Modules

Dataset

you should keep dataset and loader in dataset.py

Model

you should keep models in model.py

Gstate

gstate has global variables. You can use it to avoid unnecessary parameters, and save model and load model in an much easy method.

Experiment

you should keep experiment task in experiment.py. Experiment is the whole forward process of the experiment, and return loss. The example as following.

class E_basic(nn.Module):
    def __init__(self, predictor):
        super(E_basic, self).__init__()
        self.predictor = predictor
        gstate.clear_statics('number', 'loss', 'accuracy')

    def forward(self, x, t):
        y = self.predictor(x)
        loss = nn.CrossEntropyLoss()(y, t)
        accuracy = F.accuracy(y, t)
        gstate.summary(number=y.size(0), loss=loss.item(), accuracy=accuracy)
        return loss

Updater

If you have the special requirments different optimizers update the net, you can use updater keep some optimizers. The updater has the same interface as optim.

Trainer

Trainer is the encapsulation class which can extend extra functions.

Extensions

extensions.py contains many functions. The functions as following.

test
report_log
print_log
save_log
drop_lr
gs_best
save_best
save_trigger
print_best
basic_load

Train

combine all trainning processes

TopRust/Zcat