I'm trying to build a mnist autoencoder to learn how to work with reshaping and encoding.
The following error is thrown when i run the code:
ValueError: total size of new array must be unchanged, input_shape = [748], output_shape = [28, 28]
The code looks like this:
import tensorflow as tf
import tensorflow.keras as keras
from tensorflow.keras import layers
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255., x_test / 255.
enc_in = tf.keras.Input(shape=(28, 28))
x = layers.Flatten()(enc_in)
x = layers.Dense(128, activation='relu', name='enc_dense_1')(x)
enc_out = layers.Dense(10, activation='softmax', name='enc_out')(x)
x = layers.Dense(128, activation='relu', name='dec_dense_1')(enc_out)
x = layers.Dense(748, activation='relu', name='dec_dense_2')(x)
dec_out = layers.Reshape(target_shape=(28, 28))(x)
autoencoder = keras.Model(inputs=enc_in, outputs=dec_out)
autoencoder.compile(
optimizer='adam',
loss=keras.losses.CategoricalCrossentropy(),
metrics=['accuracy']
)
print(autoencoder.summary())
autoencoder.fit(x_train, x_train, batch_size=64, epochs=10, validation_split=0.2)
Why is tensorflow complaining about the output shape while my keras.Input shape is (28, 28) ?
The error is quite self-explanatory, you made a small mistake.
#28 * 28 = 784
x = layers.Dense(748, activation='relu', name='dec_dense_2')(x)
dec_out = layers.Reshape(target_shape=(28, 28))(x)
You accidentally wrote 748 instead of 784, and of course the reshape cannot take place due to this error.