def create_model():
    model_in = Input(shape=(1,), batch_shape=(1, 1, 900), name='seq_model_in')
    lstm_layer = LSTM(1, batch_input_shape=(1, 1, 900), stateful=stateful)(model_in)
    dense_final = Dense(900, activation='sigmoid')(lstm_layer)
    model = Model(inputs=model_in, outputs=dense_final)
    return model

model = create_model()
model.compile(loss='binary_crossentropy', optimizer='adam')

#Training loop
for epoch in range(num_epochs):
    for i in range(num_train_examples):
        for j in range(sequence_length):
            tr_loss = model.train_on_batch(train[i][j], y_true)
            #Accumulate the training loss, etc.

#This has to happen instead of the spot where I call create_model() and model.compile
if num_gpus <= 1
    print("training with 1 GPU")
    model = create_model()
else:
    print("training with multiple GPUs")
    with tf.device("/cpu:0"):
        model = create_model()
    model = multi_gpu_model(model, gpus=num_gpus)
model.compile(loss='binary_crossentropy', optimizer='adam')