Keras variable length LSTM

keras_var_length.py

from __future__ import print_function

import numpy as np
from keras.layers import Dense
from keras.layers import LSTM
from keras.models import Sequential
from numpy.random import choice

USE_SEQUENCES = False
USE_STATELESS_MODEL = False

# [1, 0, 1, 0, 1, 0, 1, 0, 1, 0]

# you can all the four possible combinations
# USE_SEQUENCES and USE_STATELESS_MODEL

max_len = 20

N_train = 100
N_test = 10
N = N_train + N_test

cutoff = 0.8
template = np.array([1, 0] * max_len)
x = []
y = []
for i in range(N):
    min_max = choice(a=range(len(template)), size=2, replace=False)
    x_sub_sequence = template[min(min_max):max(min_max)]
    y_sub_sequence = 1 - x_sub_sequence
    x.append(x_sub_sequence)
    y.append(y_sub_sequence)

N = len(x)  # update N

X_train = x[:N_train]
X_test = x[N_train:]

y_train = y[:N_train]
y_test = y[N_train:]

# STATEFUL MODEL
print('Build STATEFUL model...')
model = Sequential()
model.add(LSTM(10,
               batch_input_shape=(1, 1, 1),
               return_sequences=False,
               stateful=True))
model.add(Dense(1, activation='sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

print('Train...')
for epoch in range(15):
    mean_tr_acc = []
    mean_tr_loss = []
    for i in range(len(X_train)):
        y_true = y_train[i]
        for j in range(len(X_train[i])):
            tr_loss, tr_acc = model.train_on_batch(np.reshape(X_train[i][j], (1, 1, 1)), np.array([y_true]))
            mean_tr_acc.append(tr_acc)
            mean_tr_loss.append(tr_loss)
        model.reset_states()

    print('accuracy training = {}'.format(np.mean(mean_tr_acc)))
    print('loss training = {}'.format(np.mean(mean_tr_loss)))
    print('___________________________________')

    mean_te_acc = []
    mean_te_loss = []
    for i in range(len(X_test)):
        for j in range(len(X_test[i])):
            te_loss, te_acc = model.test_on_batch(np.reshape(X_test[i][j], (1, 1, 1)), y_test[i])
            mean_te_acc.append(te_acc)
            mean_te_loss.append(te_loss)
        model.reset_states()

        for j in range(len(X_test[i])):
            y_pred = model.predict_on_batch(np.reshape(X_test[i][j], (1, 1, 1)))
        model.reset_states()

    print('accuracy testing = {}'.format(np.mean(mean_te_acc)))
    print('loss testing = {}'.format(np.mean(mean_te_loss)))
    print('___________________________________')

I wonder if you how to support M observations at each tilmestep, i.e. batch_input_shape=(1, 1, M) ?