Pytorch Template

project for pytorch implementation example of image classification

Requirements

• python >= 3.7
• pytorch >= 1.0
• pyyaml
• scikit-learn
• wandb
• pre-commit (for pre-commit formatting, type check and testing)
• hiddenlayer
• graphviz
• python wrapper for graphviz

Please run poetry install to install the necessary packages.

You can also setup the environment using docker and docker-compose.

Dataset

Flowers Recognition Dataset Download the dataset from HERE．

Directory Structure

.
├── docs/
├── LICENSE
├── README.md
├── dataset/
│   └── flowers/
├── pyproject.toml
├── .gitignore
├── .gitattributes
├── .pre-commit-config.yaml
├── poetry.lock
├── docker-compose.yaml
├── Dockerfile
├── tests/
└── src/
    ├── csv
    ├── libs/
    ├── utils
    ├── notebook/
    ├── result/
    ├── scripts/
    │   └── experiment.sh
    ├── train.py
    └── evaluate.py

Features

• configuration class using dataclasses.dataclass (libs/config.py)

  • type check.
  • detection of unnecessary / extra parameters in a specified configuration.
  • dataclass is an immutable object, which prevents the setting from being changed by mistake.

• automatically generating configuration files (utils/make_configs.py)

  • e.g.) run this command

  python utils/make_configs.py --model resnet18 resnet30 resnet50 --learning_rate 0.001 0.0001 --dataset_name flower

  then you can get all of the combinations with model and learning_rate (total 6 config files), while the other parameters are set by default as described in libs/config.py.

  You can choose which data you use in experiment by specifying dataset_name. The lists of data for training, validation and testing are saved as csv files. You can see the paths to them in libs/dataset_csv.py and get them corresponding to dataset_name. If you want to use another dataset, please add csv files and the paths in DATASET_CSVS in libs/dataset_csv.py.

  You can also set tuple object parameters in configs like the below.

  python utils/make_configs.py --model resnet18 --topk 1 3 --topk 1 3 5

  By running this, you can get two configurations, in one of which topk parameter is (1, 3) and in the other topk parameter is (1, 3, 5).

• running all the experiments by running shell scripts (scripts/experiment.sh)

• support type annotation (typing)

• code formatting with black, isort and flake8

• visualize model for debug using hiddenlayer (src/utils/visualize_model.py)

Experiment

Please see scripts/experiment.sh for the detail. You can set configurations and run all the experiments by the below command.

sh scripts/experiment.sh

Setup dependencies

If you use local environment, then run

poetry install

If you use docker, then run

docker-compose up -d --build
docker-compose run mlserver bash

training

python train.py ./result/xxxx/config.yaml

evaluation

python evaluate.py ./result/xxxx/config.yaml validation
python evaluate.py ./result/xxxx/config.yaml test

Model visualization

python utils/visualize_model.py MODEL_NAME

Formatting

• black
• flake8
• isort

TODO

• pytorch implementation of image classification
• configuration class using dataclasses.dataclass
• auto generation of config yaml files
• shell script to run all the experiment
• support typing (type annotation)
• test code (run testing with pre-commit check)
• mypy (pre-commit check)
• formatting (pre-commit isort, black and flake8)
• calculate cyclomatic complexity / expression complexity / cognitive complexity (flake8 extension)
• CI for testing using GitHub Actions
• visualization of models
• add Dockerfile and docker-compose.yaml

License

This repository is released under the MIT License