CODEWITHSUNDEEP
neural-networkdeep-learningcaffe
Franck Dernoncourt
Franck Dernoncourt

Reputation: 83157

How can I modify the batch size on the fly, i.e. after loading an ANN in pycaffe?

I trained a neural network using Caffe. I then use it to predict outputs given some new inputs. To do so, I load the trained neural network in pycaffe along with a deploy.prototxt, which specifies the inputs:

name: "IrisNet"
input: "data"
input_dim: 1 # batch size
input_dim: 1
input_dim: 1
input_dim: 300 # number of features

input: "adbeoption"
input_dim: 1 # batch size
input_dim: 1
input_dim: 1
input_dim: 1 # number of features


layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "data"
  top: "ip1"

[...]

I load the neural network using:

my_net = caffe.Net(deploy_prototxt_filename,caffemodel_filename, caffe.TEST)

Since I don't know in advance how many inputs I will have, I would like to be able to change the batch size after I loaded the neural network (i.e. after calling caffe.Net()). How to do so?

Upvotes: 1

Views: 672

Answers (1)

Franck Dernoncourt
Franck Dernoncourt

Reputation: 83157

You can use reshape:

net.blobs['data'].reshape(data_batch_size, 1, 1, data_of_features)
net.blobs['adbeoption'].reshape(adbeoption_batch_size, 1, 1, 1)

Then you can call net.forward().

It will modify the batch size on the fly, without having to reload the ANN.

Upvotes: 1

Related Questions