Counting 2,784 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Author
Contributors
Last Commit
May. 27, 2018
Created
Mar. 13, 2018

PyTorch Template Project

PyTorch deep learning project made easy.

Requirements

  • Python 3.x
  • PyTorch

Features

  • Clear folder structure which is suitable for many deep learning projects.
  • .json config file support for more convenient parameter tuning.
  • Checkpoint saving and resuming.
  • Abstract base classes for faster development:
    • BaseTrainer handles checkpoint saving/resuming, training process logging, and more.
    • BaseDataLoader handles batch generation, data shuffling, and validation data splitting.
    • BaseModel provides basic model summary.

Folder Structure

pytorch-template/
│
├── train.py - example main
├── config.json - example config file
│
├── base/ - abstract base classes
│   ├── base_data_loader.py - abstract base class for data loaders
│   ├── base_model.py - abstract base class for models
│   └── base_trainer.py - abstract base class for trainers
│
├── data_loader/ - anything about data loading goes here
│   └── data_loaders.py
│
├── datasets/ - default datasets folder
│
├── logger/ - for training process logging
│   └── logger.py
│
├── model/ - models, losses, and metrics
│   ├── modules/ - submodules of your model
│   ├── loss.py
│   ├── metric.py
│   └── model.py
│
├── saved/ - default checkpoints folder
│
├── trainer/ - trainers
│   └── trainer.py
│
└── utils/
    ├── util.py
    └── ...

Usage

The code in this repo is an MNIST example of the template.

Config file format

Config files are in .json format:

{
  "name": "Mnist_LeNet",        // training session name
  "cuda": true,                 // use cuda
  "data_loader": {
      "data_dir": "datasets/",  // dataset path
      "batch_size": 32,         // batch size
      "shuffle": true           // shuffle data each time calling __iter__()
  },
  "validation": {
      "validation_split": 0.1,  // validation data ratio
      "shuffle": true           // shuffle training data before splitting
  },
  "optimizer_type": "Adam",
  "optimizer": {
      "lr": 0.001,              // (optional) learning rate
      "weight_decay": 0         // (optional) weight decay
  },
  "loss": "my_loss",            // loss
  "metrics": [                  // metrics
    "my_metric",
    "my_metric2"
  ],
  "trainer": {
      "epochs": 1000,           // number of training epochs
      "save_dir": "saved/",     // checkpoints are saved in save_dir/name
      "save_freq": 1,           // save checkpoints every save_freq epochs
      "verbosity": 2,           // 0: quiet, 1: per epoch, 2: full
      "monitor": "val_loss",    // monitor value for best model
      "monitor_mode": "min"     // "min" if monitor value the lower the better, otherwise "max" 
  },
  "arch": "MnistModel",         // model architecture
  "model": {}                   // model configs
}

Add addional configurations if you need.

Using config files

Modify the configurations in .json config files, then run:

python train.py --config config.json

Resuming from checkpoints

You can resume from a previously saved checkpoint by:

python train.py --resume path/to/checkpoint

Customization

Data Loader

  • Writing your own data loader
  1. Inherit BaseDataLoader

    BaseDataLoader is similar to torch.utils.data.DataLoader, you can use either of them.

    BaseDataLoader handles:

    • Generating next batch
    • Data shuffling
    • Generating validation data loader by calling BaseDataLoader.split_validation()
  2. Implementing abstract methods

    You need to implement these abstract methods:

    • _pack_data(): pack data members into a list of tuples
    • _unpack_data(): unpack packed data
    • _update_data(): updata data members
    • _n_samples(): total number of samples
  • DataLoader Usage

    BaseDataLoader is an iterator, to iterate through batches:

    for batch_idx, (x_batch, y_batch) in data_loader:
        pass
  • Example

    Please refer to data_loader/data_loaders.py for an MNIST data loading example.

Trainer

  • Writing your own trainer
  1. Inherit BaseTrainer

    BaseTrainer handles:

    • Training process logging
    • Checkpoint saving
    • Checkpoint resuming
    • Reconfigurable monitored value for saving current best
      • Controlled by the configs monitor and monitor_mode, if monitor_mode == 'min' then the trainer will save a checkpoint model_best.pth.tar when monitor is a current minimum
  2. Implementing abstract methods

    You need to implement _train_epoch() for your training process, if you need validation then you can implement _valid_epoch() as in trainer/trainer.py

  • Example

    Please refer to trainer/trainer.py for MNIST training.

Model

  • Writing your own model
  1. Inherit BaseModel

    BaseModel handles:

    • Inherited from torch.nn.Module
    • summary(): Model summary
  2. Implementing abstract methods

    Implement the foward pass method forward()

  • Example

    Please refer to model/model.py for a LeNet example.

Loss and metrics

If you need to change the loss function or metrics, first import those function in train.py, then modify "loss" and "metrics" in .json config files

Multiple metrics

You can add multiple metrics in your config files:

"metrics": ["my_metric", "my_metric2"],

Additional logging

If you have additional information to be logged, in _train_epoch() of your trainer class, merge them with log as shown below before returning:

additional_log = {"gradient_norm": g, "sensitivity": s}
log = {**log, **additional_log}
return log

Validation data

To split validation data from a data loader, call BaseDataLoader.split_validation(), it will return a validation data loader, with the number of samples according to the specified ratio in your config file.

Note: the split_validation() method will modify the original data loader Note: split_validation() will return None if "validation_split" is set to 0

Checkpoints

You can specify the name of the training session in config files:

"name": "MNIST_LeNet",

The checkpoints will be saved in save_dir/name.

The config file is saved in the same folder.

Note: checkpoints contain:

{
  'arch': arch,
  'epoch': epoch,
  'logger': self.train_logger,
  'state_dict': self.model.state_dict(),
  'optimizer': self.optimizer.state_dict(),
  'monitor_best': self.monitor_best,
  'config': self.config
}

Contributing

Feel free to contribute any kind of function or enhancement, here the coding style follows PEP8

TODOs

  • Iteration-based training (instead of epoch-based)
  • Deprecate BaseDataLoader, use torch.utils.data instesad
  • Multi-GPU support
  • Multiple optimizers and lr scheduler
  • Update the example to PyTorch 0.4 (or 1.0)
  • TensorboardX or visdom logger support
  • Configurable logging layout, checkpoint naming
  • Load settings from config files

License

This project is licensed under the MIT License. See LICENSE for more details

Acknowledgments

This project is inspired by the project Tensorflow-Project-Template by Mahmoud Gemy