Coursera
Ungraded Lab: Denoising with a CNN Autoencoder
In the final lab for this week, you will introduce noise to the Fashion MNIST dataset and train an autoencoder to reconstruct the original input 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 train and test sets a little differently this time. Instead of just normalizing the images, you will also introduce random noise and the generated images will be used as input to your model. The target or label will still be the clean images.
def map_image_with_noise(image, label):
'''Normalizes the images and generates noisy inputs.'''
image = tf.cast(image, dtype=tf.float32)
image = image / 255.0
noise_factor = 0.5
factor = noise_factor * tf.random.normal(shape=image.shape)
image_noisy = image + factor
image_noisy = tf.clip_by_value(image_noisy, 0.0, 1.0)
return image_noisy, image
BATCH_SIZE = 128
SHUFFLE_BUFFER_SIZE = 1024
train_dataset = tfds.load('fashion_mnist', as_supervised=True, split="train")
train_dataset = train_dataset.map(map_image_with_noise)
train_dataset = train_dataset.shuffle(SHUFFLE_BUFFER_SIZE).batch(BATCH_SIZE).repeat()
test_dataset = tfds.load('fashion_mnist', as_supervised=True, split="test")
test_dataset = test_dataset.map(map_image_with_noise)
test_dataset = test_dataset.batch(BATCH_SIZE).repeat()
Downloading and preparing dataset 29.45 MiB (download: 29.45 MiB, generated: 36.42 MiB, total: 65.87 MiB) to /root/tensorflow_datasets/fashion_mnist/3.0.1...
Dl Completed...: 0 url [00:00, ? url/s]
Dl Size...: 0 MiB [00:00, ? MiB/s]
Extraction completed...: 0 file [00:00, ? file/s]
Generating splits...: 0%| | 0/2 [00:00<?, ? splits/s]
Generating train examples...: 0%| | 0/60000 [00:00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/fashion_mnist/3.0.1.incompleteFF8MFS/fashion_mnist-train.tfrecord*...: 0…
Generating test examples...: 0%| | 0/10000 [00:00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/fashion_mnist/3.0.1.incompleteFF8MFS/fashion_mnist-test.tfrecord*...: 0%…
Dataset fashion_mnist downloaded and prepared to /root/tensorflow_datasets/fashion_mnist/3.0.1. Subsequent calls will reuse this data.
Build the Model
You will use the same model from the previous lab.
def encoder(inputs):
'''Defines the encoder with two Conv2D and max pooling layers.'''
conv_1 = tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), activation='relu', padding='same')(inputs)
max_pool_1 = tf.keras.layers.MaxPooling2D(pool_size=(2,2))(conv_1)
conv_2 = tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), activation='relu', padding='same')(max_pool_1)
max_pool_2 = tf.keras.layers.MaxPooling2D(pool_size=(2,2))(conv_2)
return max_pool_2
def bottle_neck(inputs):
'''Defines the bottleneck.'''
bottle_neck = tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), activation='relu', padding='same')(inputs)
encoder_visualization = tf.keras.layers.Conv2D(filters=1, kernel_size=(3,3), activation='sigmoid', padding='same')(bottle_neck)
return bottle_neck, encoder_visualization
def decoder(inputs):
'''Defines the decoder path to upsample back to the original image size.'''
conv_1 = tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), activation='relu', padding='same')(inputs)
up_sample_1 = tf.keras.layers.UpSampling2D(size=(2,2))(conv_1)
conv_2 = tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), activation='relu', padding='same')(up_sample_1)
up_sample_2 = tf.keras.layers.UpSampling2D(size=(2,2))(conv_2)
conv_3 = tf.keras.layers.Conv2D(filters=1, kernel_size=(3,3), activation='sigmoid', padding='same')(up_sample_2)
return conv_3
def convolutional_auto_encoder():
'''Builds the entire autoencoder model.'''
