A Light Pytorch Project Template for Intensive AI Research

Feature

• TensorBoard / wandb support
• Background generator is used (reason of using background generator)
  • In Windows, background generator could not be supported. So if error occurs, set false to use_background_generator in config
• Training state and network checkpoint saving, loading
  • Training state includes not only network weights, but also optimizer, step, epoch.
  • Checkpoint includes only network weights. This could be used for inference.
• Hydra and Omegaconf is supported
• Distributed Learning using Distributed Data Parallel is supported
• Config with yaml file / easy dot-style access to config
• Code lint / CI
• Code Testing with pytest

Code Structure

• assets dir: icon image of Pytorch Project Template. You can remove this directory.
• config dir: directory for config files
• dataloader dir: dataloader and dataset codes are here. Also, put dataset in meta dir.
• model dir: model.py is for wrapping network architecture. model_arch.py is for coding network architecture.
• tests dir: directory for pytest testing codes. You can check your network's flow of tensor by fixing tests/model/net_arch_test.py. Just copy & paste Net_arch.forward method to net_arch_test.py and add assert phrase to check tensor.
• utils dir:
  • utils.py is for utility. random seed setting, dot-access hyper parameter, get commit hash, etc are here.
  • writer.py is for writing logs in tensorboard / wandb.
• trainer.py file: this is for setting up and iterating epoch.

Config Structure

── configs                   <- Hydra configuration files
    ├── default.yaml               <- enter point of hydra
    ├── datamodule               <- Datamodule configs folder (add your data config under it)
    |   └── default.yaml          <- Datamodule default yaml 
    ├── hydra/job_logging        <- Hydra configs
    |   └── custom.yaml          <- customize hydra config here 
    ├── model                    <- Model configs (add your model config under it)
    |   └── default.yaml          <- Model default yaml 
    └── general                 <- general config for training
        └── default.yaml          <- General default yaml (e.g. logger)

Hightlight

• Support for multiple dataloaders and multiple models in one project
• Simply write another model.network.Network class and dataloader.dataloader.Dataset_
• Then add two more yaml file for new network and Dataloader separately under the config/model and config/datamodule
• The yaml file could be quite similar to the default.yaml under under the config/model and config/datamodule

Setup

Install requirements

• python3 (3.6, 3.7, 3.8 is tested)
• Write PyTorch version which you want to requirements.txt. (https://pytorch.org/get-started/)
• pip install -r requirements.txt

Config

• Config is written in yaml file
  • You can choose configs at config/default.yaml. Custom configs are under config/job/
• general field
  • name is train name you run.
  • working_dir is root directory for saving checkpoints, logging logs.
  • device is device mode for running your model. You can choose cpu or cuda
  • random_seed is for setting python, numpy, pytorch random seed.
  • dist is for configuring Distributed Data Parallel.
    • gpus is the number that you want to use with DDP (gpus value is used at world_size in DDP). Not using DDP when gpus is 0, using all gpus when gpus is -1.
    • timeout is seconds for timeout of process interaction in DDP. When this is set as ~, default timeout (1800 seconds) is applied in gloo mode and timeout is turned off in nccl mode.
• data field
  • Configs for Dataloader.
  • change _target_ for the dataloader(or data generator) you want to use
  • glob train_dir / test_dir with file_format for Dataloader.
  • If divide_dataset_per_gpu is true, origin dataset is divide into sub dataset for each gpu. This could mean the size of origin dataset should be multiple of number of using gpu. If this option is false, dataset is not divided but epoch goes up in multiple of number of gpus.
• model field
  • Configs for Network architecture and options for model.
  • change _target_ for the the model network you want to use
  • You can add configs in yaml format to config your network.
• log field
  • Configs for logging include tensorboard / wandb logging.
  • summary_interval and checkpoint_interval are interval of step and epoch between training logging and checkpoint saving.
  • checkpoint and logs are saved under working_dir/chkpt_dir and working_dir/trainer.log. Tensorboard logs are saving under working_dir/outputs/tensorboard
• load field
  • loading from wandb server is supported
  • wandb_load_path is Run path in overview of run. If you don't want to use wandb load, this field should be ~.
  • network_chkpt_path is path to network checkpoint file. If using wandb loading, this field should be checkpoint file name of wandb run.
  • resume_state_path is path to training state file. If using wandb loading, this field should be training state file name of wandb run.

Code lint

1. pip install -r requirements-dev.txt for install develop dependencies (this requires python 3.6 and above because of black)

2. pre-commit install for adding pre-commit to git hook

3. conda env create -f environment.yml for creating a conda venv from evironment.yml but some of the dependencies may conflict with your environment like cuda

Sample Train

• python trainer.py device=cuda
• python trainer.py log.use_wandb=true device=cpu
• python trainer.py device=cuda dist.gpus=-1 train.num_epoch=100

TO DO

1. Implement progress bar in Multi-GPU trainning
2. Multi-GPU trainning with hydra.utils.instantiate()

Inspired by

• https://github.com/open-mmlab/mmsr
• https://github.com/ryul99/pytorch-project-template
• https://github.com/ashleve/lightning-hydra-template
• https://github.com/allenai/allennlp (test case writing)