In this part of the assignment, you’ll complete the train_model function in train.py, which trains a BERT-based model for intent classification and slot filling. This task will deepen your understanding of the training process, how loss functions work for multi-task learning, and how to use backpropagation to optimize a model.

Understanding Training

For more general and preliminary information check out: [TBC]Preliminaries and Quiz Materials.

1. What is Training?

   • Training is the process of optimizing a model to make better predictions by minimizing a loss function.

   • For your BERT-based model:

     • Intent Classification: Predict the intent of a user query (e.g., addFilter or recipeRequest).

     • Slot Filling: Assign BIO tags to each token in the query (e.g., B-shortTimeKeyWord for "fast").

2. How Does the Training Loop Work?

   • Forward Pass: The input data flows through the model to generate predictions.

   • Loss Calculation: The predictions are compared with ground truth labels to calculate the loss.

   • Backward Pass (Backpropagation): The model adjusts its weights based on the loss to improve predictions in the next iteration.

3. What Are We Optimizing?

   • Intent Loss: Measures how well the model predicts the intent.

   • Slot Loss: Measures how well the model predicts slot tags for each token.

4. Why Do We Use Two Loss Functions?

   • Your model performs two tasks simultaneously, so you need separate losses for each task. These are combined to train the model in a balanced way.

Steps to Complete the train_model Function

The train_model function is incomplete, and your task is to fill in the missing pieces. This function is the backbone of the training process for a dual-task model that handles intent classification and slot filling. Through this exercise, you’ll learn how to integrate loss functions, perform forward passes, and optimize a model’s weights effectively.

Assignment: Completing the Training Function

Objective

The train_model function is incomplete, and your task is to fill in the missing pieces. This function is the backbone of the training process for a dual-task model that handles intent classification and slot filling. Through this exercise, you’ll learn how to integrate loss functions, perform forward passes, and optimize a model’s weights effectively.

Understanding the Missing Slots

1. Loss Functions

   • Two loss functions are needed:

     • Intent Loss: Measures how well the model predicts the intent.

     • Slot Loss: Measures how well the model tags tokens for slots using BIO tags.

   • Hints:

     • Think about which loss functions are commonly used for classification tasks.

     • Both tasks deal with categorical data, so the loss functions should compare predicted probabilities with actual labels.

     • Look into PyTorch’s Loss Functions documentation for insights. (Super Hint: our preferred loss function is in the first 5 loss functions on this documentation list)

   Questions to Consider:

   • How do loss functions compute the difference between predicted and true labels?

   • Can the same type of loss function work for both intents and slots? Why?

2. Optimizer

   • The optimizer updates the model’s weights to minimize the loss.

   • Hints:

     • The optimizer needs access to the model’s parameters.

     • A learning rate is crucial—too high might cause erratic updates; too low might slow down training.

     • Explore PyTorch’s Optimization documentation to understand popular optimizers.

   Questions to Consider:

   • Why do we use an optimizer instead of manually adjusting weights?

   • How does the learning rate impact training?

3. Forward Pass

   • The forward pass feeds the input data into the model to generate predictions:

     • intent_logits: Predicted logits for intent classification.

     • slot_logits: Predicted logits for slot tagging.

   • Hints:

     • The model expects tokenized inputs and an attention mask.

     • Think about what the model outputs when given inputs for two tasks.

   Questions to Consider:

   • What are logits, and how are they different from probabilities?

   • How does the model simultaneously handle intent and slot predictions?

4. Loss Calculation

   • After the forward pass, compute the losses:

     • Intent Loss: Compare intent_logits with intent_labels.

     • Slot Loss: Compare slot_logits with slot_labels.

   • Hints:

     • slot_logits and slot_labels may need reshaping to align their dimensions for the loss function.

     • Combine both losses into a single loss for backpropagation. Think about how best to combine them… this is arguably a hyperparameter.

   Questions to Consider:

   • Why might slot logits need to be flattened before calculating the loss?

   • How does combining intent and slot losses impact model training?

5. Backpropagation

   • Backpropagation updates the model weights based on the calculated loss.

   • Steps:

     • Zero the gradients using optimizer.zero_grad() to prevent accumulation.

     • Use loss.backward() to compute gradients.

     • Apply the gradients to update the weights with optimizer.step().

   Questions to Consider:

   • Why do we need to zero the gradients before backpropagation?

   • What happens if you skip the optimizer.step() step?

6. Loss Logging

   • After updating the weights, print the loss for each epoch or batch to monitor training progress.

   • Hints:

     • Loss values should decrease over epochs if the model is learning effectively.

     • Keep track of both intent and slot losses separately to ensure balanced training.

   Questions to Consider:

   • What does it mean if the loss doesn’t decrease?

   • Why is it helpful to log intent and slot losses separately?

Hints for Success

• Loss Functions: Look into loss functions designed for classification tasks.

• Reshaping Slot Logits: Flatten the logits and labels to make them compatible with the loss function.

• Optimizer: The Adam optimizer is a common choice for deep learning models.

By completing this assignment, you’ll gain hands-on experience with key components of model training, preparing you for more advanced tasks. Take your time, think critically, and ask for clarification if needed!