Save the date! Google I/O returns May 18-20 Register now

Training a neural network on MNIST with Keras

This simple example demonstrate how to plug TFDS into a Keras model.

View on TensorFlow.org View source on GitHub Download notebook
import tensorflow as tf
import tensorflow_datasets as tfds

Step 1: Create your input pipeline

Build efficient input pipeline using advices from:

Load MNIST

Load with the following arguments:

  • shuffle_files: The MNIST data is only stored in a single file, but for larger datasets with multiple files on disk, it's good practice to shuffle them when training.
  • as_supervised: Returns tuple (img, label) instead of dict {'image': img, 'label': label}
(ds_train, ds_test), ds_info = tfds.load(
    'mnist',
    split=['train', 'test'],
    shuffle_files=True,
    as_supervised=True,
    with_info=True,
)

Build training pipeline

Apply the following transormations:

  • ds.map: TFDS provide the images as tf.uint8, while the model expect tf.float32, so normalize images
  • ds.cache As the dataset fit in memory, cache before shuffling for better performance.
    Note: Random transformations should be applied after caching
  • ds.shuffle: For true randomness, set the shuffle buffer to the full dataset size.
    Note: For bigger datasets which do not fit in memory, a standard value is 1000 if your system allows it.
  • ds.batch: Batch after shuffling to get unique batches at each epoch.
  • ds.prefetch: Good practice to end the pipeline by prefetching for performances.
def normalize_img(image, label):
  """Normalizes images: `uint8` -> `float32`."""
  return tf.cast(image, tf.float32) / 255., label

ds_train = ds_train.map(
    normalize_img, num_parallel_calls=tf.data.experimental.AUTOTUNE)
ds_train = ds_train.cache()
ds_train = ds_train.shuffle(ds_info.splits['train'].num_examples)
ds_train = ds_train.batch(128)
ds_train = ds_train.prefetch(tf.data.experimental.AUTOTUNE)

Build evaluation pipeline

Testing pipeline is similar to the training pipeline, with small differences:

  • No ds.shuffle() call
  • Caching is done after batching (as batches can be the same between epoch)
ds_test = ds_test.map(
    normalize_img, num_parallel_calls=tf.data.experimental.AUTOTUNE)
ds_test = ds_test.batch(128)
ds_test = ds_test.cache()
ds_test = ds_test.prefetch(tf.data.experimental.AUTOTUNE)

Step 2: Create and train the model

Plug the input pipeline into Keras.

model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(input_shape=(28, 28)),
  tf.keras.layers.Dense(128,activation='relu'),
  tf.keras.layers.Dense(10)
])
model.compile(
    optimizer=tf.keras.optimizers.Adam(0.001),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()],
)

model.fit(
    ds_train,
    epochs=6,
    validation_data=ds_test,
)
Epoch 1/6
469/469 [==============================] - 4s 4ms/step - loss: 0.6240 - sparse_categorical_accuracy: 0.8288 - val_loss: 0.2043 - val_sparse_categorical_accuracy: 0.9424
Epoch 2/6
469/469 [==============================] - 1s 2ms/step - loss: 0.1796 - sparse_categorical_accuracy: 0.9499 - val_loss: 0.1395 - val_sparse_categorical_accuracy: 0.9598
Epoch 3/6
469/469 [==============================] - 1s 2ms/step - loss: 0.1215 - sparse_categorical_accuracy: 0.9642 - val_loss: 0.1137 - val_sparse_categorical_accuracy: 0.9678
Epoch 4/6
469/469 [==============================] - 1s 2ms/step - loss: 0.0968 - sparse_categorical_accuracy: 0.9724 - val_loss: 0.0974 - val_sparse_categorical_accuracy: 0.9707
Epoch 5/6
469/469 [==============================] - 1s 2ms/step - loss: 0.0774 - sparse_categorical_accuracy: 0.9775 - val_loss: 0.0852 - val_sparse_categorical_accuracy: 0.9766
Epoch 6/6
469/469 [==============================] - 1s 2ms/step - loss: 0.0631 - sparse_categorical_accuracy: 0.9811 - val_loss: 0.0868 - val_sparse_categorical_accuracy: 0.9735
<tensorflow.python.keras.callbacks.History at 0x7f70782baa58>