Caffe: How to change the batch size dynamically?

Created on 3 Jul 2015 · 7Comments · Source: BVLC/caffe

In pycaffe particular

Is there any ways we can change the batch_size dynamically without modifying the prototxt file ?

I noticed the method reshape, but apparently it does not really change the shape, as after I run forward(), the shape will have be altered back to its original shape

Many thanks

Source

chwangaa

Most helpful comment

This example may be helpful to you. See the code in forward_pass():

for chunk in [caffe_images[x:x+batch_size] for x in xrange(0, len(caffe_images), batch_size)]:
    new_shape = (len(chunk),) + tuple(dims)
    if net.blobs['data'].data.shape != new_shape:
        net.blobs['data'].reshape(*new_shape)
    for index, image in enumerate(chunk):
        image_data = transformer.preprocess('data', image)
        net.blobs['data'].data[index] = image_data
    output = net.forward()[net.outputs[-1]]

lukeyeager on 6 Jul 2015

👍5

All 7 comments

Basically I am trying to see the relationship between the batch size and the running time

chwangaa on 3 Jul 2015

You need to reshape the input blob and then call reshape on the net, before calling forward.

seanbell on 5 Jul 2015

I did the following
net.blobs['data'].reshape(BLOCK_SIZE, DIMENSION, HEIGHT, WIDTH) net.blobs['label'].reshape(BLOCK_SIZE, ) net.reshape()

The shape of every layer changes correctly. However, as soon as I run forward, the shape will automatically change back :-(

chwangaa on 5 Jul 2015

This example may be helpful to you. See the code in forward_pass():

for chunk in [caffe_images[x:x+batch_size] for x in xrange(0, len(caffe_images), batch_size)]:
    new_shape = (len(chunk),) + tuple(dims)
    if net.blobs['data'].data.shape != new_shape:
        net.blobs['data'].reshape(*new_shape)
    for index, image in enumerate(chunk):
        image_data = transformer.preprocess('data', image)
        net.blobs['data'].data[index] = image_data
    output = net.forward()[net.outputs[-1]]

lukeyeager on 6 Jul 2015

👍5

Somehow I cannot manage to have it work. I am not trying to classify here. So I used the method suggested above to change the batchSize to 1000. However, the running time of forward does not even slightly alter, (i.e. when I modify the value in prototxt file the running time will be longer). Therefore I think the above method does not work

chwangaa on 7 Jul 2015

Please discuss usage on caffe-users; I've answered in the thread Regarding Change BatchSize in python.