Skip to content

Latest commit

 

History

History
50 lines (31 loc) · 2.64 KB

HW2.md

File metadata and controls

50 lines (31 loc) · 2.64 KB

Homework: Fine-Tuning a Pre-trained Model on the Butterfly Dataset

Objective

In this homework, you will fine-tune a pre-trained image generation model using the butterfly dataset. The goal is to understand the process of fine-tuning and observe how the model adapts to a new dataset.

Instructions

1. Load and Prepare the Dataset

  • Download the butterfly dataset.
  • Create a dataloader to sample a batch of images.
  • Preprocess the images by resizing them, applying random horizontal flips, converting to tensors, and normalizing.

2. Fine-Tune the Model

  • Set the optimization target to image_pipe.unet.parameters().
  • Create a training loop to update the model weights using the preprocessed dataset.
  • Implement gradient accumulation to handle small batch sizes and GPU memory constraints.

3. Considerations

  • Batch Size: Use a small batch size due to GPU memory constraints. You can adjust the image size to speed up training or allow for larger batches.
  • Gradient Accumulation: Accumulate gradients over several batches before updating the model parameters. This simulates a larger batch size and helps in stabilizing the training process.
  • Logging and Monitoring: Generate samples during training to visually inspect the model's progress. Use logging tools like Weights and Biases (wandb) or TensorBoard to track training metrics.

4. Optional: Logging with Weights and Biases

  • Initialize a wandb project and log loss values during training.
  • Track the average loss per epoch for better monitoring.

Submission

Submit a single Jupyter Notebook file (.ipynb) with the following:

  1. Code Implementation: The complete code used to fine-tune the model, including:

    • Loading the dataset
    • Preprocessing
    • Setting up the training loop
    • Applying gradient accumulation

  2. Training Logs: Logs showing the training progress, including loss values per epoch. If using wandb, include a link to your project dashboard.

  3. Generated Samples: A few images generated by the model at different stages of training (e.g., after each epoch).

Resources

For detailed code and step-by-step instructions, refer to the Fine-Tuning Guide on GitHub. This notebook contains the complete code and explanations needed to fine-tune the model on the butterfly dataset and monitor its performance.

By completing this homework, you will gain practical experience in fine-tuning a pre-trained image generation model and understanding the considerations involved in adapting models to new datasets. Good luck and have fun experimenting!