pytorch-adapt 0.0.35

No project description provided

News

November 19: Git repo is now public

Documentation

• Documentation
• Installation instructions
• List of algorithms/papers implemented

Google Colab Examples

See the examples folder for notebooks you can download or run on Google Colab.

Overview

This library consists of 11 modules:

Module	Description
Adapters	Wrappers for training and inference steps
Containers	Dictionaries for simplifying object creation
Datasets	Commonly used datasets and tools for domain adaptation
Frameworks	Wrappers for training/testing pipelines
Hooks	Modular building blocks for domain adaptation algorithms
Layers	Loss functions and helper layers
Meta Validators	Post-processing of metrics, for hyperparameter optimization
Models	Architectures used for benchmarking and in examples
Utils	Various tools
Validators	Metrics for determining and estimating accuracy
Weighters	Functions for weighting losses

How to...

Use in vanilla PyTorch

from pytorch_adapt.hooks import DANNHook
from pytorch_adapt.utils.common_functions import batch_to_device

# Assuming that models, optimizers, and dataloader are already created.
hook = DANNHook(optimizers)
for data in tqdm(dataloader):
    data = batch_to_device(data, device)
    # Optimization is done inside the hook.
    # The returned loss is for logging.
    loss, _ = hook({}, {**models, **data})

Build complex algorithms

Let's customize DANNHook with:

• virtual adversarial training
• entropy conditioning

from pytorch_adapt.hooks import EntropyReducer, MeanReducer, VATHook

# G and C are the Generator and Classifier models
misc = {"combined_model": torch.nn.Sequential(G, C)}
reducer = EntropyReducer(
    apply_to=["src_domain_loss", "target_domain_loss"], default_reducer=MeanReducer()
)
hook = DANNHook(optimizers, reducer=reducer, post_g=[VATHook()])
for data in tqdm(dataloader):
    data = batch_to_device(data, device)
    loss, _ = hook({}, {**models, **data, **misc})

Wrap with your favorite PyTorch framework

For additional functionality, adapters can be wrapped with a framework (currently just PyTorch Ignite).

from pytorch_adapt.adapters import DANN
from pytorch_adapt.containers import Models, Optimizers
from pytorch_adapt.datasets import DataloaderCreator
from pytorch_adapt.frameworks.ignite import Ignite

# Assume G, C and D are existing models
models_cont = Models(models)
# Override the default optimizer for G and C
optimizers_cont = Optimizers((torch.optim.Adam, {"lr": 0.123}), keys=["G", "C"])
adapter = DANN(models=models_cont, optimizers=optimizers_cont)

dc = DataloaderCreator(num_workers=2)
trainer = Ignite(adapter)
trainer.run(datasets, dataloader_creator=dc)

Wrappers for other frameworks (e.g. PyTorch Lightning and Catalyst) are planned to be added.

Check your model's performance

You can do this in vanilla PyTorch:

from pytorch_adapt.validators import SNDValidator

# Assuming predictions have been collected
target_train = {"preds": preds}
validator = SNDValidator()
score = validator.score(epoch=1, target_train=target_train)

You can also do this using a framework wrapper:

validator = SNDValidator()
trainer = Ignite(adapter, validator=validator)
trainer.run(datasets, dataloader_creator=dc)

Run the above examples

See this notebook and the examples page for other notebooks.

Installation

Pip

pip install pytorch-adapt

To get the latest dev version:

pip install pytorch-adapt --pre

To use pytorch_adapt.frameworks.ignite:

pip install pytorch-adapt[ignite]

Conda

Coming soon...

Dependencies

Required dependencies:

• numpy
• torch >= 1.6
• torchvision
• torchmetrics
• pytorch-metric-learning >= 1.0.0.dev5

Acknowledgements

Contributors

Pull requests are welcome!

Advisors

Thank you to Ser-Nam Lim, and my research advisor, Professor Serge Belongie.

Logo

Thanks to Jeff Musgrave for designing the logo.