Re-sizing extracted features for use in caffe

[saumya-jetley]

So as to load only a portion of the caffe net for GPU based training, I am extracting features at layer L1 and re-using them in a model built L1 onwards.
To load the extracted (sequential) features, I must re-structure them to the blob dimension at L1. Is there a better way to do this than to manually set the blob dimensions while extracting, like so:

//snippet from the 'root-to-caffe/tools/extract_features' code
int h = h1; //Addition 1
int w = w1; //Addition 2
int channels = c1; //Addition 3
for (int n = 0; n < batch_size; ++n) {
datum.set_height(h);
datum.set_width(w);
datum.set_channels(channels);
datum.clear_data();

I have added the top three lines into the extract_features.cpp code to extract a blob of dimension h1xw1xc1; where h1-height, w1-width,c1-number of channels.

Please feel free to correct me or make suitable suggestions.
thanks.

[jyegerlehner]

That's the same thing I ran into. So I suggest: yes, just make those changes locally in the code.

[shelhamer]

Fixed in #1457 -- merged in c942dc1.

[saumya-jetley]

oh, somehow i wasn't able to track down that issue. Thanks a lot guys! that helps.