GLC

Unofficial implementation of Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise (NIPS 18) in PyTorch.

Usage

(See example.ipynb for a walkthrough on MNIST)

from datasets import GoldCorrectionDataset
from glc import CorrectionGenerator, GoldCorrectionLossFunction

c_gen = CorrectionGenerator(simulate=True, dataset=trn_ds, randomization_strength=1.0)

# Fetch both corrupted and clean datasets if in simuate mode
trusted_dataset, untrusted_dataset = c_gen.fetch_datasets()

"""
Train the model on untrusted_dataset
"""
# Generate correction matrix
label_correction_matrix = c_gen.generate_correction_matrix(trainer.model, 32)

# Wrap trusted and untrusted dataset together using GoldCorrectionDataset class
gold_ds = GoldCorrectionDataset(trusted_dataset, untrusted_dataset)
gold_dl = DataLoader(gold_ds, batch_size=32, shuffle=True)

# Modified loss function
gold_loss = GoldCorrectionLossFunction(label_correction_matrix)

"""
Train using gold_ds and gold_loss the model, until convergence
"""

Results

MNIST

Regular training on trusted data (~5% of entire data) -> 61.12 accuracy

Gold Loss Correction with 5% trusted -> 95.45 accuracy (All samples in untrusted data (95% of total data) is corrupted by randomly assigning labels)