indici per le categorie\n", "x = range(len(categories))\n", "\n", "# Larghezza delle barre\n", "width = 0.3\n", "\n", "plt.figure(figsize=(12, 6))\n", "plt.bar(x, test_metrics_df['F1-Score'], width=width, label='Primo Modello', align='center', alpha=0.7)\n", "plt.bar([i + width for i in x], test_metrics_df_2['F1-Score'], width=width, label='Secondo Modello', align='center', alpha=0.7)\n", "plt.bar([i + 2 * width for i in x], test_metrics_df_3['F1-Score'], width=width, label='Terzo Modello', align='center', alpha=0.7)\n", "\n", "plt.xticks([i + width for i in x], categories, fontsize=14, rotation=45, ha='right', color=\"#b81414\")\n", "plt.yticks(fontsize=14, color=\"#b81414\")\n", "plt.xlabel(\"Categorie\", fontsize=16, color=\"black\")\n", "plt.ylabel('F1-Score', fontsize=16, color=\"black\")\n", "plt.title('Confronto F1-Score tra Modelli', fontsize=18)\n", "plt.legend()\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "# Indici per le categorie\n", "x = range(len(categories))\n", "\n", "# Larghezza delle barre\n", "width = 0.3\n", "\n", "plt.figure(figsize=(12, 6))\n", "plt.bar(x, test_metrics_df['Accuracy'], width=width, label='Primo Modello', align='center', alpha=0.7)\n", "plt.bar([i + width for i in x], test_metrics_df_2['Accuracy'], width=width, label='Secondo Modello', align='center', alpha=0.7)\n", "plt.bar([i + 2 * width for i in x], test_metrics_df_3['Accuracy'], width=width, label='Terzo Modello', align='center', alpha=0.7)\n", "\n", "plt.title('Confronto Accuracy tra Modelli', fontsize=18)\n", "plt.xticks([i + width for i in x], categories, fontsize=14, rotation=45, ha='right', color=\"#b81414\")\n", "plt.yticks(fontsize=14, color=\"#b81414\")\n", "plt.xlabel(\"Categorie\", fontsize=16, color=\"black\")\n", "plt.ylabel('Accuracy', fontsize=16, color=\"black\")\n", "plt.legend(loc='center left', bbox_to_anchor=(1, 0.5), fontsize=12)\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "# Indici per le categorie\n", "x = range(len(categories))\n", "\n", "# Larghezza delle barre\n", "width = 0.3\n", "\n", "plt.figure(figsize=(12, 6))\n", "plt.bar(x, test_metrics_df['Precision'], width=width, label='Primo Modello', align='center', alpha=0.7)\n", "plt.bar([i + width for i in x], test_metrics_df_2['Precision'], width=width, label='Secondo Modello', align='center', alpha=0.7)\n", "plt.bar([i + 2 * width for i in x], test_metrics_df_3['Precision'], width=width, label='Terzo Modello', align='center', alpha=0.7)\n", "\n", "plt.title('Confronto Precision tra Modelli', fontsize=18)\n", "plt.xticks([i + width for i in x], categories, fontsize=14, rotation=45, ha='right', color=\"#b81414\")\n", "plt.yticks(fontsize=14, color=\"#b81414\")\n", "plt.xlabel(\"Categorie\", fontsize=16, color=\"black\")\n", "plt.ylabel('Precision', fontsize=16, color=\"black\")\n", "\n", "plt.legend()\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "# Indici per le categorie\n", "x = range(len(categories))\n", "\n", "# Larghezza delle barre\n", "width = 0.3\n", "\n", "plt.figure(figsize=(12, 6))\n", "plt.bar(x, global_accuracy_1, width=width, label='Primo Modello', align='center', alpha=0.7)\n", "plt.bar([i + width for i in x], global_accuracy_2, width=width, label='Secondo Modello', align='center', alpha=0.7)\n", "plt.bar([i + 2 * width for i in x], global_accuracy_3, width=width, label='Terzo Modello', align='center', alpha=0.7)\n", "\n", "plt.title('Confronto Global accuracy tra Modelli', fontsize=18)\n", "plt.xticks([i + width for i in x], categories, fontsize=14, rotation=45, ha='right', color=\"#b81414\")\n", "plt.yticks(fontsize=14, color=\"#b81414\")\n", "plt.xlabel(\"Categorie\", fontsize=16, color=\"black\")\n", "plt.ylabel('Global Accuracy', fontsize=16, color=\"black\")\n", "\n", "plt.legend(loc='center left', bbox_to_anchor=(1, 0.5), fontsize=12)\n", "\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Dai risultati ottenuti si può notare che il primo modello anche essendo il più semplice ha performance migliori, di poco superiori al secondo ma decisamente migliori del terzo il quale anche avendo una struttura più complessa non riesce a catturare le dinamiche dei dati in modo efficace, risultando meno accurato e meno robusto nella classificazione

