theloop 🔄

This is simpliest and useful train loop for pytorch. You can easy train your model on any dataset and dataloader.

Installing

git clone https://github.com/alxmamaev/theloop
cd theloop
python setup.py install

Examples 🔬

• CIFAR10 example
• IMDB example (in progress)
• Metric learning example (in progress)
• Autoencoder example (in progress)

We have a cool logs 😎

   _____ _______       _____ _______   _______ _    _ ______ _      ____   ____  _____
  / ____|__   __|/\   |  __ \__   __| |__   __| |  | |  ____| |    / __ \ / __ \|  __ \
 | (___    | |  /  \  | |__) | | |       | |  | |__| | |__  | |   | |  | | |  | | |__) |
  \___ \   | | / /\ \ |  _  /  | |       | |  |  __  |  __| | |   | |  | | |  | |  ___/
  ____) |  | |/ ____ \| | \ \  | |       | |  | |  | | |____| |___| |__| | |__| | |
 |_____/   |_/_/    \_\_|  \_\ |_|       |_|  |_|  |_|______|______\____/ \____/|_|

 EXPERIMENT NAME: experiment
 EXPERIMENT ID: 4364
 NUM EPOCH: 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+


  |￣￣￣￣￣￣|
    EPOCH: 0  
  |＿＿＿＿＿＿|
(\__/) ||
(•ㅅ•) ||
/ 　 づ
BATCH 1562; ITER 1562: 100%|██████████| 1563/1563 [03:59<00:00,  3.99s/it, loss=1.63]
+-------------------+
|   EPOCH METRICS   |
+----------+--------+
|  Metric  | Vlaue  |
+----------+--------+
| accuracy | 0.6173 |
+----------+--------+
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

  |￣￣￣￣￣￣|
    THE END  
  |＿＿＿＿＿＿|
(\__/) ||
(•ㅅ•) ||
/ 　 づ
+-------------------+
|    BEST METRICS   |
+----------+--------+
|  Metric  | Vlaue  |
+----------+--------+
| accuracy | 0.6914 |
+----------+--------+

Features 📊

• Cool logs 😎
• Tensorboard logging
• Checkpoint saving
• Validation
• Selecting the best checkpoint
• Scheduler support
• Tqdm notebook support
• Early stopping by ^C

Documentation 🗂

Parameters:

• model - your nn.Module model
• criterion - string with name of loss or loss class
• batch_callback - function of batch callback (look a examples)
• val_callback - function of batch callback (look a examples)
• optimizer - string with name of optimizer or optimizer class
• optimizer_param - dict of key-value parameters that will be push into optimizer
• scheduler - string with name of scheduler or scheduler class
• scheduler_params - dict of key-value parameters that will be push into scheduler
• device - string with name of acceleration device (default "cpu")
• val_rate - int rate of iteration when validation was starting
• logdir - string path to log directory
• name - string name of your experiment
• loss_key - string key of loss in dict returned by batch_callback
• val_criterion_key - string key of validation in dict returned by val_callback
• val_criterion_mode - string mode of selecting best checkpoint ("max" or "min")
• use_best_model - bool use best (by validation metric) model as final model
• use_tqdm_notebook - bool use tqdm_notebook instead tqdm

theloop.a

Train model from dataset

• train_dataset - torch Dataset for training
• val_dataset - torch Dataset for validation
• batch_size - int batch size
• n_workers - int number of workers
• shuffle - bool shuffle dataset
• n_epoch - int number of epoch

theloop.ka

Train model from dataloader

• train_dataloader - torch DataLoader for training
• val_dataloader - torch DataLoader for validation
• n_epoch - int number of epoch

To do

• Early stopping by validation metrics
• Pushing images to Tensorboard
• Data parallel