DeepNoise

A PyTorch framework for learning with noisy labels.
Report Bug · Request Feature

---

DeepNoise is a PyTorch framework for learning with noisy labels in the context of deep learning. It provides implementations for well-known algorithms, common datasets in the literature, and noise-related utilities for learning from noisy labels.

Note that the DeepNoise framework is in an early stage of development. We plan to implement more algorithms, datasets, better logging and visualizations, and more noise-related utilities. The current API and structure of the framework might change in the future. All suggestions and contributions are welcome.

Built with

• PyTorch

Getting Started

Installation

git clone https://github.com/MohammedAlkhrashi/DeepNoise.git
cd DeepNoise
pip install -r requriments.txt

Usage

Through the provided training tool and config files

cd DeepNoise
python train.py --cfg_path configs/algorithms/default_sl.py

or

Through framework API:

1. Define a noise injection strategy

from DeepNoise.noise_injectors import SymmetricNoiseInjector
noise_injector = SymmetricNoiseInjector(noise_prob=0.4)

2. Define a torch dataset

from DeepNoise.datasets.cifar import NoisyCIFAR10
train_set = NoisyCIFAR10(
    noise_injector=noise_injector,
    train=True,
    download=True,
    transforms=train_transform,
    root=data_root,
)
test_set = NoisyCIFAR10(
    train=False, download=True, transforms=test_transforms, root=data_root
)

3. Define callbacks and initialize wandb

from DeepNoise.callbacks.lr_scheduler import StepLRScheduler
from DeepNoise.callbacks.statistics import SimpleStatistics
callbacks = [
    SimpleStatistics(),
    StepLRScheduler(optimizer, milestones=[80, 120], gamma=0.1),
]

wandb.init(project="DeepNoise")

4. Define a trainer

from DeepNoise.algorithms.symmetric_loss import SymmetericLossTrainer
trainer = SymmetericLossTrainer(
    model=model,
    optimizer=optimizer,
    train_loader=train_loader,
    val_loader=test_loader,
    epochs=epochs,
    callbacks=callbacks,
)

5. Finally, train on your noisy data using a robust algorithm!

trainer.start()

Check example.py for a full running example.

Todo

• Uplaod DeepNoise to PyPI
• Add a documentation website
• Improve Logging
• Add resouces used

Authors

• @MohammedAlkhrashi
• @HishamAlyahya

Acknowledgement

• README template by @DeeshanSharma