Is fit_generator in Keras supposed to reset the generator after each epoch?

Question

I am trying to use fit_generator with a custom generator to read in data that's too big for memory. There are 1.25 million rows I want to train on, so I have the generator yield 50,000 rows at a time. fit_generator has 25 steps_per_epoch, which I thought would bring in those 1.25MM per epoch. I added a print statement so that I could see how much offset the process was doing, and I found that it exceeded the max when it got a few steps into epoch 2. There are a total of 1.75 million records in that file, and once it passes 10 steps, it gets an index error in the create_feature_matrix call (because it brings in no rows).

def get_next_data_batch():
    import gc
    nrows = 50000
    skiprows = 0

    while True:
        d = pd.read_csv(file_loc,skiprows=range(1,skiprows),nrows=nrows,index_col=0)
        print(skiprows)
        x,y = create_feature_matrix(d)
        yield x,y
        skiprows = skiprows + nrows
        gc.collect()
get_data = get_next_data_batch()

... set up a Keras NN ...

model.fit_generator(get_next_data_batch(), epochs=100,steps_per_epoch=25,verbose=1,workers=4,callbacks=callbacks_list)

Am I using fit_generator wrong or is there some change that needs to be made to my custom generator to get this to work?

Answer 1

No - fit_generator doesn't reset generator, it's simply continuing calling it. In order to achieve the behavior you want you may try the following:

def get_next_data_batch(nb_of_calls_before_reset=25):
    import gc
    nrows = 50000
    skiprows = 0
    nb_calls = 0

    while True:
        d = pd.read_csv(file_loc,skiprows=range(1,skiprows),nrows=nrows,index_col=0)
        print(skiprows)
        x,y = create_feature_matrix(d)
        yield x,y
        nb_calls += 1
        if nb_calls == nb_of_calls_before_reset:
            skiprows = 0
        else:
            skiprows = skiprows + nrows
        gc.collect()