Is it possible to have dynamic batchsize in keras?

Question

Keras codes I have looked or wrote have fixed batchsize during training (i.e. 32, 64, 128 ...). I am wondering if it is possible to have dynamic batchsize. (For example, 104 in the first iteration, 82 in the next iteration, 95 in next, and so on.)

I am currently using tensorflow backend.

DJK · Accepted Answer · 2017-11-06T12:27:03.423

4

It is possible if you train on a loop vs training with fit. an example

from random import shuffle    

dataSlices = [(0,104),(104,186),(186,218)]

for epochs in range(0,10):
    shuffle(dataSlices)
    for i in dataSlices:
        x,y = X[i[0]:i[1],:],Y[i[0]:i[1],:]
        model.fit(x,y,epochs=1,batchsize=x.shape[0])
        #OR as suggest by Daniel Moller
        #model.train_on_batch(x,y)

This would assume your data is 2d numpy arrays. This idea can be further expanded to use a fit_generator() inplace of the for loop if you so choose (see docs).

edited Nov 06 '17 at 12:27

answered Nov 06 '17 at 00:56

DJK

8,924
4
24
40

2

You can probably use `train_on_batch` for that instead of `fit`. (This may eliminate some overhead). – Daniel Möller Nov 06 '17 at 11:39
Thanks, djk47463 and Daniel Möller! It helps. – user3377018 Nov 07 '17 at 02:58

score 0 · Answer 2 · answered Apr 24 '19 at 21:16

0

Use None to set any batch size. In particular this answer explains it.

answered Apr 24 '19 at 21:16

prosti

42,291
14
186
151

Is it possible to have dynamic batchsize in keras?

2 Answers2