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Applico una funzione focal loss al primo modello per ridurre l'impatto delle classi di maggioranza sul calcolo della perdita.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Dà maggiore peso ai campioni che il modello fatica a classificare correttamente.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "load_model_path = '/home/gap/Scrivania/Filtro_anti_hater/anti_hater_model_stratify_400_128_6__1.keras' \n", "model_reload_1 = load_model(load_model_path)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "def focal_loss(alpha=0.25, gamma=2.0):\n", " def loss(y_true, y_pred):\n", " y_true = tf.cast(y_true, tf.float32)\n", " y_pred = tf.clip_by_value(y_pred, 1e-8, 1 - 1e-8)\n", " ce_loss = -y_true * tf.math.log(y_pred)\n", " focal = alpha * tf.pow(1 - y_pred, gamma) * ce_loss\n", " return tf.reduce_mean(tf.reduce_sum(focal, axis=-1))\n", " return loss\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model_reload_1.compile(\n", " optimizer='adam',\n", " loss=focal_loss(alpha=0.25, gamma=2.0),\n", " metrics=['accuracy']\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Model: \"sequential\"\n
\n
\n
\n
\n
\n
\n
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓\n┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃\n┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩\n│ embedding_layer (Embedding)     │ (None, 400, 128)       │    27,208,064 │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ bidirectional_lstm              │ (None, 128)            │        98,816 │\n│ (Bidirectional)                 │                        │               │\n├─────────────────────────────────┼────────────────────────┼───────────────┤\n│ output_layer (Dense)            │ (None, 6)              │           774 │\n└─────────────────────────────────┴────────────────────────┴───────────────┘\n
\n
\n
\n
\n
\n
\n
 Total params: 27,307,654 (104.17 MB)\n
\n
\n
\n
\n
\n
\n
 Trainable params: 27,307,654 (104.17 MB)\n
\n
\n
\n
\n
\n
\n
 Non-trainable params: 0 (0.00 B)\n
\n
\n
\n
\n
" ] } } ], "source": [ "model_reload_1.summary()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "reduce_lr = ReduceLROnPlateau(\n", " monitor='val_loss',\n", " factor=0.2,\n", " patience=2, \n", " min_lr=1e-6,\n", " verbose=1\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "early_stopping = EarlyStopping(\n", " monitor='val_loss', # metrica da monitorare\n", " patience=3, # numero di epoche da aspettare prima di fermarsi\n", " restore_best_weights=True, # ripristina i migliori pesi\n", " mode='min', # minimizzare la loss\n", " min_delta=0.001, # cambiamento minimo da considerare come miglioramento\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model_checkpoint_path = \"anti_hater_model_reload__1.keras\"\n", "\n", "model_checkpoint = ModelCheckpoint(\n", " filepath=model_checkpoint_path, # Percorso per salvare il modello\n", " monitor='val_loss', # Metrica da monitorare\n", " save_best_only=True, # Salvo solo il modello migliore\n", " save_weights_only=False, # Salvo l'intero modello (inclusa l'architettura)\n", " mode='min', # minimizzare la val_loss\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Epoch 1/10\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 144ms/step - accuracy: 0.6887 - loss: 0.0831\nEpoch 1: val_loss improved from inf to 0.29047, saving model to anti_hater_model_reload__1.keras\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 929s 147ms/step - accuracy: 0.6887 - loss: 0.0831 - val_accuracy: 0.9000 - val_loss: 0.2905 - learning_rate: 8.0000e-06\nEpoch 2/10\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 142ms/step - accuracy: 0.6953 - loss: 0.0815\nEpoch 2: val_loss did not improve from 0.29047\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 917s 145ms/step - accuracy: 0.6953 - loss: 0.0815 - val_accuracy: 0.8984 - val_loss: 0.2989 - learning_rate: 8.0000e-06\nEpoch 3/10\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 142ms/step - accuracy: 0.6923 - loss: 0.0820\nEpoch 3: ReduceLROnPlateau reducing learning rate to 1.6000001778593287e-06.\n\nEpoch 3: val_loss did not improve from 0.29047\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 914s 145ms/step - accuracy: 0.6923 - loss: 0.0820 - val_accuracy: 0.9004 - val_loss: 0.2987 - learning_rate: 8.0000e-06\nEpoch 4/10\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 0s 143ms/step - accuracy: 0.6902 - loss: 0.0812\nEpoch 4: val_loss did not improve from 0.29047\n6313/6313 ━━━━━━━━━━━━━━━━━━━━ 924s 146ms/step - accuracy: 0.6902 - loss: 0.0812 - val_accuracy: 0.8999 - val_loss: 0.3015 - learning_rate: 1.6000e-06\nEpoch 4: early stopping\nRestoring model weights from the end of the best epoch: 1.\nIl modello è stato salvato in: anti_hater_model_reload__1.keras\n
\n
\n
\n
\n
" ] } } ], "source": [ "history = model.fit(\n", " X_train_balanced,\n", " y_train_balanced,\n", " validation_data=(X_val_mlsmote, y_val_mlsmote), # Usa i dati di validazione originali\n", " epochs=10,\n", " batch_size=32,\n", " callbacks=[early_stopping, reduce_lr, model_checkpoint], \n", " verbose=1\n", ")\n", "\n", "print(f\"Il modello è stato salvato in: {model_checkpoint_path}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Valutazione del quarto modello (primo modello + focal loss) sul test set

