tests_keras.py

from __future__ import division

from .tests_tqdm import importorskip, mark

pytestmark = mark.slow


@mark.filterwarnings("ignore:.*:DeprecationWarning")
def test_keras(capsys):
    """Test tqdm.keras.TqdmCallback"""
    TqdmCallback = importorskip('tqdm.keras').TqdmCallback
    np = importorskip('numpy')
    try:
        import keras as K
    except ImportError:
        K = importorskip('tensorflow.keras')

    # 1D autoencoder
    dtype = np.float32
    model = K.models.Sequential([
        K.layers.InputLayer((1, 1), dtype=dtype), K.layers.Conv1D(1, 1)])
    model.compile("adam", "mse")
    x = np.random.rand(100, 1, 1).astype(dtype)
    batch_size = 10
    batches = len(x) / batch_size
    epochs = 5

    # just epoch (no batch) progress
    model.fit(
        x,
        x,
        epochs=epochs,
        batch_size=batch_size,
        verbose=False,
        callbacks=[
            TqdmCallback(
                epochs,
                desc="training",
                data_size=len(x),
                batch_size=batch_size,
                verbose=0)])
    _, res = capsys.readouterr()
    assert "training: " in res
    assert "{epochs}/{epochs}".format(epochs=epochs) in res
    assert "{batches}/{batches}".format(batches=batches) not in res

    # full (epoch and batch) progress
    model.fit(
        x,
        x,
        epochs=epochs,
        batch_size=batch_size,
        verbose=False,
        callbacks=[
            TqdmCallback(
                epochs,
                desc="training",
                data_size=len(x),
                batch_size=batch_size,
                verbose=2)])
    _, res = capsys.readouterr()
    assert "training: " in res
    assert "{epochs}/{epochs}".format(epochs=epochs) in res
    assert "{batches}/{batches}".format(batches=batches) in res

    # auto-detect epochs and batches
    model.fit(
        x,
        x,
        epochs=epochs,
        batch_size=batch_size,
        verbose=False,
        callbacks=[TqdmCallback(desc="training", verbose=2)])
    _, res = capsys.readouterr()
    assert "training: " in res
    assert "{epochs}/{epochs}".format(epochs=epochs) in res
    assert "{batches}/{batches}".format(batches=batches) in res

    # continue training (start from epoch != 0)
    initial_epoch = 3
    model.fit(
        x,
        x,
        initial_epoch=initial_epoch,
        epochs=epochs,
        batch_size=batch_size,
        verbose=False,
        callbacks=[TqdmCallback(desc="training", verbose=0,
                                miniters=1, mininterval=0, maxinterval=0)])
    _, res = capsys.readouterr()
    assert "training: " in res
    assert "{epochs}/{epochs}".format(epochs=initial_epoch - 1) not in res
    assert "{epochs}/{epochs}".format(epochs=epochs) in res