Problem with Fillers in matcaffe in OSX

[sguada]

This is a weird problem.
When using imagenet_deploy.txt containing weight_filler and/or bias_filler in a layer then the net.init() process don't continue.

When not using weight_filler and/or bias_filler in the layers then net.init() works well but then when restoring the trained model there is dimension mismatch in fc6 layer.

[shelhamer]

Sergio, you already raised this issue in #175. Please close one of these.

[shelhamer]

Also, I'm afraid you're not even on dev since imagenet_deploy.prototxt doesn't exist anymore, having been replaced by models/imagenet.prototxt. Or is this bug on master?

[sguada]

@shelhamer sorry for the confusion with imagenet_deploy.prototxt and imagenet.prototxt. I know there is no imagenet_deploy.prototxt anymore, but imagenet.prototxt doesn't currently work (since it include the fillers) with matcaffe, so I created a temporal imagenet_deploy.prototxt which doesn't include fillers, but then I get a problem when restoring the net due to the dimensions.

[sguada]

@shelhamer I'm tracking down the error, it seems that it is related to OSX, since in Ubuntu it doesn't happen.

If one removes the fillers from imagenet.prototxt then matcaffe_demo works in OSX, otherwise it doesn't. Although it takes a lot of time to set_mode to CPU (I cannot test GPU due to memory constraints), while it should be fast since class Caffe is a singleton (isn't it? @Yangqing). These problems don't happen in Ubuntu.

Once the mode is set it is fast to process the images.

[sguada]

@Yangqing the part of the code that make it slow is the initialization of CUBLAS in common.cpp

  if (cublasCreate(&cublas_handle_) != CUBLAS_STATUS_SUCCESS) {
    LOG(ERROR) << "Cannot create Cublas handle. Cublas won't be available.";
  }

Does anyone knows why this happen? This makes matlab to use 300MB extra of memory. When in reality it is never used in CPU mode.
This is something to keep in mind for CPU-ONLY version #3

[shelhamer]

@sguada is this still a problem or ancient history?

[shelhamer]

Closing; pretty sure this is solved by matcaffe improvements and the resolution of the CUDA set device call order issue in #507 #521.