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nValutazione sul Test Set:\n748/748 ━━━━━━━━━━━━━━━━━━━━ 20s 27ms/step\n\nMetriche del Test Set:\n        Category  Accuracy  F1-Score  Precision\n0          toxic  0.866227  0.510547   0.392849\n1   severe_toxic  0.893006  0.064304   0.035341\n2        obscene  0.882771  0.407767   0.278788\n3         threat  0.898438  0.006539   0.003378\n4         insult  0.874081  0.347619   0.232686\n5  identity_hate  0.891920  0.024142   0.013083\n\nInferenza su alcuni esempi del Test Set:\nCommento #1:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.14 0.   0.01 0.   0.02 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #2:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.02 0.   0.   0.   0.   0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #3:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.02 0.   0.01 0.   0.01 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #4:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.02 0.01 0.01 0.01 0.01 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #5:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #6:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.03 0.   0.   0.   0.01 0.  ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #7:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.03 0.03 0.03 0.03 0.03 0.03])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #8:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.26 0.24 0.29 0.22 0.19 0.21])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #9:\n - Predetto: [1 1 1 0 1 0] (probabilità: [1.   0.65 0.98 0.32 0.96 0.49])\n - Vero: [1 1 1 0 1 0]\n--------------------------------------------------\nCommento #10:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.15 0.14 0.12 0.21 0.16 0.16])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #11:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.05 0.06 0.05 0.05 0.05 0.05])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #12:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0. 0. 0. 0. 0. 0.])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #13:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.21 0.08 0.12 0.09 0.12 0.1 ])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #14:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.01 0.01 0.01 0.01 0.01 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\nCommento #15:\n - Predetto: [0 0 0 0 0 0] (probabilità: [0.07 0.01 0.02 0.01 0.02 0.01])\n - Vero: [0 0 0 0 0 0]\n--------------------------------------------------\n
\n
\n
\n
\n
" ] } } ], "source": [ "print(\"\\nValutazione sul Test Set:\")\n", "y_test_pred = model_reload_1.predict(X_test_mlsmote) # Predizioni sul test set\n", "y_test_pred_binary = (y_test_pred > 0.5).astype(int) # Converto probabilità in etichette binarie\n", "\n", "test_metrics_df_reload_1 = pd.DataFrame()\n", "test_metrics_df_reload_1['Category'] = categories\n", "\n", "test_accuracies = []\n", "test_f1_scores = []\n", "test_precisions = []\n", "\n", "for i in range(len(categories)):\n", " test_acc = accuracy_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_f1 = f1_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " test_prec = precision_score(y_test_mlsmote[:, i], y_test_pred_binary[:, i])\n", " \n", " test_accuracies.append(test_acc)\n", " test_f1_scores.append(test_f1)\n", " test_precisions.append(test_prec)\n", "\n", "test_metrics_df_reload_1['Accuracy'] = test_accuracies\n", "test_metrics_df_reload_1['F1-Score'] = test_f1_scores\n", "test_metrics_df_reload_1['Precision'] = test_precisions\n", "\n", "print(\"\\nMetriche del Test Set:\")\n", "print(test_metrics_df_reload_1)\n", "\n", "\n", "print(\"\\nInferenza su alcuni esempi del Test Set:\")\n", "\n", "num_examples = 15\n", "for idx in range(num_examples):\n", " comment = X_test_mlsmote[idx]\n", " true_labels = y_test_mlsmote[idx]\n", " predicted_probs = y_test_pred[idx]\n", " predicted_labels = y_test_pred_binary[idx]\n", " \n", " print(f\"Commento #{idx + 1}:\")\n", " print(f\" - Predetto: {predicted_labels} (probabilità: {predicted_probs.round(2)})\")\n", " print(f\" - Vero: {true_labels}\")\n", " print(\"-\" * 50)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Precisione Globale (Global Accuracy):\n0.8805\n
\n
\n
\n
\n
" ] } } ], "source": [ "global_hamming_loss = hamming_loss(y_test_mlsmote, y_test_pred_binary)\n", "\n", "global_accuracy_reload_1 = 1 - global_hamming_loss\n", "\n", "print_colored(f\"Precisione Globale (Global Accuracy):\", \"blue\")\n", "print(f\"{global_accuracy_reload_1:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per toxic:\n[[19064  2581]\n [  621  1670]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per severe_toxic:\n[[21287  2402]\n [  159    88]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per obscene:\n[[20164  2499]\n [  307   966]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per threat:\n[[21497  2360]\n [   71     8]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per insult:\n[[20119  2648]\n [  366   803]]\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
Confusion Matrix per identity_hate:\n[[21317  2414]\n [  173    32]]\n
\n
\n
\n
\n
" ] } } ], "source": [ "conf_matrices = multilabel_confusion_matrix(y_test_mlsmote, y_test_pred_binary)\n", "\n", "if conf_matrices.shape[0] != len(categories):\n", " print(\"Errore: Il numero di confusion matrix non corrisponde al numero di categorie.\")\n", "else:\n", " for i, category in enumerate(categories):\n", " plt.figure(figsize=(6, 4))\n", " \n", " conf_matrix = conf_matrices[i]\n", " \n", " labels = np.array([['TN', 'FP'], ['FN', 'TP']])\n", " annotated_matrix = np.empty_like(conf_matrix, dtype=object)\n", " for row in range(conf_matrix.shape[0]):\n", " for col in range(conf_matrix.shape[1]):\n", " annotated_matrix[row, col] = f\"{labels[row, col]}: {conf_matrix[row, col]}\"\n", " \n", " sns.heatmap(conf_matrix, annot=annotated_matrix, fmt='', cmap='Blues', cbar=False)\n", " plt.title(f\"Confusion Matrix per {category}\")\n", " plt.xlabel(\"Predizioni\")\n", " plt.ylabel(\"Valori Reali\")\n", " plt.show()\n", " \n", " print(f\"Confusion Matrix per {category}:\")\n", " print(conf_matrix)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " fpr, tpr, _ = roc_curve(y_test_mlsmote[:, i], y_test_pred[:, i])\n", " roc_auc = auc(fpr, tpr)\n", " \n", " plt.figure(figsize=(6, 4))\n", " plt.plot(fpr, tpr, label=f'AUC = {roc_auc:.2f}')\n", " plt.plot([0, 1], [0, 1], 'k--', label='Random Guessing')\n", " plt.title(f\"ROC Curve per {category}\")\n", " plt.xlabel('False Positive Rate')\n", " plt.ylabel('True Positive Rate')\n", " plt.legend(loc='lower right')\n", " plt.grid(alpha=0.3)\n", " plt.show()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
\nPrecision-Recall per toxic (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.10        1.00        \n0.00      0.11        1.00        \n0.01      0.12        0.99        \n0.01      0.14        0.98        \n0.03      0.17        0.96        \n0.07      0.20        0.93        \n0.16      0.25        0.87        \n0.37      0.34        0.78        \n0.75      0.53        0.61        \n1.00      1.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per severe_toxic (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.02        1.00        \n0.01      0.02        1.00        \n0.02      0.02        0.99        \n0.06      0.03        0.96        \n0.18      0.04        0.82        \n0.47      0.04        0.38        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per obscene (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.05        1.00        \n0.00      0.06        1.00        \n0.00      0.07        0.99        \n0.01      0.08        0.98        \n0.01      0.09        0.98        \n0.03      0.11        0.96        \n0.09      0.15        0.92        \n0.26      0.20        0.85        \n0.63      0.33        0.69        \n1.00      1.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per threat (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.00        1.00        \n0.00      0.00        0.99        \n0.01      0.01        0.96        \n0.02      0.01        0.89        \n0.06      0.01        0.73        \n0.17      0.01        0.43        \n0.45      0.00        0.13        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per insult (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.05        1.00        \n0.00      0.05        1.00        \n0.00      0.06        0.99        \n0.01      0.07        0.99        \n0.01      0.09        0.98        \n0.04      0.11        0.96        \n0.10      0.13        0.91        \n0.27      0.18        0.81        \n0.62      0.27        0.62        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
\n
\n
\nPrecision-Recall per identity_hate (valori campionati):\nThreshold Precision   Recall      \n----------------------------------\n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        1.00        \n0.00      0.01        0.99        \n0.01      0.02        0.98        \n0.02      0.02        0.91        \n0.07      0.02        0.82        \n0.19      0.02        0.57        \n0.47      0.01        0.18        \n1.00      0.00        0.00        \n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " precision, recall, thresholds = precision_recall_curve(y_test_mlsmote[:, i], y_test_pred[:, i])\n", " \n", " print(f\"\\nPrecision-Recall per {category} (valori campionati):\")\n", " print(f\"{'Threshold':<10}{'Precision':<12}{'Recall':<12}\")\n", " print(\"-\" * 34)\n", " \n", " sampled_indices = np.linspace(0, len(thresholds) - 1, 10, dtype=int) # Campiona 10 valori\n", " for idx in sampled_indices:\n", " print(f\"{thresholds[idx]:<10.2f}{precision[idx]:<12.2f}{recall[idx]:<12.2f}\")\n", " \n", " plt.figure(figsize=(6, 4))\n", " plt.plot(recall, precision, label='Precision-Recall Curve')\n", " plt.title(f\"Precision-Recall Curve per {category}\")\n", " plt.xlabel('Recall')\n", " plt.ylabel('Precision')\n", " plt.legend(loc='upper right')\n", " plt.grid(alpha=0.3)\n", " plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

