unnir / cvae Goto Github PK

I've noticed that when you display the training loss:

    print('====> Epoch: {} Average loss: {:.4f}'.format(
          epoch, train_loss / len(train_loader.dataset)))

you normalize that by the length of the entire dataset. Couldn't you instead divide it by the number of batches you passing for that epoch, in order to have an idea of the average loss based on batches?

I would like to contribute to this project, there are many aspects that can be improved. Here are the changes I propose for the code:

1. Determine the training device automatically.
   

2. Pass arguments to the script instead of declaring them in the code (e.g., batch_size, latent_size, epochs).
   

3. Compute the loss in the model's forward method for faster computations.
   

4. Remove the unnecessary one-hot encoding of labels since it's automatically done inside BCE.
   

5. Add a progress bar for both training and evaluation.
   

6. Modularize the code into multiple files to make it easier to explore and improve.
   

Additionally, there are several enhancements that could improve model training and usage:

1. Use Automatic Mixed Precision (AMP) for faster training on GPUs.
   

2. Implement a learning rate scheduler.
   

3. Add a loss graph visualizer like Wandb.
   

4. Add a method to save/load the model weights.
   

5. Add a requirements.txt file
   

These changes, along with other smaller improvements, will contribute to the overall enhancement of the project.

In loss_function instead of doing

    KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
    return BCE + K

shouldn’t we normalize the KLD term over the batch instead? For example doing torch.mean()