A Toolkit for Training, Tracking and Saving PyTorch Models

gpu-memory, logger, pytorch, tensorboard, toolkit
pip install torch-scope==0.5.6



License Documentation Status Downloads PyPI version

A Toolkit for training pytorch models, which has three features:

  • Tracking environments, dependency, implementations and checkpoints;
  • Providing a logger wrapper with two handlers (to std and file);
  • Supporting automatic device selection;
  • Providing a tensorboard wrapper;
  • Providing a spreadsheet writer to automatically summarizing notes and results;

We are in an early-release beta. Expect some adventures and rough edges.

Quick Links


To install via pypi:

pip install torch-scope

To build from source:

pip install git+https://github.com/LiyuanLucasLiu/Torch-Scope


git clone https://github.com/LiyuanLucasLiu/Torch-Scope.git
cd Torch-Scope
python setup.py install


An example is provided as below, please read the doc for a detailed api explaination.

  • set up the git in the server & add all source file to the git
  • use tensorboard to track the model stats (tensorboard --logdir PATH/log/ --port ####)
from torch_scope import wrapper
logger = logging.getLogger(__name__)

if __name__ == '__main__':

    parser = argparse.ArgumentParser()

    parser.add_argument('--checkpoint_path', type=str, ...)
    parser.add_argument('--name', type=str, ...)
    parser.add_argument('--gpu', type=str, ...)
    args = parser.parse_args()

    pw = wrapper(os.path.join(args.checkpoint_path, args.name), name = args.log_dir, enable_git_track = False)
    # Or if the current folder is binded with git, you can turn on the git tracking as below
    # pw = wrapper(os.path.join(args.checkpoint_path, args.name), name = args.log_dir, enable_git_track = True)
    # if you properly set the path to credential_path and want to use spreadsheet writer, turn on sheet tracking as below
    # pw = wrapper(os.path.join(args.checkpoint_path, args.name), name = args.log_dir, \
    #             enable_git_track=args.git_tracking, sheet_track_name=args.spreadsheet_name, \ 
    #             credential_path="/data/work/jingbo/ll2/Torch-Scope/torch-scope-8acf12bee10f.json")
    gpu_index = pw.auto_device() if 'auto' == args.gpu else int(args.gpu)
    device = torch.device("cuda:" + str(gpu_index) if gpu_index >= 0 else "cpu")

    pw.save_configue(args) # dump the config to config.json

    # if the spreadsheet writer is enabled, you can add a description about the current model
    # pw.add_description(args.description) 

    logger.info(str(args)) # would be plotted to std & file if level is 'info' or lower


    batch_index = 0

    for index in range(epoch):


    	for instance in ... :

    		loss = ...

    		tot_loss += loss.detach()

    		if batch_index % ... = 0:
    			pw.add_loss_vs_batch({'loss': tot_loss / ..., ...}, batch_index, False)
    			pw.add_model_parameter_stats(model, batch_index, save=True)
    			pw.add_model_update_stats(model, batch_index)
    			tot_loss = 0

    		batch_index += 1

    	dev_score = ...
    	pw.add_loss_vs_batch({'dev_score': dev_score, ...}, index, True)

    	if dev_score > best_score:
    		pw.save_checkpoint(model, optimizer, is_best = True)
    		best_score = dev_score
    		pw.save_checkpoint(model, optimizer, is_best = False)

Advanced Usage

Auto Device

Git Tracking

Spreadsheet Logging

Share the spreadsheet with the following account torch-scope@torch-scope.iam.gserviceaccount.com. And access the table with its name.