I valori di val_accuracy, Auc Roc e Global Accuracy non sono bassi ma alcune categorie presentano valori bassissimi di F1-score e Precision. Il modello tende a classificare molti commenti innocui come tossici.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Confronto tra modelli

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Categoria toxic\n   Primo     Modello:   Accuracy: 0.5272        F1-Score: 0.5272        Precision: 0.4039       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8695        F1-Score: 0.5242        Precision: 0.4026       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.2909        F1-Score: 0.1494        Precision: 0.0844       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8662        F1-Score: 0.5105        Precision: 0.3928       Global Accuracy: 0.8844    \n\nCategoria severe_toxic\n   Primo     Modello:   Accuracy: 0.0634        F1-Score: 0.0634        Precision: 0.0350       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8907        F1-Score: 0.0697        Precision: 0.0382       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.6331        F1-Score: 0.0247        Precision: 0.0127       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8930        F1-Score: 0.0643        Precision: 0.0353       Global Accuracy: 0.8844    \n\nCategoria obscene\n   Primo     Modello:   Accuracy: 0.4232        F1-Score: 0.4232        Precision: 0.2911       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8844        F1-Score: 0.4077        Precision: 0.2802       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.8436        F1-Score: 0.2306        Precision: 0.1561       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8828        F1-Score: 0.4078        Precision: 0.2788       Global Accuracy: 0.8844    \n\nCategoria threat\n   Primo     Modello:   Accuracy: 0.0076        F1-Score: 0.0076        Precision: 0.0040       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8959        F1-Score: 0.0080        Precision: 0.0041       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.0954        F1-Score: 0.0072        Precision: 0.0036       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8984        F1-Score: 0.0065        Precision: 0.0034       Global Accuracy: 0.8844    \n\nCategoria insult\n   Primo     Modello:   Accuracy: 0.3581        F1-Score: 0.3581        Precision: 0.2420       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8764        F1-Score: 0.3467        Precision: 0.2336       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.9159        F1-Score: 0.0214        Precision: 0.0247       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8741        F1-Score: 0.3476        Precision: 0.2327       Global Accuracy: 0.8844    \n\nCategoria identity_hate\n   Primo     Modello:   Accuracy: 0.0276        F1-Score: 0.0276        Precision: 0.0150       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8915        F1-Score: 0.0226        Precision: 0.0122       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.3409        F1-Score: 0.0105        Precision: 0.0053       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8919        F1-Score: 0.0241        Precision: 0.0131       Global Accuracy: 0.8844    \n\n
\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " print_colored(f\"Categoria {category}\\n\".ljust(16), \"blue\", end=\"\")\n", " \n", " # Primo Modello\n", " print_colored(f\" Primo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy = test_metrics_df['Accuracy'][i]\n", " f1_score = test_metrics_df['F1-Score'][i]\n", " precision = test_metrics_df['Precision'][i]\n", " print_colored(f\" Accuracy: {f1_score:<10.4f} \", bg_color=\"blue\" if accuracy == max(accuracy, test_metrics_df_2['Accuracy'][i], test_metrics_df_3['Accuracy'][i], test_metrics_df_reload_1['Accuracy'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score:<10.4f} \", bg_color=\"blue\" if f1_score == max(f1_score, test_metrics_df_2['F1-Score'][i], test_metrics_df_3['F1-Score'][i], test_metrics_df_reload_1['F1-Score'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision:<10.4f}\", bg_color=\"blue\" if precision == max(precision, test_metrics_df_2['Precision'][i], test_metrics_df_3['Precision'][i], test_metrics_df_reload_1['Precision'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_1:<10.4f}\", \"black\")\n", "\n", " # Secondo Modello\n", " print_colored(f\" Secondo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_2 = test_metrics_df_2['Accuracy'][i]\n", " f1_score_2 = test_metrics_df_2['F1-Score'][i]\n", " precision_2 = test_metrics_df_2['Precision'][i]\n", " print_colored(f\" Accuracy: {accuracy_2:<10.4f} \", bg_color=\"blue\" if accuracy_2 == max(accuracy, accuracy_2, test_metrics_df_3['Accuracy'][i], test_metrics_df_reload_1['Accuracy'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score_2:<10.4f} \", bg_color=\"blue\" if f1_score_2 == max(f1_score, f1_score_2, test_metrics_df_3['F1-Score'][i], test_metrics_df_reload_1['F1-Score'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision_2:<10.4f}\", bg_color=\"blue\" if precision_2 == max(precision, precision_2, test_metrics_df_3['Precision'][i], test_metrics_df_reload_1['Precision'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_1:<10.4f}\", \"black\")\n", "\n", " # Terzo Modello\n", " print_colored(f\" Terzo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_3 = test_metrics_df_3['Accuracy'][i]\n", " f1_score_3 = test_metrics_df_3['F1-Score'][i]\n", " precision_3 = test_metrics_df_3['Precision'][i]\n", " print_colored(f\" Accuracy: {accuracy_3:<10.4f} \", bg_color=\"blue\" if accuracy_3 == max(accuracy, accuracy_2, accuracy_3, test_metrics_df_reload_1['Accuracy'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score_3:<10.4f} \", bg_color=\"blue\" if f1_score_3 == max(f1_score, f1_score_2, f1_score_3, test_metrics_df_reload_1['F1-Score'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision_3:<10.4f}\", bg_color=\"blue\" if precision_3 == max(precision, precision_2, precision_3, test_metrics_df_reload_1['Precision'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_3:<10.4f}\", \"black\")\n", "\n", " # Quarto Modello\n", " print_colored(f\" Quarto Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_reload_1 = test_metrics_df_reload_1['Accuracy'][i]\n", " f1_score_reload_1 = test_metrics_df_reload_1['F1-Score'][i]\n", " precision_reload_1 = test_metrics_df_reload_1['Precision'][i]\n", " print_colored(f\" Accuracy: {accuracy_reload_1:<10.4f} \", bg_color=\"blue\" if accuracy_reload_1 == max(accuracy, accuracy_2, accuracy_3, test_metrics_df_reload_1['Accuracy'][i]) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" F1-Score: {f1_score_reload_1:<10.4f} \", bg_color=\"blue\" if f1_score_reload_1 == max(f1_score, f1_score_2, f1_score_3, f1_score_reload_1) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Precision: {precision_reload_1:<10.4f}\", bg_color=\"blue\" if precision_reload_1 == max(precision, precision_2, precision_3, precision_reload_1) else \"\", color=\"black\", end=\"\")\n", " print_colored(f\" Global Accuracy: {global_accuracy_reload_1:<10.4f}\\n\", \"black\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Stesso confronto, ho usato il testo bianco su sfondo blu ma il testo è poco leggibile.