Coursera

Ungraded Lab: MNIST Deep Autoencoder

Welcome back! In this lab, you will extend the shallow autoencoder you built in the previous exercise. The model here will have a deeper network so it can handle more complex images.

Imports

try:
  # %tensorflow_version only exists in Colab.
  %tensorflow_version 2.x
except Exception:
  pass

import tensorflow as tf
import tensorflow_datasets as tfds

import numpy as np
import matplotlib.pyplot as plt

Colab only includes TensorFlow 2.x; %tensorflow_version has no effect.

Prepare the Dataset

You will prepare the MNIST dataset just like in the previous lab.

def map_image(image, label):
  '''Normalizes and flattens the image. Returns image as input and label.'''
  image = tf.cast(image, dtype=tf.float32)
  image = image / 255.0
  image = tf.reshape(image, shape=(784,))

  return image, image

# Load the train and test sets from TFDS

BATCH_SIZE = 128
SHUFFLE_BUFFER_SIZE = 1024

train_dataset = tfds.load('mnist', as_supervised=True, split="train")
train_dataset = train_dataset.map(map_image)
train_dataset = train_dataset.shuffle(SHUFFLE_BUFFER_SIZE).batch(BATCH_SIZE).repeat()

test_dataset = tfds.load('mnist', as_supervised=True, split="test")
test_dataset = test_dataset.map(map_image)
test_dataset = test_dataset.batch(BATCH_SIZE).repeat()

Downloading and preparing dataset 11.06 MiB (download: 11.06 MiB, generated: 21.00 MiB, total: 32.06 MiB) to /root/tensorflow_datasets/mnist/3.0.1...



Dl Completed...:   0%|          | 0/5 [00:00<?, ? file/s]


Dataset mnist downloaded and prepared to /root/tensorflow_datasets/mnist/3.0.1. Subsequent calls will reuse this data.

Build the Model

As mentioned, you will have a deeper network for the autoencoder. Compare the layers here with that of the shallow network you built in the previous lab.

def deep_autoencoder():
  '''Builds the encoder and decoder using Dense layers.'''
  encoder = tf.keras.layers.Dense(units=128, activation='relu')(inputs)
  encoder = tf.keras.layers.Dense(units=64, activation='relu')(encoder)
  encoder = tf.keras.layers.Dense(units=32, activation='relu')(encoder)

  decoder = tf.keras.layers.Dense(units=64, activation='relu')(encoder)
  decoder = tf.keras.layers.Dense(units=128, activation='relu')(decoder)
  decoder = tf.keras.layers.Dense(units=784, activation='sigmoid')(decoder)

  return encoder, decoder

# set the input tensor
inputs =  tf.keras.layers.Input(shape=(784,))

# get the encoder and decoder output
deep_encoder_output, deep_autoencoder_output = deep_autoencoder()

# setup the encoder because you will visualize its output later
deep_encoder_model = tf.keras.Model(inputs=inputs, outputs=deep_encoder_output)

# setup the autoencoder
deep_autoencoder_model = tf.keras.Model(inputs=inputs, outputs=deep_autoencoder_output)

Compile and Train the Model

train_steps = 60000 // BATCH_SIZE

deep_autoencoder_model.compile(optimizer=tf.keras.optimizers.Adam(), loss='binary_crossentropy')
deep_auto_history = deep_autoencoder_model.fit(train_dataset, steps_per_epoch=train_steps, epochs=50)

Epoch 1/50
468/468 [==============================] - 22s 23ms/step - loss: 0.1973
Epoch 2/50
468/468 [==============================] - 6s 14ms/step - loss: 0.1278
Epoch 3/50
468/468 [==============================] - 6s 12ms/step - loss: 0.1168
Epoch 4/50
468/468 [==============================] - 6s 12ms/step - loss: 0.1100
Epoch 5/50
468/468 [==============================] - 4s 9ms/step - loss: 0.1059
Epoch 6/50
468/468 [==============================] - 6s 12ms/step - loss: 0.1027
Epoch 7/50
468/468 [==============================] - 4s 9ms/step - loss: 0.1002
Epoch 8/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0982
Epoch 9/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0964
Epoch 10/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0949
Epoch 11/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0937
Epoch 12/50
468/468 [==============================] - 6s 12ms/step - loss: 0.0928
Epoch 13/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0918
Epoch 14/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0911
Epoch 15/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0904
Epoch 16/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0898
Epoch 17/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0891
Epoch 18/50
468/468 [==============================] - 4s 10ms/step - loss: 0.0885
Epoch 19/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0879
Epoch 20/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0874
Epoch 21/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0869
Epoch 22/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0866
Epoch 23/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0862
Epoch 24/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0860
Epoch 25/50
468/468 [==============================] - 5s 11ms/step - loss: 0.0856
Epoch 26/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0855
Epoch 27/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0853
Epoch 28/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0850
Epoch 29/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0848
Epoch 30/50
468/468 [==============================] - 3s 7ms/step - loss: 0.0847
Epoch 31/50
468/468 [==============================] - 5s 11ms/step - loss: 0.0846
Epoch 32/50
468/468 [==============================] - 4s 7ms/step - loss: 0.0844
Epoch 33/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0843
Epoch 34/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0842
Epoch 35/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0841
Epoch 36/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0840
Epoch 37/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0838
Epoch 38/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0837
Epoch 39/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0837
Epoch 40/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0836
Epoch 41/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0835
Epoch 42/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0834
Epoch 43/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0833
Epoch 44/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0833
Epoch 45/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0832
Epoch 46/50
468/468 [==============================] - 5s 10ms/step - loss: 0.0831
Epoch 47/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0830
Epoch 48/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0829
Epoch 49/50
468/468 [==============================] - 4s 8ms/step - loss: 0.0829
Epoch 50/50
468/468 [==============================] - 4s 9ms/step - loss: 0.0828

Display sample results

See the results using the model you just trained.

def display_one_row(disp_images, offset, shape=(28, 28)):
  '''Display sample outputs in one row.'''
  for idx, test_image in enumerate(disp_images):
    plt.subplot(3, 10, offset + idx + 1)
    plt.xticks([])
    plt.yticks([])
    test_image = np.reshape(test_image, shape)
    plt.imshow(test_image, cmap='gray')


def display_results(disp_input_images, disp_encoded, disp_predicted, enc_shape=(8,4)):
  '''Displays the input, encoded, and decoded output values.'''
  plt.figure(figsize=(15, 5))
  display_one_row(disp_input_images, 0, shape=(28,28,))
  display_one_row(disp_encoded, 10, shape=enc_shape)
  display_one_row(disp_predicted, 20, shape=(28,28,))

# take 1 batch of the dataset
test_dataset = test_dataset.take(1)

# take the input images and put them in a list
output_samples = []
for input_image, image in tfds.as_numpy(test_dataset):
      output_samples = input_image

# pick 10 random numbers to be used as indices to the list above
idxs = np.random.choice(BATCH_SIZE, size=10)

# get the encoder output
encoded_predicted = deep_encoder_model.predict(test_dataset)

# get a prediction for the test batch
deep_predicted = deep_autoencoder_model.predict(test_dataset)

# display the 10 samples, encodings and decoded values!
display_results(output_samples[idxs], encoded_predicted[idxs], deep_predicted[idxs])

1/1 [==============================] - 0s 180ms/step
1/1 [==============================] - 0s 113ms/step