minimal failure cases, only on tensorflow backend

fail.py

from keras.layers import Input, Dense
from keras.models import Model, Sequential
from keras.datasets import mnist
from keras.layers.normalization import BatchNormalization as BN

autoencoder1 = Sequential([
    Dense(128, activation='relu',input_shape=(784,)),
    BN(),
    Dense(784, activation='relu'),
])

### fails on tensorflow, both CPU or GPU
### it works w/o problem with theano

x = Input(shape=(784,))
y = autoencoder1(x)
autoencoder2 = Model(input=x,output=y)
autoencoder = autoencoder2

autoencoder.compile(optimizer='rmsprop', loss='mse')


################################################################

# train the VAE on MNIST digits
(x_train, _), (x_test, y_test) = mnist.load_data()

x_train = x_train.astype('float32') / 255.
print(x_train.shape)
x_train = x_train.reshape((x_train.shape[0],784))
x_test = x_test.astype('float32') / 255.
x_test = x_test.reshape((x_test.shape[0],784))

print('x_train.shape:', x_train.shape)

from keras.callbacks import CSVLogger, ReduceLROnPlateau, EarlyStopping

autoencoder.fit(x_train, x_train,
        shuffle=True,
        nb_epoch=1,
        batch_size=256,
        validation_data=(x_test, x_test),
        callbacks=[CSVLogger("vae-conv-deconv/loss.csv"),
                   # EarlyStopping(patience=6,verbose=1,mode='min'),
                   ReduceLROnPlateau(verbose=1,patience=20,factor=0.5,mode='min',epsilon=0.0001)
        ])

h5 = "ae.h5"
autoencoder.save(h5)

del autoencoder

import keras.models

autoencoder = keras.models.load_model(h5)
autoencoder.summary()

reuse-variables-fail-load.py

from keras.layers import Input, Dense
from keras.models import Model, Sequential
from keras.datasets import mnist
from keras.layers.normalization import BatchNormalization as BN

autoencoder1 = Sequential([
    Dense(128, activation='relu',input_shape=(784,)),
    BN(),
    Dense(784, activation='relu'),
])

### with reuse_variables, construction is ok

import tensorflow as tf
x = Input(shape=(784,))
tf.get_variable_scope().reuse_variables()
y = autoencoder1(x)
autoencoder2 = Model(input=x,output=y)
autoencoder = autoencoder2

autoencoder.compile(optimizer='rmsprop', loss='mse')


################################################################

# train the VAE on MNIST digits
(x_train, _), (x_test, y_test) = mnist.load_data()

x_train = x_train.astype('float32') / 255.
print(x_train.shape)
x_train = x_train.reshape((x_train.shape[0],784))
x_test = x_test.astype('float32') / 255.
x_test = x_test.reshape((x_test.shape[0],784))

print('x_train.shape:', x_train.shape)

from keras.callbacks import CSVLogger, ReduceLROnPlateau, EarlyStopping

autoencoder.fit(x_train, x_train,
        shuffle=True,
        nb_epoch=1,
        batch_size=256,
        validation_data=(x_test, x_test),
        callbacks=[CSVLogger("vae-conv-deconv/loss.csv"),
                   # EarlyStopping(patience=6,verbose=1,mode='min'),
                   ReduceLROnPlateau(verbose=1,patience=20,factor=0.5,mode='min',epsilon=0.0001)
        ])

h5 = "ae.h5"
autoencoder.save(h5)

del autoencoder

import keras.models

### fails to load, regardless of reuse_variables
# tf.get_variable_scope().reuse_variables()
autoencoder = keras.models.load_model(h5)
autoencoder.summary()

work.py

from keras.layers import Input, Dense
from keras.models import Model, Sequential
from keras.datasets import mnist
from keras.layers.normalization import BatchNormalization as BN

autoencoder1 = Sequential([
    Dense(128, activation='relu',input_shape=(784,)),
    BN(),
    Dense(784, activation='relu'),
])

# this work
autoencoder = autoencoder1

autoencoder.compile(optimizer='rmsprop', loss='mse')


################################################################

# train the VAE on MNIST digits
(x_train, _), (x_test, y_test) = mnist.load_data()

x_train = x_train.astype('float32') / 255.
print(x_train.shape)
x_train = x_train.reshape((x_train.shape[0],784))
x_test = x_test.astype('float32') / 255.
x_test = x_test.reshape((x_test.shape[0],784))

print('x_train.shape:', x_train.shape)

from keras.callbacks import CSVLogger, ReduceLROnPlateau, EarlyStopping

autoencoder.fit(x_train, x_train,
        shuffle=True,
        nb_epoch=1,
        batch_size=256,
        validation_data=(x_test, x_test),
        callbacks=[CSVLogger("vae-conv-deconv/loss.csv"),
                   # EarlyStopping(patience=6,verbose=1,mode='min'),
                   ReduceLROnPlateau(verbose=1,patience=20,factor=0.5,mode='min',epsilon=0.0001)
        ])

h5 = "ae.h5"
autoencoder.save(h5)

del autoencoder

import keras.models

autoencoder = keras.models.load_model(h5)
autoencoder.summary()