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Categoria toxic\n   Primo     Modello:   Accuracy: 0.8697        F1-Score: 0.5272        Precision: 0.4039       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8695        F1-Score: 0.5242        Precision: 0.4026       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.2909        F1-Score: 0.1494        Precision: 0.0844       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8662        F1-Score: 0.5105        Precision: 0.3928       Global Accuracy: 0.8844    \n\nCategoria severe_toxic\n   Primo     Modello:   Accuracy: 0.8975        F1-Score: 0.0634        Precision: 0.0350       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8907        F1-Score: 0.0697        Precision: 0.0382       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.6331        F1-Score: 0.0247        Precision: 0.0127       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8930        F1-Score: 0.0643        Precision: 0.0353       Global Accuracy: 0.8844    \n\nCategoria obscene\n   Primo     Modello:   Accuracy: 0.8877        F1-Score: 0.4232        Precision: 0.2911       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8844        F1-Score: 0.4077        Precision: 0.2802       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.8436        F1-Score: 0.2306        Precision: 0.1561       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8828        F1-Score: 0.4078        Precision: 0.2788       Global Accuracy: 0.8844    \n\nCategoria threat\n   Primo     Modello:   Accuracy: 0.9023        F1-Score: 0.0076        Precision: 0.0040       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8959        F1-Score: 0.0080        Precision: 0.0041       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.0954        F1-Score: 0.0072        Precision: 0.0036       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8984        F1-Score: 0.0065        Precision: 0.0034       Global Accuracy: 0.8844    \n\nCategoria insult\n   Primo     Modello:   Accuracy: 0.8794        F1-Score: 0.3581        Precision: 0.2420       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8764        F1-Score: 0.3467        Precision: 0.2336       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.9159        F1-Score: 0.0214        Precision: 0.0247       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8741        F1-Score: 0.3476        Precision: 0.2327       Global Accuracy: 0.8844    \n\nCategoria identity_hate\n   Primo     Modello:   Accuracy: 0.8969        F1-Score: 0.0276        Precision: 0.0150       Global Accuracy: 0.8889    \n   Secondo   Modello:   Accuracy: 0.8915        F1-Score: 0.0226        Precision: 0.0122       Global Accuracy: 0.8889    \n   Terzo     Modello:   Accuracy: 0.3409        F1-Score: 0.0105        Precision: 0.0053       Global Accuracy: 0.8889    \n   Quarto    Modello:   Accuracy: 0.8919        F1-Score: 0.0241        Precision: 0.0131       Global Accuracy: 0.8844    \n\n
\n
\n
\n
\n
" ] } } ], "source": [ "for i, category in enumerate(categories):\n", " print_colored(f\"Categoria {category}\\n\".ljust(16), \"blue\", end=\"\")\n", " \n", " # Primo Modello\n", " print_colored(f\" Primo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy = test_metrics_df['Accuracy'][i]\n", " f1_score = test_metrics_df['F1-Score'][i]\n", " precision = test_metrics_df['Precision'][i]\n", " \n", " print_colored(\n", " f\" Accuracy: {accuracy:<10.4f} \",\n", " bg_color=\"blue\" if accuracy == max(accuracy, test_metrics_df_2['Accuracy'][i], test_metrics_df_3['Accuracy'][i], test_metrics_df_reload_1['Accuracy'][i]) else \"\",\n", " color=\"white\" if accuracy == max(accuracy, test_metrics_df_2['Accuracy'][i], test_metrics_df_3['Accuracy'][i], test_metrics_df_reload_1['Accuracy'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" F1-Score: {f1_score:<10.4f} \",\n", " bg_color=\"blue\" if f1_score == max(f1_score, test_metrics_df_2['F1-Score'][i], test_metrics_df_3['F1-Score'][i], test_metrics_df_reload_1['F1-Score'][i]) else \"\",\n", " color=\"white\" if f1_score == max(f1_score, test_metrics_df_2['F1-Score'][i], test_metrics_df_3['F1-Score'][i], test_metrics_df_reload_1['F1-Score'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" Precision: {precision:<10.4f}\",\n", " bg_color=\"blue\" if precision == max(precision, test_metrics_df_2['Precision'][i], test_metrics_df_3['Precision'][i], test_metrics_df_reload_1['Precision'][i]) else \"\",\n", " color=\"white\" if precision == max(precision, test_metrics_df_2['Precision'][i], test_metrics_df_3['Precision'][i], test_metrics_df_reload_1['Precision'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(f\" Global Accuracy: {global_accuracy_1:<10.4f}\", \"black\")\n", "\n", " \n", " \n", " \n", " \n", " \n", " # Secondo Modello\n", " print_colored(f\" Secondo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_2 = test_metrics_df_2['Accuracy'][i]\n", " f1_score_2 = test_metrics_df_2['F1-Score'][i]\n", " precision_2 = test_metrics_df_2['Precision'][i]\n", " \n", " print_colored(\n", " f\" Accuracy: {accuracy_2:<10.4f} \",\n", " bg_color=\"blue\" if accuracy_2 == max(accuracy, accuracy_2, test_metrics_df_3['Accuracy'][i], test_metrics_df_reload_1['Accuracy'][i]) else \"\",\n", " color=\"white\" if accuracy_2 == max(accuracy, accuracy_2, test_metrics_df_3['Accuracy'][i], test_metrics_df_reload_1['Accuracy'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" F1-Score: {f1_score_2:<10.4f} \",\n", " bg_color=\"blue\" if f1_score_2 == max(f1_score, f1_score_2, test_metrics_df_3['F1-Score'][i], test_metrics_df_reload_1['F1-Score'][i]) else \"\",\n", " color=\"white\" if f1_score_2 == max(f1_score, f1_score_2, test_metrics_df_3['F1-Score'][i], test_metrics_df_reload_1['F1-Score'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" Precision: {precision_2:<10.4f}\",\n", " bg_color=\"blue\" if precision_2 == max(precision, precision_2, test_metrics_df_3['Precision'][i], test_metrics_df_reload_1['Precision'][i]) else \"\",\n", " color=\"white\" if precision_2 == max(precision, precision_2, test_metrics_df_3['Precision'][i], test_metrics_df_reload_1['Precision'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(f\" Global Accuracy: {global_accuracy_1:<10.4f}\", \"black\")\n", "\n", " \n", " \n", " \n", " \n", " \n", " # Terzo Modello\n", " print_colored(f\" Terzo Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_3 = test_metrics_df_3['Accuracy'][i]\n", " f1_score_3 = test_metrics_df_3['F1-Score'][i]\n", " precision_3 = test_metrics_df_3['Precision'][i]\n", " \n", " print_colored(\n", " f\" Accuracy: {accuracy_3:<10.4f} \",\n", " bg_color=\"blue\" if accuracy_3 == max(accuracy, accuracy_2, accuracy_3, test_metrics_df_reload_1['Accuracy'][i]) else \"\",\n", " color=\"white\" if accuracy_3 == max(accuracy, accuracy_2, accuracy_3, test_metrics_df_reload_1['Accuracy'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" F1-Score: {f1_score_3:<10.4f} \",\n", " bg_color=\"blue\" if f1_score_3 == max(f1_score, f1_score_2, f1_score_3, test_metrics_df_reload_1['F1-Score'][i]) else \"\",\n", " color=\"white\" if f1_score_3 == max(f1_score, f1_score_2, f1_score_3, test_metrics_df_reload_1['F1-Score'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" Precision: {precision_3:<10.4f}\",\n", " bg_color=\"blue\" if precision_3 == max(precision, precision_2, precision_3, test_metrics_df_reload_1['Precision'][i]) else \"\",\n", " color=\"white\" if precision_3 == max(precision, precision_2, precision_3, test_metrics_df_reload_1['Precision'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(f\" Global Accuracy: {global_accuracy_3:<10.4f}\", \"black\")\n", "\n", " \n", " \n", " \n", " \n", " \n", " # Quarto Modello\n", " print_colored(f\" Quarto Modello:\".ljust(10), \"red\", end=\"\")\n", " accuracy_reload_1 = test_metrics_df_reload_1['Accuracy'][i]\n", " f1_score_reload_1 = test_metrics_df_reload_1['F1-Score'][i]\n", " precision_reload_1 = test_metrics_df_reload_1['Precision'][i]\n", " \n", " print_colored(\n", " f\" Accuracy: {accuracy_reload_1:<10.4f} \",\n", " bg_color=\"blue\" if accuracy_reload_1 == max(accuracy, accuracy_2, accuracy_3, test_metrics_df_reload_1['Accuracy'][i]) else \"\",\n", " color=\"white\" if accuracy_reload_1 == max(accuracy, accuracy_2, accuracy_3, test_metrics_df_reload_1['Accuracy'][i]) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" F1-Score: {f1_score_reload_1:<10.4f} \",\n", " bg_color=\"blue\" if f1_score_reload_1 == max(f1_score, f1_score_2, f1_score_3, f1_score_reload_1) else \"\",\n", " color=\"white\" if f1_score_reload_1 == max(f1_score, f1_score_2, f1_score_3, f1_score_reload_1) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(\n", " f\" Precision: {precision_reload_1:<10.4f}\",\n", " bg_color=\"blue\" if precision_reload_1 == max(precision, precision_2, precision_3, precision_reload_1) else \"\",\n", " color=\"white\" if precision_reload_1 == max(precision, precision_2, precision_3, precision_reload_1) else \"black\",\n", " end=\"\"\n", " )\n", " print_colored(f\" Global Accuracy: {global_accuracy_reload_1:<10.4f}\\n\", \"black\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Analizzo alcuni commenti creati da me e altri raccolti dal web utilizzando il primo modello, con l'obiettivo di valutare le sue capacità predittive.