PyTorch module to use OpenFace's nn4.small2.v1.t7 model
Clone or download
thnkim Update loadOpenFace.py
Fixed a bug in LRN to support DataParallel
Latest commit 6a5b3ba Nov 12, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
README.md Update README.md May 13, 2017
SpatialCrossMapLRN_temp.py Update SpatialCrossMapLRN_temp.py Nov 12, 2017
loadOpenFace.py Update loadOpenFace.py Nov 12, 2018
openface.pth Add files via upload May 13, 2017
openface_20180119.pth uploaded newly converted model. Jan 18, 2018

README.md

OpenFace for Pytorch

I made a dirty code to use OpenFace in PyTorch. I converted 'nn4.small2.v1.t7' to a .hdf5 file using 'torch-hdf5'. Then I read layer informations from .hdf5 file, which can be displayed as follows:

nn.Sequential {
  [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> (13) -> (14) -> (15) -> (16) -> (17) -> (18) -> (19) -> (20) -> (21) -> (22) -> (23) -> (24) -> (25) -> (26) -> output]
  (1): nn.SpatialConvolutionMM(3 -> 64, 7x7, 2,2, 3,3)
  (2): nn.SpatialBatchNormalization (4D) (64)
  (3): nn.ReLU
  (4): nn.SpatialMaxPooling(3x3, 2,2, 1,1)
  (5): nn.SpatialCrossMapLRN
...

Then I manually coded layers in PyTorch (see loadOpenFace.py) with some tentative layers code which may be supported by PyTorch officially laters (SpatialCrossMapLRN_temp.py, adopted from PyTorch's nn.legacy). The final model is 'openface.pth' (which may need to be renamed to 'openface_nn4_small2_v1.pth'), which can be loaded by codes in loadOpenFace.py.

Please see main section of loadOpenFace.py for how-to-use. Simply,

net = prepareOpenFace(useCuda=True, gpuDevice=0, useMultiGPU=False).eval()
feature = net(input_tensor)    # input_tensor should be (batch_size, 3, 96, 96)