Installation

Official note on installation https://caffe.berkeleyvision.org/installation.html

Installing CPU version on CHPC

  • Install Miniconda

  • Install caffe using condaconda install -c intel caffe

lsb_release -d Description: CentOS release 6.10 (Final)

Building GPU version on CHPC

(Not succeeded yet…. aborted)

  1. Check gcc version to be 4.8.5 module load gcc-4.8.5
  2. Check cmake version to newer than 3 module load cmake-3.14.5

Reference

http://www.andrewjanowczyk.com/installing-caffe-on-the-ohio-super-computing-osc-ruby-cluster/

http://homepage.cs.uiowa.edu/~zli79/notes/Installation_Caffe_on_Argon_Cluster.pdf

Installing GPU version on CHPC

https://anaconda.org/anaconda/caffe-gpu

Just use miniconda, and conda install -c anaconda caffe-gpu

Installing GPU version on Windows machine

The easiest way is using the the conda export file to copy the same environment. See the attached caffe-torch tf-torch.yml for more information.

conda create --name caffe36 --file caffe36_spec-file.txt
conda create -f caffe-torch.yml

Then it will download and install all the packages with the required source and version

conda activate caffe36

Check onenote for more debugging note.

If some dlls are missing MSVCP140D.DLL, VCRUNTIME140D.DLL, download them from net and put it in C:\Windows\System32.

Testing Installation

If on CHPC, start by submitting a gpu job qsub -I -l nodes=1:ppn=1:gpus=1,walltime=01:00:00

After downloading some weights for models by using

python download_model_binary.py ../models/bvlc_reference_caffenet/

We have enough material to test loading a network into memory.

Test CPU installation,

import caffe
import os.path
caffe.set_mode_cpu()
net_weights = os.path.join("/home/binxu/caffe/models/bvlc_reference_caffenet/", 'bvlc_reference_caffenet.caffemodel')
net_definition = os.path.join("/home/binxu/caffe/models/bvlc_reference_caffenet/", 'deploy.prototxt')
net = caffe.Net(net_definition, caffe.TEST, weights=net_weights)

net.blobs['data'].data[...] = 1
net.forward(end='fc6')

Test GPU installation, just use caffe.set_mode_gpu()

Berkeley Model zoo

Converting Caffe Model to PyTorch

As some operations are much easier in some newer frameworks like PyTorch, there is a need to convert existing model in Caffe to PyTorch.

There is some existing packages target at interpret proto of caffe and build network according to that in PyTorch, e.g. pytorch-caffe. However, to support some layer specific to Caffe, we need to add some modification to the package! See LRN

Data loading and transformation

From the documentation of caffe.io.Transformer we can see that

Format input for Caffe:

  • convert to single
  • resize to input dimensions (preserving number of channels)
  • transpose dimensions to K x H x W
  • reorder channels (for instance color to BGR)
  • scale raw input (e.g. from [0, 1] to [0, 255] for ImageNet models)
  • subtract mean
  • scale feature

Parameters in_ : name of input blob to preprocess for data : (H’ x W’ x K) ndarray

Returns caffe_in : (K x H x W) ndarray for input to a Net

The input convention for Caffe is 4 dimension array in the order of [B, C, H, W]. And for color image, the color channel is in the order of BGR.

So the output from a network built in Caffe have to be transformed by a function to visualize.

Pipelining Caffe networks in PyTorch

However, if that output is directly sent into the other network, then the 2 networks can pipe into each other except for one thing.

Standard preprocess

  • resize
  • transpose the axis to C,H,W
  • swap the channel to BGR
  • scale by * 255
  • centralize by -meanBGR_value

Standard deprocess

  • de-centralize by + meanBGR_value
  • scale by / 255
  • swap the channel to RGB
  • transpose the axis to H,W,C
  • clip the tensor to be within (0,1)

So the 2 process are inverse to each other except for clip part.

Local Response Normalization

This is an ancient module exists in AlexNet and CaffeNet, but largely abandoned in more recent network architectures. At first, it’s inspired by lateral inhibition just like those happened in biological visual networks. However, it’s not very stable, and help only in certain cases but not others.

Thus, it’s not officially supported in PyTorch for a long time, but currently it is supported.

Caffe LRN Doc

PyTorch LRN Doc