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Carico il modello salvato\n", "model = tf.keras.models.load_model('anti_hater_model_stratify_400_128_6__1.keras')\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Carico il tokenizer\n", "with open('tokenizer.pickle', 'rb') as handle:\n", " tokenizer = pickle.load(handle)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "def analyze_comment(comment, max_length=400):\n", " # Pre-processo il commento\n", " tokenized_sequence = tokenizer.texts_to_sequences([comment]) \n", " padded_sequence = pad_sequences(tokenized_sequence, maxlen=max_length, padding='post') \n", "\n", " prediction = model.predict(padded_sequence)\n", "\n", " results = {category: prob for category, prob in zip(categories, prediction[0])}\n", "\n", " return results\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"You are a stupid person!\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 63ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.9486\nsevere_toxic: 0.7094\nobscene: 0.8833\nthreat: 0.6735\ninsult: 0.8477\nidentity_hate: 0.7213\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"Your content on Instagram sucks, if I catch you on the street I will insult you\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 62ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.9503\nsevere_toxic: 0.9584\nobscene: 0.9531\nthreat: 0.9617\ninsult: 0.9590\nidentity_hate: 0.9627\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"You are beautiful\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 22ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.1012\nsevere_toxic: 0.0941\nobscene: 0.0793\nthreat: 0.0819\ninsult: 0.0853\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"\"\"Well done, if you combine your passion for images with knowledge of programming and data science, \n", "you could develop highly sought-after skills. \"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 26ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.0174\nsevere_toxic: 0.0163\nobscene: 0.0150\nthreat: 0.0144\ninsult: 0.0156\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"\"\"Let's look at an example that shows how quantum computers can succeed where classical computers fail:\n", "A classical computer might be great for difficult tasks like sorting a large database of molecules. \n", "But it struggles to solve more complex problems, such as simulating the behavior of those molecules.Today, \n", "if scientists want to know how a molecule behaves, they must synthesize it and experiment with it in the real world. \n", "If they want to know how a small change affects its behavior, they usually have to synthesize the new version and \n", "repeat the experiment. This is a costly and time-consuming process that impedes progress in fields as diverse as \n", "medicine and semiconductor design.\"\"\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 25ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.0706\nsevere_toxic: 0.0356\nobscene: 0.0472\nthreat: 0.0407\ninsult: 0.0502\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"You're ridiculous\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 77ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.0998\nsevere_toxic: 0.0634\nobscene: 0.0669\nthreat: 0.0623\ninsult: 0.0718\nidentity_hate: 0.0623\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"Son of a bitch\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 65ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.9887\nsevere_toxic: 0.5457\nobscene: 0.9543\nthreat: 0.3647\ninsult: 0.8988\nidentity_hate: 0.4566\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"You ugly nigger bastard\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 69ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.9925\nsevere_toxic: 0.3588\nobscene: 0.9485\nthreat: 0.1647\ninsult: 0.8878\nidentity_hate: 0.2600\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"I'm busy with work and I can't travel, but thanks to your travels I feel like I'm there, you make me dream.\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 79ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.1587\nsevere_toxic: 0.0870\nobscene: 0.1056\nthreat: 0.0879\ninsult: 0.1155\nidentity_hate: 0.0950\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"\"\"I organized this trip following the advice in your article, thanks for the wonderful places you made me \n", "discover.!!!!!!!!\"\"\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 67ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.3077\nsevere_toxic: 0.2967\nobscene: 0.2847\nthreat: 0.3261\ninsult: 0.2862\nidentity_hate: 0.3086\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"\"\"With that look you are obscene, you should work in a red light TV instead of making travel videos, \n", "shame on you.!!!!!!!!\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 58ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.1680\nsevere_toxic: 0.0843\nobscene: 0.0988\nthreat: 0.0835\ninsult: 0.1119\nidentity_hate: 0.0916\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"Don't write nonsense, go to work and get a serious job!\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 54ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.3813\nsevere_toxic: 0.1116\nobscene: 0.1711\nthreat: 0.1141\ninsult: 0.1754\nidentity_hate: 0.1152\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"Don't write nonsense, go to work and get a serious job bitch\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 68ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.9757\nsevere_toxic: 0.2428\nobscene: 0.8726\nthreat: 0.1480\ninsult: 0.7902\nidentity_hate: 0.2171\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"You are a great person and you bring really interesting and wonderful content. Thanks for what you do\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 63ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.1906\nsevere_toxic: 0.1337\nobscene: 0.1332\nthreat: 0.1368\ninsult: 0.1540\nidentity_hate: 0.1330\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"Next time I meet you on the street I will stab you\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 27ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.5969\nsevere_toxic: 0.6090\nobscene: 0.5700\nthreat: 0.6031\ninsult: 0.5946\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"Next time I meet you on the street I'll kill you\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 20ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.7290\nsevere_toxic: 0.6938\nobscene: 0.6725\nthreat: 0.6865\ninsult: 0.6866\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# Esempio di utilizzo\n", "comment = \"I want to offer you a coffee\"\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 20ms/step\n
\n
\n
\n
\n
" ] } } ], "source": [ "results = analyze_comment(comment)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n
Analisi del commento:\ntoxic: 0.1705\nsevere_toxic: 0.1416\nobscene: 0.1421\nthreat: 0.1484\ninsult: 0.1484\n
\n
\n
\n
\n
" ] } } ], "source": [ "print_colored(\"Analisi del commento:\", \"blue\")\n", "for category, prob in results.items():\n", " print(f\"{category}: {prob:.4f}\")\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Spiegazione generale di una rete neurale

