minimum working example for an issue with TimeDistributed's current compute_mask

distribute_mask_issues_mwe.py

from __future__ import print_function
from keras.layers import Input, Embedding, LSTM, Activation
from keras.layers import TimeDistributed as Distribute
import numpy as np

def outmask(x):
    return x._keras_history[0].outbound_nodes[0].output_masks[0]

def inmask(x):
    return x._keras_history[0].inbound_nodes[0].output_masks[0]

def pp(s,x):
    pstr = "For x={}\n\t x.ndim={}\n\t outmask.ndim={}\n\t inmask.ndim={}"
    print(pstr.format(s, x.ndim, outmask(x).ndim, inmask(x).ndim))

batch_shape = (10,10,10,10)
X_in = Input(batch_shape=batch_shape, dtype='int32')
X_emb = Embedding(input_dim=5,output_dim=10, mask_zero=True)(X_in)
X_reduced = Distribute(LSTM(10))(X_emb)
out = Activation('relu')(X_reduced)

print("==== input gets embedded =====")
pp("X_emb", X_emb)
print("==== a reduction from LSTM ====")
pp("X_reduction, ", X_reduced)