inputs = tf.keras.layers.Input(shape=(28, 28, 1,))
encoder_output = encoder(inputs)
bottleneck_output, encoder_visualization = bottle_neck(encoder_output)
decoder_output = decoder(bottleneck_output)
model = tf.keras.Model(inputs =inputs, outputs=decoder_output)
encoder_model = tf.keras.Model(inputs=inputs, outputs=encoder_visualization)
return model, encoder_model
convolutional_model, convolutional_encoder_model = convolutional_auto_encoder()
convolutional_model.summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 28, 28, 1)] 0
conv2d (Conv2D) (None, 28, 28, 64) 640
max_pooling2d (MaxPooling2 (None, 14, 14, 64) 0
D)
conv2d_1 (Conv2D) (None, 14, 14, 128) 73856
max_pooling2d_1 (MaxPoolin (None, 7, 7, 128) 0
g2D)
conv2d_2 (Conv2D) (None, 7, 7, 256) 295168
conv2d_4 (Conv2D) (None, 7, 7, 128) 295040
up_sampling2d (UpSampling2 (None, 14, 14, 128) 0
D)
conv2d_5 (Conv2D) (None, 14, 14, 64) 73792
up_sampling2d_1 (UpSamplin (None, 28, 28, 64) 0
g2D)
conv2d_6 (Conv2D) (None, 28, 28, 1) 577
=================================================================
Total params: 739073 (2.82 MB)
Trainable params: 739073 (2.82 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
Compile and Train the Model
train_steps = 60000 // BATCH_SIZE
valid_steps = 60000 // BATCH_SIZE
convolutional_model.compile(optimizer=tf.keras.optimizers.Adam(), loss='binary_crossentropy')
conv_model_history = convolutional_model.fit(train_dataset, steps_per_epoch=train_steps, validation_data=test_dataset, validation_steps=valid_steps, epochs=40)
Epoch 1/40
468/468 [==============================] - 29s 37ms/step - loss: 0.3213 - val_loss: 0.3017
Epoch 2/40
468/468 [==============================] - 14s 28ms/step - loss: 0.2954 - val_loss: 0.2946
Epoch 3/40
468/468 [==============================] - 13s 28ms/step - loss: 0.2911 - val_loss: 0.2915
Epoch 4/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2884 - val_loss: 0.2892
Epoch 5/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2863 - val_loss: 0.2876
Epoch 6/40
468/468 [==============================] - 14s 30ms/step - loss: 0.2852 - val_loss: 0.2868
Epoch 7/40
468/468 [==============================] - 13s 28ms/step - loss: 0.2842 - val_loss: 0.2857
Epoch 8/40
468/468 [==============================] - 13s 28ms/step - loss: 0.2833 - val_loss: 0.2850
Epoch 9/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2827 - val_loss: 0.2847
Epoch 10/40
468/468 [==============================] - 14s 30ms/step - loss: 0.2823 - val_loss: 0.2846
Epoch 11/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2818 - val_loss: 0.2835
Epoch 12/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2814 - val_loss: 0.2831
Epoch 13/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2812 - val_loss: 0.2831
Epoch 14/40
468/468 [==============================] - 12s 27ms/step - loss: 0.2808 - val_loss: 0.2825
Epoch 15/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2805 - val_loss: 0.2826
Epoch 16/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2802 - val_loss: 0.2824
Epoch 17/40
468/468 [==============================] - 12s 27ms/step - loss: 0.2801 - val_loss: 0.2824
Epoch 18/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2799 - val_loss: 0.2819
Epoch 19/40
468/468 [==============================] - 12s 27ms/step - loss: 0.2798 - val_loss: 0.2817
Epoch 20/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2795 - val_loss: 0.2818
Epoch 21/40
468/468 [==============================] - 12s 27ms/step - loss: 0.2794 - val_loss: 0.2818
Epoch 22/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2792 - val_loss: 0.2817
Epoch 23/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2792 - val_loss: 0.2825
Epoch 24/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2791 - val_loss: 0.2813
Epoch 25/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2790 - val_loss: 0.2813
Epoch 26/40
468/468 [==============================] - 14s 29ms/step - loss: 0.2790 - val_loss: 0.2813
Epoch 27/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2787 - val_loss: 0.2810
Epoch 28/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2787 - val_loss: 0.2810
Epoch 29/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2787 - val_loss: 0.2808
Epoch 30/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2785 - val_loss: 0.2808
Epoch 31/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2785 - val_loss: 0.2810
Epoch 32/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2785 - val_loss: 0.2809
Epoch 33/40
468/468 [==============================] - 12s 27ms/step - loss: 0.2783 - val_loss: 0.2806
Epoch 34/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2783 - val_loss: 0.2805
Epoch 35/40
468/468 [==============================] - 12s 27ms/step - loss: 0.2783 - val_loss: 0.2807
Epoch 36/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2781 - val_loss: 0.2807
Epoch 37/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2782 - val_loss: 0.2806
Epoch 38/40
468/468 [==============================] - 13s 27ms/step - loss: 0.2781 - val_loss: 0.2803
Epoch 39/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2781 - val_loss: 0.2807
Epoch 40/40
468/468 [==============================] - 12s 26ms/step - loss: 0.2779 - val_loss: 0.2803
Display sample results
Let’s see if the model can generate the clean image from noisy inputs.
def display_one_row(disp_images, offset, shape=(28, 28)):
'''Display sample outputs in one row.'''
for idx, noisy_image in enumerate(disp_images):
plt.subplot(3, 10, offset + idx + 1)
plt.xticks([])
plt.yticks([])
noisy_image = np.reshape(noisy_image, shape)
plt.imshow(noisy_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 indices
idxs = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
# prepare test samples as a batch of 10 images
conv_output_samples = np.array(output_samples[idxs])
conv_output_samples = np.reshape(conv_output_samples, (10, 28, 28, 1))
# get the encoder ouput
encoded = convolutional_encoder_model.predict(conv_output_samples)
# get a prediction for some values in the dataset
predicted = convolutional_model.predict(conv_output_samples)
# display the samples, encodings and decoded values!
display_results(conv_output_samples, encoded, predicted, enc_shape=(7,7))
1/1 [==============================] - 0s 239ms/step
1/1 [==============================] - 0s 175ms/step
