reset_states() failure in a stateful network with initial_states set and training in batch - TypeError: 'NoneType' object is not subscriptable

Following is the stacktrace -

Traceback (most recent call last):
  File "/usr/local/Cellar/python/3.6.5_1/Frameworks/Python.framework/Versions/3.6/lib/python3.6/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/usr/local/Cellar/python/3.6.5_1/Frameworks/Python.framework/Versions/3.6/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/Users/manoj/Desktop/repos/yane/yane/LSTM/manytomanyLSTM.py", line 137, in <module>
    incremental_train(space)
  File "/Users/manoj/Desktop/repos/yane/yane/LSTM/manytomanyLSTM.py", line 128, in incremental_train
    model.reset_states()
  File "/usr/local/lib/python3.6/site-packages/keras/engine/topology.py", line 1968, in reset_states
    layer.reset_states()
  File "/usr/local/lib/python3.6/site-packages/keras/layers/recurrent.py", line 681, in reset_states
    batch_size = self.input_spec[0].shape[0]        
TypeError: 'NoneType' object is not subscriptable

[huan]

I ran into this issue today with TF 2.0 preview.

Raised an issue at https://github.com/tensorflow/tensorflow/issues/25985

[manojrege]

Steps to reproduce: If the initial_states is removed, the code works.

from keras.models import Input, Model
from keras.layers import Dense, LSTM

epochs = 1
trainY = np.random.rand(32 * 32, 1)

interleaved_trainX = np.random.rand(32 * 32, 1, 1)
interleaved_trainY = np.random.rand(32 * 32, 1, 1)

encoder_inputs = Input(batch_shape=(32, 1, 1), name='ENCODER_INPUT')
encoder = LSTM(256,
               return_sequences=True,
               stateful=True,
               name='LSTM',
               return_state=True)
encoder_outputs, state_h, state_c = encoder(encoder_inputs)
encoder_states = [state_h, state_c]

decoder_inputs = Input(batch_shape=(32, 1, 1),
                       name='DECODER_INPUT')
decoder = LSTM(256,
               return_sequences=True,
               stateful=True,
               name='DECODER_LSTM')
decoder_out = decoder(decoder_inputs, initial_state=encoder_states)
decoder_2 = LSTM(256,
                 return_sequences=False,
                 stateful=True,
                 batch_size=32,
                 name='DECODER_LSTM2')(decoder_out)
decoder_outputs = Dense(1, name='DENSE')(decoder_2)

model = Model([encoder_inputs, decoder_inputs], decoder_outputs)

model.compile(loss='mse', optimizer='adam', metrics=['mean_absolute_percentage_error'])
for i in range(epochs):
    print('Epoch', i, '/', epochs)
    model.fit([interleaved_trainX, interleaved_trainY], trainY,
              batch_size=32,
              validation_data=([interleaved_trainX, interleaved_trainY], trainY),
              verbose=1,
              epochs=1,
              shuffle=False)
    model.reset_states()

loss, mape = model.evaluate([interleaved_trainX, interleaved_trainY],
                            trainY,
                            batch_size=32,
                            verbose=1)

print('Test accuracy:', loss)
print('Test MAPE:', mape)