BVLC / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
34.12k stars 18.69k forks source link

Write padding aware im2col? #36

Closed mavenlin closed 10 years ago

mavenlin commented 10 years ago

@Yangqing Is there any specific reason for an individual padding layer but not supporting padding in im2col? My concern is that the padding layer causes unnecessary extra memory usage which can be avoided. Is it necessary to implement padding aware im2col function?

Yangqing commented 10 years ago

Just legacy reason, I guess... I was translating codes quickly from decaf to caffe and did not worry too much about memory consumption back then. Yeah, it would be good to have a padding aware im2col function to save a (rather unnecessary) copy of the data.

sguada commented 10 years ago

It would be even better to have a more memory efficient convolutional layer that won't require im2col, but that it is going to be harder.

Sergio

2014-02-03 Yangqing Jia notifications@github.com:

Just legacy reason, I guess... I was translating codes quickly from decaf to caffe and did not worry too much about memory consumption back then. Yeah, it would be good to have a padding aware im2col function to save a (rather unnecessary) copy of the data.

Reply to this email directly or view it on GitHubhttps://github.com/BVLC/caffe/issues/36#issuecomment-34001933 .

Yangqing commented 10 years ago

I assume Forrest is working along that direction (not sure).

Yangqing

On Mon, Feb 3, 2014 at 1:40 PM, Sergio Guadarrama notifications@github.comwrote:

It would be even better to have a more memory efficient convolutional layer that won't require im2col, but that it is going to be harder.

Sergio

2014-02-03 Yangqing Jia notifications@github.com:

Just legacy reason, I guess... I was translating codes quickly from decaf to caffe and did not worry too much about memory consumption back then. Yeah, it would be good to have a padding aware im2col function to save a (rather unnecessary) copy of the data.

Reply to this email directly or view it on GitHub< https://github.com/BVLC/caffe/issues/36#issuecomment-34001933> .

Reply to this email directly or view it on GitHubhttps://github.com/BVLC/caffe/issues/36#issuecomment-34003395 .

sguada commented 10 years ago

I was talking with Forrest, and he plans to work on that in a week or two.

I found this examples about convolutions http://www.evl.uic.edu/sjames/cs525/final.html

zimenglan-sysu commented 10 years ago

@Yangqing how to convert the 2 dimension convolution to 3 dimension convolution in the im2col.cpp and conv_layer.cpp files? Thinks