v0.6.0

What's New

• TorchTrainer can now accept models in which loss functions are included (e.g., ArcFace, LLMs). You must pass criterion = None to the TorchTrainer.train() and calculate the loss inside your training hook.

Sample code

def forward_train(self, trainer, inputs):
    embeddings, loss = trainer.model.*inputs)
    return loss, embeddings.detach()