☢️ Learning with Retrospection ☢️

About

• LearningWithRetrospection.py contains class LWR which implements the algorithm in this paper

Algorithm

Usage

lwr=LWR(
k=1, # Number of Epochs (Interval) to update soft labelsupdate_rate=0.9, # The rate at which True Label weightage is decayednum_batches_per_epoch=len(dataset) //batch_size,
dataset_length=len(dataset),
output_shape=(10, ), # Number of Classestau=5, # Temperature -- Just leave it to 5 if you don't know what you're doingmax_epochs=20, # Max number of epochssoftmax_dim=1# Axis for softmax
)
data, target=data.to(device), target.to(device)
optimizer.zero_grad()
output=model(data)
loss=lwr(batch_idx, output, target, eval=False) # LWR expects LOGITSloss.backward()
optimizer.step()