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

1. Abbiamo un input

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

L'input è costituito da un vettore (x) di dati numerici (ad esempio, pixel di un'immagine, valori numerici di una tabella, ecc.).
Questo input viene normalizzato (scalato tra 0 e 1) per rendere l'addestramento più stabile. I dati vengono analizzati e trasferiti al primo strato nascosto.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

2 . Passaggio dell'input ai neuroni del primo strato nascosto

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Ogni neurone del primo strato nascosto riceve tutti i valori dell'input.
Per ciascun valore dell'input xi il neurone lo moltiplica per un peso associato wi(valore inizialmente casuale).
Il neurone calcola quindi una somma ponderata di tutti questi valori:

z=w1x1 + w2x2 + ... + wnxn + b

b è il bias, che è un valore costante inizialmente casuale aggiunto per aumentare la flessibilità del modello. Esso permette alla rete di rappresentare relazioni più complesse, spostando la somma ponderata lontano dallo zero.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

3. Passaggio dell'informazione al primo strato nascosto

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

L'output di ciascun neurone del primo strato (z) viene attivato attraverso una funzione di attivazione a=f(z).
La funzione di attivazione introduce non linearità permettendo alla rete di apprendere relazioni più complesse come curve, superfici o confini non lineari tra classi ed è fondamentale nelle reti neurali in quanto un modello lineare z=w1x1 + w2x2 + ... + wnxn è semplice e può separare solo dati che siano divisibili da una linea retta, quindi si possono trovare solo a destra o a sinistra di questa retta. Se i dati invece non sono separabili linearmente (ad esempio, due classi che formano un cerchio una dentro l’altra), un modello lineare fallirà.
Le funzioni di attivazione trasformano l'output lineare di un neurone z=wx+b in una forma non lineare.
Senza funzioni di attivazione non lineari una rete neurale è equivalente a una singola trasformazione lineare anche se ha più strati perdendo la capacità di apprendere relazioni non lineari nei dati. Questo la rende equivalente a un singolo percettrone limitandone drasticamente le capacità di apprendimento e rendendola inadatta a risolvere problemi complessi.
Le funzioni di attivazione più usate sono la Step function, la Tanh, la Relu, Leaky Relu, Elu e Swish per gli strati nascosti, Sigmoide e Softmax nello strato finale.

