Update app.py

app.py

@@ -1,79 +1,32 @@
-import pandas as pd
-from sklearn.model_selection import train_test_split
 from transformers import BertTokenizer, BertForSequenceClassification, Trainer, TrainingArguments
 from datasets import Dataset
 import torch
-from sklearn.metrics import accuracy_score
+from sklearn.model_selection import train_test_split
+import pandas as pd
 
-# Load the CSV data
+# Load data (use your own dataset CSV here)
 df = pd.read_csv("dishTV_customer_service_with_address_and_rules_accurate_v2.csv")
-
-# Clean the dataset by dropping rows with NaN values in important columns
 df = df.dropna(subset=['Agent Utterance', 'Customer Utterance', 'Category', 'Rule Followed', 'Question Asked', 'Question Answered'])
-
-# Merge Agent and Customer Utterances into a single conversation text
 df['Conversation'] = df['Agent Utterance'] + " " + df['Customer Utterance']
-
-# Define mappings for categories and labels
-category_mapping = {
+df['Category'] = df['Category'].map({
     'Greeting': 0,
     'Addressing Issue': 1,
     'Feedback': 2,
     'Resolution': 3,
     'Address': 4
-}
-
-# Map categories to numeric labels
-df['Category'] = df['Category'].map(category_mapping)
-
-# Rule validation functions to check whether each rule was followed by the agent and whether the customer answered
-def validate_rules(row):
-    missed_rules = []
-    missed_answers = []
-
-    # Rule checks for the agent
-    if 'hello' not in row['Agent Utterance'].lower() and 'hi' not in row['Agent Utterance'].lower():
-        missed_rules.append('Greeting')
-    
-    if 'address' not in row['Agent Utterance'].lower():
-        missed_rules.append('Address')
-    
-    if 'feedback' not in row['Agent Utterance'].lower():
-        missed_rules.append('Feedback')
-    
-    if 'resolved' not in row['Agent Utterance'].lower() and 'fix' not in row['Agent Utterance'].lower():
-        missed_rules.append('Resolution')
-
-    # Check if customer answered relevant questions
-    if 'address' in row['Agent Utterance'].lower() and ('address' not in row['Customer Utterance'].lower()):
-        missed_answers.append('Customer Address Answer')
-    
-    if 'feedback' in row['Agent Utterance'].lower() and ('yes' not in row['Customer Utterance'].lower() and 'no' not in row['Customer Utterance'].lower()):
-        missed_answers.append('Customer Feedback Answer')
-
-    # Returning the result as compliant or non-compliant
-    if len(missed_rules) == 0 and len(missed_answers) == 0:
-        return 1, []  # Compliant
-    else:
-        return 0, missed_rules + missed_answers  # Non-Compliant
-
-# Apply the rule validation to each row
-df[['Compliant', 'Missed Rules/Answers']] = df.apply(lambda row: pd.Series(validate_rules(row)), axis=1)
-
-# Splitting the data into training and validation datasets
+})
+
+# Split data
 train_texts, val_texts, train_labels, val_labels = train_test_split(df['Conversation'].tolist(), df['Compliant'].tolist(), test_size=0.2)
 
-# Load pre-trained BERT tokenizer
+# Tokenizer
 tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
 
-# Tokenize the input data
-def tokenize_function(examples):
-    return tokenizer(examples, padding="max_length", truncation=True, max_length=512)
-
+# Tokenize the inputs
 train_encodings = tokenizer(train_texts, truncation=True, padding=True, max_length=512)
 val_encodings = tokenizer(val_texts, truncation=True, padding=True, max_length=512)
 
-# Create Dataset objects for PyTorch
+# Create PyTorch datasets
 train_dataset = Dataset.from_dict({
     'input_ids': train_encodings['input_ids'],
     'attention_mask': train_encodings['attention_mask'],
@@ -86,46 +39,56 @@ val_dataset = Dataset.from_dict({
     'labels': val_labels
 })
 
-# Load pre-trained BERT model
-model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)  # Binary classification (Compliant vs Non-Compliant)
+# Check dataset sizes
+print(f"Training dataset size: {len(train_dataset)}")
+print(f"Validation dataset size: {len(val_dataset)}")
+
+# Model
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
 
-# Define compute_metrics function for evaluation
-def compute_metrics(p):
-    predictions, labels = p
-    predictions = torch.argmax(predictions, axis=-1)
-    return {'accuracy': accuracy_score(labels, predictions)}
+# Calculate steps per epoch
+steps_per_epoch = len(train_dataset) // 8  # Assuming batch size = 8
+num_train_epochs = 3  # Desired number of epochs
+max_steps = steps_per_epoch * num_train_epochs
 
-# Define training arguments for the Trainer
+# Define training arguments
 training_args = TrainingArguments(
     output_dir='./results',
-    evaluation_strategy='epoch',
+    evaluation_strategy='epoch',  # Evaluate at the end of each epoch
     learning_rate=2e-5,
-    per_device_train_batch_size=8,
+    per_device_train_batch_size=8,  # You can adjust batch size here
     per_device_eval_batch_size=8,
-    num_train_epochs=3,
+    num_train_epochs=num_train_epochs,  # Setting epochs to 3
     weight_decay=0.01,
     logging_dir='./logs',
+    logging_steps=500,
+    save_steps=1000,
+    load_best_model_at_end=True,
+    metric_for_best_model="accuracy",
+    max_steps=max_steps,  # Limit the total steps
 )
 
-# Initialize Trainer
 trainer = Trainer(
     model=model,
     args=training_args,
     train_dataset=train_dataset,
     eval_dataset=val_dataset,
-    compute_metrics=compute_metrics
+    compute_metrics=lambda p: {'accuracy': (p.predictions.argmax(axis=-1) == p.label_ids).mean()}
 )
 
-# Train the model
+# Start training
+print(f"Starting training for {num_train_epochs} epochs...")
 trainer.train()
 
 # Evaluate the model
+print("Evaluating model...")
 eval_results = trainer.evaluate()
 print(f"Evaluation results: {eval_results}")
 
+
 # Save the trained model
-model.save_pretrained('./dishTV_bert_model')
-tokenizer.save_pretrained('./dishTV_bert_model')
+model.save_pretrained('dishTV_bert_model')
+tokenizer.save_pretrained('dishTV_bert_model')
 
 # Testing the model with an example
 def predict(text):