How to load a model from model-only checkpoints?

[ShuhuaGao]

By default, the best model and the last model are saved by catalyst as checkpoints. For example, I have the following files in the checkpoints folder:

model.0049.pth
model.best.pth
model.last.pth
model.storage.json

I then use the following code to read the checkpoint.

checkpoint = utils.load_checkpoint(path=os.path.join(model_dir, 'model.best.pth'))
checkpoint.keys()
# the output is: odict_keys(['model.0.weight', 'model.0.bias', 'model.2.weight', 'model.2.bias', 'model.4.weight', 'model.4.bias'])

The obtained checkpoint contains weights and bias values for each linear layer (the underlying net is just a MLP).

Now the question is how I can restore the network model with checkpoint. I can't work with utils.unpack_checkpoint(checkpoint, best_model) since it depends on the 'model_state_dict' key but our checkpoint above does not have it.

Is there any method in utils handling this task? Or do I have to make it manually with PyTorch methods?

[ShuhuaGao]

Sorry for my ignorance..

In the model-only saving mode, the model.*.pth file contains exactly the state_dict of a model in pytorch. Supposing we have defined our model, simply do it like this to restore the trained model:

model.load_state_dict(torch.load(os.path.join(model_dir, 'model.best.pth'))