\n
" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [ { "output_type": "display_data", "metadata": {}, "data": { "text/html": [ "
\n
\n
\n
\n
\n
\n
\n\"No\n
\n
\n
\n
" ] } } ], "source": [ "fig, ax = plt.subplots(figsize=(12, 12))\n", "\n", "# Colori pastello\n", "colors = {\n", " \"yellow\": \"#f8f32b\", \n", " \"purple\": \"#899ad5\", \n", " \"blu\": \"#00aaff\" \n", "}\n", "\n", "# Posizioni dei cerchi\n", "nodes = {\n", " 'x1': (1, 5),\n", " 'x2': (1, 3),\n", " 'xn': (1, 1),\n", " 'Neurone 1': (4, 5),\n", " 'Neurone 2': (4, 3),\n", " 'Neurone 3': (4, 1),\n", " 'Neurone 4': (7, 5),\n", " 'Neurone 5': (7, 3),\n", " 'Neurone 6': (7, 1),\n", " 'y': (10, 3)\n", "}\n", "\n", "# Nodi\n", "for label, (x, y) in nodes.items():\n", " if \"Neurone\" in label: # Colorare i neuroni da 1 a 6 in rosso pastello\n", " color = colors[\"purple\"]\n", " elif label == 'y': # Colorare il nodo di output in verde pastello\n", " color = colors[\"blu\"]\n", " else: # Colorare gli input in giallo pastello\n", " color = colors[\"yellow\"]\n", " circle = plt.Circle((x, y), 0.5, color=color, zorder=2)\n", " ax.add_patch(circle)\n", " ax.text(x, y, label, fontsize=10, ha='center', va='center', zorder=3)\n", "\n", "# Tre pallini sotto il nodo x2\n", "dots_x2 = [(1, 2.2), (1, 1.95), (1, 1.7)] # Posizioni dei pallini\n", "for x, y in dots_x2:\n", " circle = plt.Circle((x, y), 0.05, color='black', zorder=2) \n", " ax.add_patch(circle)\n", "\n", "connections = [\n", " ('x1', 'Neurone 1'), ('x1', 'Neurone 2'), ('x1', 'Neurone 3'),\n", " ('x2', 'Neurone 1'), ('x2', 'Neurone 2'), ('x2', 'Neurone 3'),\n", " ('xn', 'Neurone 1'), ('xn', 'Neurone 2'), ('xn', 'Neurone 3'),\n", " ('Neurone 1', 'Neurone 4'), ('Neurone 1', 'Neurone 5'), ('Neurone 1', 'Neurone 6'),\n", " ('Neurone 2', 'Neurone 4'), ('Neurone 2', 'Neurone 5'), ('Neurone 2', 'Neurone 6'),\n", " ('Neurone 3', 'Neurone 4'), ('Neurone 3', 'Neurone 5'), ('Neurone 3', 'Neurone 6'),\n", " ('Neurone 4', 'y'), ('Neurone 5', 'y'), ('Neurone 6', 'y')\n", "]\n", "weights = {\n", " ('x1', 'Neurone 1'): 'w1', ('x1', 'Neurone 2'): 'w2', ('x1', 'Neurone 3'): 'w3',\n", " ('x2', 'Neurone 1'): 'w4', ('x2', 'Neurone 2'): 'w5', ('x2', 'Neurone 3'): 'w6',\n", " ('xn', 'Neurone 1'): 'w7', ('xn', 'Neurone 2'): 'w8', ('xn', 'Neurone 3'): 'w9',\n", " ('Neurone 1', 'Neurone 4'): 'w10', ('Neurone 1', 'Neurone 5'): 'w11', ('Neurone 1', 'Neurone 6'): 'w12',\n", " ('Neurone 2', 'Neurone 4'): 'w13', ('Neurone 2', 'Neurone 5'): 'w14', ('Neurone 2', 'Neurone 6'): 'w15',\n", " ('Neurone 3', 'Neurone 4'): 'w16', ('Neurone 3', 'Neurone 5'): 'w17', ('Neurone 3', 'Neurone 6'): 'w18',\n", " ('Neurone 4', 'y'): 'w19', ('Neurone 5', 'y'): 'w20', ('Neurone 6', 'y'): 'w21'\n", "}\n", "\n", "for start, end in connections:\n", " x_start, y_start = nodes[start]\n", " x_end, y_end = nodes[end]\n", " \n", " # Freccia\n", " ax.annotate(\n", " '', xy=(x_end, y_end), xytext=(x_start, y_start),\n", " arrowprops=dict(arrowstyle=\"->\", color='black', lw=1.5), zorder=1\n", " )\n", " \n", " # Angolo di inclinazione della linea\n", " dx = x_end - x_start\n", " dy = y_end - y_start\n", " angle = np.degrees(np.arctan2(dy, dx))\n", " \n", " # Weights\n", " if (start, end) in weights:\n", " label_x = x_start + dx * 0.68 \n", " label_y = y_start + dy * 0.70\n", " ax.text(label_x, label_y, weights[(start, end)], fontsize=14, color='black', \n", " rotation=angle+1, rotation_mode='anchor') \n", "\n", "ax.set_xlim(0, 11)\n", "ax.set_ylim(0, 6)\n", "ax.set_aspect('equal')\n", "ax.axis('off')\n", "\n", "plt.show()\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

4. Loss function e funzione di costo

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Lo strato finale produce un valore predetto y pred basato sull'input iniziale e sul flusso attraverso la rete e viene confrontato con y true che è il valore corretto per quell'input e viene fornito dal dataset di addestramento, ad esempio, se stiamo classificando immagini di gatti e cani, y true sarà \"1\" per i gatti e \"0\" per i cani, questo perchè in un dataset ogni input ha un'etichetta associata.
La loss function e la funzione di costo quantificano quindi la differenza tra un valore previsto, ovvero l'output del modello per un dato input e il valore effettivo. Se le previsioni di un modello sono accurate la perdita è piccola, se le loro previsioni sono imprecise e quindi c'è parecchia differenza tra y pred e y true la perdita è grande.
C'è una sottile differenza però tra le due funzioni ed è dovuta al fatto che la loss function si riferisce all'errore di un esempio di addestramento, mentre una funzione di costo calcola l'errore medio in un intero set di addestramento.

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

5. Backpropagation

\n
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "
\n

Durante la fase di backpropagation, procedendo a ritroso dall' output agli input, si calcola il gradiente dell'errore rispetto agli output di ogni strato nello stesso modo in cui viene calcolato per l'ultimo strato sempre utilizzando la funzione di costo. Una volta ottenuti i gradienti per ogni strato si calcolano i gradienti dei pesi e dei bias i quali vengono utlizzati da un algoritmo di ottimizzazione, ad esempio Adam, per aggiornare proprio i parametri della rete neurale. Dopo questi aggiornamenti inizia una nuova epoca di addestramento.

\n
" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }