For reproducing keras fit_generator bug

test_sequence.py

from keras.utils import Sequence
from keras.models import Sequential
from keras.layers import Dense
from keras.utils import to_categorical
import pybedtools as pbt
import numpy as np
import random as rn
import keras
import tensorflow
import os

class DummySequence(Sequence):

    def __init__(self, x_set, y_set, batch_size):
        self.x, self.y = x_set, y_set
        self.batch_size = batch_size

    def __len__(self):
        return int(np.ceil(len(self.x) / float(self.batch_size)))

    def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) * self.batch_size]

        return np.array(batch_x), np.array(batch_y)

    def on_epoch_end(self):
        a = pbt.example_bedtool('a.bed')
        b = pbt.example_bedtool('b.bed')
        c = pbt.BedTool(a.cat(b))

if __name__ == '__main__':
    seed = 1337
    np.random.seed(seed)
    rn.seed(seed)
    tensorflow.set_random_seed(seed)

    x = np.random.random((100, 3))
    y = to_categorical(np.random.random(100) > .5).astype(int)

    x2 = np.random.random((100, 3))
    y2 = to_categorical(np.random.random(100) > .5).astype(int)

    seq = DummySequence(x, y, 10)
    seq2 = DummySequence(x2, y2, 10)
    import pickle
    f = open('temp.pkl','wb')
    pickle.dump(seq, f)
    model = Sequential()
    model.add(Dense(32, input_dim=3))
    model.add(Dense(2, activation='softmax'))
    model.compile(optimizer='rmsprop',
                  loss='categorical_crossentropy',
                  metrics=['accuracy'])

    output_dir = 'temp'
    callbacks = [
        keras.callbacks.TensorBoard(log_dir=output_dir,
                                    histogram_freq=0, write_graph=True, write_images=False),
        keras.callbacks.ModelCheckpoint(os.path.join(output_dir+'/', "weights.h5"),
                                        verbose=0, save_weights_only=False, monitor='val_loss')
    ]

    model.fit_generator(generator=seq, validation_data=seq2, workers=3, use_multiprocessing=True, epochs=5, callbacks=callbacks)