Pytorch. Snapshots Weights Averaging.

Okay, you have a number of checkpoints from a train-loop for 100500 epochs. Or you’ve carried some experiments with one architecture and changed only global parameters and now you have 7 saved .pth models in your folder.

But all of these models still can’t achieve 90% accuracy … 0.5–0.7% are missing.

Then how to reach the desired accuracy of 90%? The answer is to combine all models & average weights from snapshots.

Explanation.

Get a dictionary for each snapshot: parameters’ names and values.

for snapshot_path in list_of_snapshots_paths:
    model = load_net(path=snapshot_path)
    snapshots_weights[snapshot_path] = 
                             dict(model.named_parameters())

Iterate on each parameter and set in new state_dict averaged value.

custom_params += snapshot_params[name].data        dict_params[name].data.copy_(custom_params/N)

Load new state_dict into the model.

model_dict = model.state_dict()
model_dict.update(dict_params)
model.load_state_dict(model_dict)