lrcurve

Creates a learning-curve plot for Jupyter/Colab notebooks that is updated in real-time.

There is a framework agnostic interface lrcurve.PlotLearningCurve that works well with PyTorch and Tensorflow and a keras wrapper lrcurve.KerasLearningCurve that uses the keras callback interface.

lrcurve works with python 3.6 or newer and is distributed under the MIT license.

Install

pip install -U lrcurve

API

• lrcurve.PlotLearningCurve
• lrcurve.KerasLearningCurve

Examples

• Keras example
• Framework agnostic example
• PyTorch example

Keras example

from lrcurve import KerasLearningCurve

model.compile(optimizer=keras.optimizers.Adam(),
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=[keras.metrics.SparseCategoricalAccuracy()])

model.fit(train.x, train.y,
          epochs=100,
          verbose=0,
          validation_data=(validation.x, validation.y),
          callbacks=[KerasLearningCurve()])

Framework agnostic example

with PlotLearningCurve() as plot:
    for i in range(100):
        plot.append(i, {
            'loss': math.exp(-(i+1)/10),
            'val_loss': math.exp(-i/10)
        })
        plot.draw()
        time.sleep(0.1)

PyTorch example

from lrcurve import PlotLearningCurve

plot = PlotLearningCurve(
    mappings = {
        'loss': { 'line': 'train', 'facet': 'loss' },
        'val_loss': { 'line': 'validation', 'facet': 'loss' },
        'acc': { 'line': 'train', 'facet': 'acc' },
        'val_acc': { 'line': 'validation', 'facet': 'acc' }
    },
    facet_config = {
        'loss': { 'name': 'Cross-Entropy', 'limit': [0, None], 'scale': 'linear' },
        'acc': { 'name': 'Accuracy', 'limit': [0, 1], 'scale': 'linear' }
    },
    xaxis_config = { 'name': 'Epoch', 'limit': [0, 500] }
)

with plot:
    # optimize model
    for epoch in range(500):
        # compute loss
        z_test = network(x_test)
        loss_test = criterion(z_test, y_test)

        optimizer.zero_grad()
        z_train = network(x_train)
        loss_train = criterion(z_train, y_train)
        loss_train.backward()
        optimizer.step()

        # compute accuacy
        accuacy_test = sklearn.metrics.accuracy_score(torch.argmax(z_test, 1).numpy(), y_test)
        accuacy_train = sklearn.metrics.accuracy_score(torch.argmax(z_train, 1).numpy(), y_train)

        # append and update
        plot.append(epoch, {
            'loss': loss_train,
            'val_loss': loss_test,
            'acc': accuacy_train,
            'val_acc': accuacy_test
        })
        plot.draw()

Sponsor

Sponsored by NearForm Research.