reloading

A Python utility to reload a loop body from source on each iteration without losing state

Useful for editing source code during training of deep learning models. This lets you e.g. add logging, print statistics or save the model without restarting the training and, therefore, without losing the training progress.

Install

pip install reloading

Usage

To reload the body of a for loop from source before each iteration, simply wrap the iterator with reloading, e.g.

from reloading import reloading

for i in reloading(range(10)):
    # here could be your training loop
    print(i)

To reload a function from source before each execution, decorate the function definition with @reloading, e.g.

from reloading import reloading

@reloading
def some_function():
    pass

Examples

Here are the short snippets of how to use reloading with your favourite library. For complete examples, check out the examples folder.

PyTorch

for epoch in reloading(range(NB_EPOCHS)):
    # the code inside this outer loop will be reloaded before each epoch

    for images, targets in dataloader:
        optimiser.zero_grad()
        predictions = model(images)
        loss = F.cross_entropy(predictions, targets)
        loss.backward()
        optimiser.step()

Here is a full PyTorch example.

fastai

@reloading
def update_learner(learner):
    # this function will be reloaded from source before each epoch so that you
    # can make changes to the learner while the training is running
    pass

class LearnerUpdater(LearnerCallback):
    def on_epoch_begin(self, **kwargs):
        update_learner(self.learn)

path = untar_data(URLs.MNIST_SAMPLE)
data = ImageDataBunch.from_folder(path)
learn = cnn_learner(data, models.resnet18, metrics=accuracy, 
                    callback_fns=[LearnerUpdater])
learn.fit(10)

Here is a full fastai example.

Keras

@reloading
def update_model(model):
    # this function will be reloaded from source before each epoch so that you
    # can make changes to the model while the training is running using
    # K.set_value()
    pass

class ModelUpdater(Callback):
    def on_epoch_begin(self, epoch, logs=None):
        update_model(self.model)

model = Sequential()
model.add(Dense(64, activation='relu', input_dim=20))
model.add(Dense(10, activation='softmax'))

sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
              optimizer=sgd,
              metrics=['accuracy'])

model.fit(x_train, y_train,
          epochs=200,
          batch_size=128,
          callbacks=[ModelUpdater()])

Here is a full Keras example.

TensorFlow

for epoch in reloading(range(NB_EPOCHS)):
    # the code inside this outer loop will be reloaded from source
    # before each epoch so that you can change it during training
  
    train_loss.reset_states()
    train_accuracy.reset_states()
    test_loss.reset_states()
    test_accuracy.reset_states()
  
    for images, labels in tqdm(train_ds):
      train_step(images, labels)
  
    for test_images, test_labels in tqdm(test_ds):
      test_step(test_images, test_labels)

Here is a full TensorFlow example.