jaw.utils package

Submodules

jaw.utils.computation module

get_device()

Wrap the pytorch device selection.

Returns:

a torch device with GPU calculation if cuda is available, with CPU calculation else.

Ret type:

torch.device.

jaw.utils.progress_bar module

Simple progress bar. All credits to Kangliu : https://github.com/kuangliu/pytorch-cifar/blob/master/utils.py.

format_time(seconds)

Transform second into days, hours or minutes if needed. Format the time as follow : day, hours, min, sec, ms.

Parameters:

seconds (int.) – Seconds elapsed since the training launching.

Returns:

the string of the formated elapsed time.

progress_bar(current, total, msg=None)

Print a progress bar.

Note

Call this function inside the test() function. Example :

def test(model, loader, f_loss, optimizer, device):

    ...

    for i, (inputs, targets) in enumerate(loader):
        inputs, targets = inputs.to(device), targets.to(device)

        outputs = model(inputs)
        loss = f_loss(outputs, targets)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        N += inputs.shape[0]
        tot_loss += inputs.shape[0] * f_loss(outputs, targets).item()
        predicted_targets = outputs.argmax(dim=1)
        correct += (predicted_targets == targets).sum().item()

        progress_bar(i, len(loader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'
                 % (tot_loss/(i+1), 100.*correct/N, correct, N))

    return tot_loss/N, correct/N

Parameters:

• current (int.) – The position of the current example inside the dataset.

• total (int.) – Number of total examples inside the dataset.

• msg (str.) – Message to print aside the progress bar.

Returns:

the string of the formated elapsed time.

jaw.utils.tracking module

class ModelCheckpoint(filepath, model)

Utility class for saving a model when a training produce a new better.

update(loss)

Check if the current model is better than the previous saved. If true, overwrite the best model by the current one.

Parameters:

loss (float.) – Validation loss of the current model.

generate_unique_logpath(logdir, prefix)

Generate a unique log file for a new model saving.

Parameters:

• logdir (str.) – The path of the directory where the logs will be saved.

• prefix (str.) – Prefix name of the log file.