Added ChannelConstantFiller

[sguada]

Added the option to fill a blob using a different constant value per channel.

This change allows to fill the mean_image of a data_layer with the mean of the RGB values.

[shelhamer]

Is this better than passing a mean image file with the same value at every pixel?

[sguada]

I think this way is better because one can define it in the prototxt,
change when need it and don't need to create one mean image file for every
different size or value.

Sergio

2014-02-11 Evan Shelhamer notifications@github.com:

Is this better than passing a mean image file with the same value at every
pixel?

Reply to this email directly or view it on GitHubhttps://github.com//pull/96#issuecomment-34835135
.

[shelhamer]

Ok, good motivation. Could you include an example network using this instead of the meanfile field? Can this be included in deployment nets like examples/imagenet_deploy.prototxt to automatically subtract the mean from inputs and avoid the need to do it in user code (like the wrappers)?

[sguada]

I have wrote some prototxt as examples on how to use the bias_filler to fill the data_mean instead of using meanfile.
I have performed some test to see the effect in validation accuracy and found the following interesting results:

==> logs/imagenet_test_zero_mean.log <==
E0212 16:52:36.766688  6231 test_net.cpp:54] Test accuracy:0.49616

==> logs/imagenet_test_constant_mean.log <==
E0212 16:48:28.909536  5106 test_net.cpp:54] Test accuracy:0.56152

==> logs/imagenet_test_channelconstant_mean.log <==
E0212 17:10:05.759047 11184 test_net.cpp:54] Test accuracy:0.57048

==> logs/imagenet_test_mean_file.log <==
E0212 17:01:32.425643  8061 test_net.cpp:54] Test accuracy:0.57278

What means that if one don't remove the mean from the images the performance drops substantially 7.66% . If one remove the average pixel value from the image the performance drops 1.13%. While if one remove the average RGB values of the training-data from the image the performance only drops 0.23%

So for anyone using a wrapper they can just remove the mean RGB values and lose very little in performance.

[shelhamer]

Could this be better as a DataProcessing layer #148? With a MeanSubtraction layer one could define a mean image, single number, or channel as desired and use it for training, test, and deployment. A workaround is to add an InnerProduct layer immediately after the input with the bias filled to the negative mean, but that doesn't feel clean.

[shelhamer]

@sguada let's wait for #148 since this is more appropriately implemented in a MeanSubtractionLayer data preprocessing layer. Thanks for the constant / channel constant / mean image analysis. We should add that to the docs somewhere.