Recreates the bug with multi-GPU TensorFlow in Kur

parallel_bug.py

import keras.models as M
import keras.layers as L
from kur.utils.parallelism import make_parallel

# Pretend to have some 32 x 32 images.
input = x = L.Input(shape=(32, 32))

# Shape: (samples, 32, 32)

x = L.TimeDistributed(
	L.Dense(100)
)(x)

# Shape: (samples, 32, 100)

x = L.LSTM(50, return_sequences=True)(x)

# Shape: (samples, 32, 50)

x = L.TimeDistributed(
	L.Dense(20)
)(x)

# Shape: (samples, 32, 20)

# Compile the model.
model = M.Model(input=input, output=x)
model.summary()

# Try to parallelize it.
# It will fail for any value of `gpu_count`.
model = make_parallel(model, 1)