News

• [2022-3-22]: **v0.2.2 has been released:
  • Fix some bugs.
• [2021-11-3]: **v0.2.0 has been released:
  • New features:
    • Change the format of link configuration.
• [2021-10-27]: **v0.1.4 has been released:
  • New features:
    • Add contrastive representation learning methods (MoCo-V2).
• [2021-10-24]: **v0.1.2 has been released:
  • New features:
    • Add distributed (DDP) support.
• [2021-10-7]: **v0.1.1 has been released:
  • New features:
    • Change the Cars196 loading method.
• [2021-9-15]: **v0.1.0 has been released:
  • New features:
    • output_wrapper and pipeline setting are decomposed for convenience.
    • Pipeline will be stored in the experiment folder using a directed graph.
• [2021-9-13]: **v0.0.1 has been released:
  • New features:
    • config.yaml will be created to store the configuration in the experiment folder.**
• [2021-9-6]: v0.0.0 has been released.

Introduction

GeDML is an easy-to-use generalized deep metric learning library, which contains:

• State-of-the-art DML algorithms: We contrain 18+ losses functions and 6+ sampling strategies, and divide these algorithms into three categories (i.e., collectors, selectors, and losses).
• Bridge bewteen DML and SSL: We attempt to bridge the gap between deep metric learning and self-supervised learning through specially designed modules, such as collectors.
• Auxiliary modules to assist in building: We also encapsulates the upper interface for users to start programs quickly and separates the codes and configs for managing hyper-parameters conveniently.

Installation

Pip

pip install gedml

Quickstart

Demo 1: deep metric learning

CUDA_VISIBLE_DEVICES=0 python demo.py \
--data_path <path_to_data> \
--save_path <path_to_save> \
--eval_exclude f1_score NMI AMI \
--device 0 --batch_size 128 --test_batch_size 128 \
--setting proxy_anchor --splits_to_eval test --embeddings_dim 128 \
--lr_trunk 0.0001 --lr_embedder 0.0001 --lr_collector 0.01 \
--dataset cub200 --delete_old \

Demo 2: contrastive representation learning

CUDA_VISIBLE_DEVICES=0 python demo.py \
--data_path <path_to_data> \
--save_path <path_to_save> \
--eval_exclude f1_score NMI AMI \
--device 0 --batch_size 128 --test_batch_size 128 \
--setting mocov2 --splits_to_eval test --embeddings_dim 128 \
--lr_trunk 0.015 --lr_embedder 0.015 \
--dataset imagenet --delete_old \

If you want to use our code to conduct DML or CRL experiments, please refer to the up-to-date and most detailed configurations below: :point_down:

• If you use the command line, you can run sample_run.sh to try this project.
• If you debug with VS Code, you can refer to launch.json to set .vscode.

API

Initialization

Use ParserWithConvert to get parameters

>>> from gedml.launcher.misc import ParserWithConvert
>>> csv_path = ...
>>> parser = ParserWithConvert(csv_path=csv_path, name="...")
>>> opt, convert_dict = parser.render()

Use ConfigHandler to create all objects.

>>> from gedml.launcher.creators import ConfigHandler
>>> link_path = ...
>>> assert_path = ...
>>> param_path = ...
>>> config_handler = ConfigHandler(
    convert_dict=convert_dict,
    link_path=link_path,
    assert_path=assert_path,
    params_path=param_path,
    is_confirm_first=True
)
>>> config_handler.get_params_dict()
>>> objects_dict = config_handler.create_all()

Start

Use manager to automatically call trainer and tester.

>>> from gedml.launcher.misc import utils
>>> manager = utils.get_default(objects_dict, "managers")
>>> manager.run()

Or directly use trainer and tester.

>>> from gedml.launcher.misc import utils
>>> trainer = utils.get_default(objects_dict, "trainers")
>>> tester = utils.get_default(objects_dict, "testers")
>>> recorder = utils.get_default(objects_dict, "recorders")
# start to train
>>> utils.func_params_mediator(
    [objects_dict],
    trainer.__call__
)
# start to test
>>> metrics = utils.func_params_mediator(
    [
        {"recorders": recorder},
        objects_dict,
    ],
    tester.__call__
)

Document

For more information, please refer to: :point_right: Docs :book:

Some specific guidances:

• How to write config
• How to define custom module
• ...

Configs

We will continually update the optimal parameters of different configs in TsinghuaCloud

Framework

This project is modular in design. The pipeline diagram is as follows:

Code structure

• tests: Debug files.
• examples: Demos.
  • config: Folder which stores links, param, wrapper, args.yaml and assert.yaml.
  • demo.py: A demo python file.
• src/gedml: Main code.
  • core: Losses, selectors, collectors, etc.
  • client: Tmux manager.
  • config: Config files which define the hyper-string etc.
  • launcher: Manager, Trainer, Tester, etc.
  • recorder: Recorder.

Method

Collectors

Losses

classifier-based

pair-based

proxy-based

Selectors

Code Reference

• KevinMusgrave / pytorch-metric-learning
• KevinMusgrave / powerful-benchmarker
• Confusezius / Deep-Metric-Learning-Baselines
• facebookresearch / moco
• PatrickHua / SimSiam
• ujjwaltiwari / Deep_Variational_Metric_Learning
• idstcv / SoftTriple
• wzzheng / HDML
• google-research / simclr
• kunhe / FastAP-metric-learning
• wy1iu / LargeMargin_Softmax_Loss
• tjddus9597 / Proxy-Anchor-CVPR2020
• facebookresearch / deit

TODO:

• assert parameters.
• write github action to automate unit-test, package publish and docs building.
• add cross-validation splits protocol.
• distributed tester for matrix-form input.
• add metrics module.
• how to improve the running efficiency.

IMPORTANT TODO:

• re-define pipeline setting!!!
• simplify distribution setting!!