Bypassing Google Colab GPU Limitations

This repository provides a solution for bypassing the GPU limitations in Google Colab by utilizing multiple Google accounts. By following this guide, you can continuously train your models using Colab GPUs.

Solution Steps

1. Create Multiple Google Accounts: Set up 5 to 6 Google accounts.
2. Share Project Folder:
   • Create a folder in Google Drive.
   • Place the .ipynb file of your project in this folder.
   • Share the folder with all the accounts created in the previous step.
3. Use Sample Code:
   • Utilize the sample code provided in this repository to save and load checkpoints for your model, including the model itself, the optimizer, the scheduler, min-loss, or any other necessary data.
   • Ensure that the path for checkpoints is in your shared folder (from the second step).
4. Run and Switch Accounts:
   • Run the code on one of your accounts. When the session ends or you reach the GPU usage limit, switch to another account and run the code again.
   • The model will continue training from the last saved checkpoint.

This method allows you to train your model as long as you want by simply switching accounts every approximately 5 hours.

Sample Code

The provided code sample demonstrates how to save and load checkpoints for your model. It includes:

• Mounting Google Drive
• Defining paths for saving the best model and checkpoints
• Loading checkpoints
• Training and evaluating the model and saving the checkpoints

Code Example

import torch
from torch.optim.lr_scheduler import ReduceLROnPlateau
from google.colab import drive

drive.mount('/content/drive')

BEST_MODEL_PATH = '/content/drive/MyDrive/MyModel/model_best.pt'
DRIVE_CHP_PATH = '/content/drive/MyDrive/MyModel/model_last_checkpoint.pth'


class Trainer():
    def __init__(self, config, dataset, checkpoint=None):
        # Initialization
        if checkpoint is None:
            self.model = self.get_model(self.device)
            self.optimizer = torch.optim.AdamW(self.model.parameters(), config.lr, weight_decay=config.wd)
            self.scheduler = ReduceLROnPlateau(self.optimizer, mode='min', factor=0.6, patience=3, min_lr=1e-6, threshold=0.01)
            self.epoch, self.minloss = 0, float('inf')
        else:
            self.model, self.optimizer, self.scheduler, self.epoch, self.minloss = self.load_checkpoint(checkpoint)
            print('\rCheckpoint loaded successfully'); print('-' * 50)

    # Add other necessary methods here...

    def load_checkpoint(self, checkpoint):
        minloss = checkpoint['minloss']
        epoch = checkpoint['epoch']
        model = checkpoint['model']
        optimizer = checkpoint['optimizer']
        scheduler = checkpoint['scheduler']
        return model, optimizer, scheduler, epoch, minloss

    def save_checkpoint(self, model, optimizer, scheduler, epoch, minloss, save_path):
        checkpoint = {
            'minloss': minloss,
            'epoch': epoch,
            'model': model,
            'optimizer': optimizer,
            'scheduler': scheduler}
        torch.save(checkpoint, save_path)

    def train_and_evaluate(self, n_epochs):
        for epoch in range(self.epoch, n_epochs):
            # ...
            train_loss = self.train(self.model, self.train_dataloader, self.optimizer, epoch, self.device)
            test_loss = self.evaluate(self.model, self.test_dataloader, self.device)
            self.save_checkpoint(self.model, self.optimizer, self.scheduler, epoch+1, self.minloss, DRIVE_CHP_PATH )
            # ...


if __name__ == '__main__':
    # Initialize trainer
    path = DRIVE_CHP_PATH
    checkpoint = torch.load(path) if os.path.isfile(path) else None
    trainer = Trainer(config, {'train_data': train_data, 'test_data': test_data}, checkpoint)

    # Train and evaluate the model
    trainer.train_and_evaluate(50)

Contact

If you encounter any issues or have questions about using this solution, feel free to reach out.