Transform neural network error with loadcaffe? - torch

Question

Wanna try the project https://github.com/chuanli11/CNNMRF. After cloning and installing loadcaffe by luarocks. Then run the comand:
qlua cnnmrf.lua -content_name master_entrance_portrait
Got this error:
mkdir: cannot create directory ‘data/result/’: File exists
mkdir: cannot create directory ‘data/result/trans/’: File exists
mkdir: cannot create directory ‘data/result/trans/MRF/’: File exists
mkdir: cannot create directory ‘data/result/trans/MRF/’: File exists
[libprotobuf WARNING google/protobuf/io/coded_stream.cc:505] Reading dangerously large protocol message. If the message turns out to be larger than 1073741824 bytes, parsing will be halted for security reasons. To increase the limit (or to disable these warnings), see CodedInputStream::SetTotalBytesLimit() in google/protobuf/io/coded_stream.h.
[libprotobuf WARNING google/protobuf/io/coded_stream.cc:78] The total number of bytes read was 574671192
Successfully loaded data/models/VGG_ILSVRC_19_layers.caffemodel
qlua: ...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:19: nOutputPlane should be divisible by nGroups
stack traceback:
[C]: at 0x7fec4cde29d0
[C]: in function 'assert'
...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:19: in function '__init'
...yne/Downloads/torch/install/share/lua/5.1/torch/init.lua:91: in function <...yne/Downloads/torch/install/share/lua/5.1/torch/init.lua:87>
[C]: in function 'SpatialConvolution'
data/models/VGG_ILSVRC_19_layers_deploy.prototxt.lua:3: in main chunk
[C]: in function 'dofile'
...oads/torch/install/share/lua/5.1/loadcaffe/loadcaffe.lua:24: in function 'load'
./transfer_CNNMRF_wrapper.lua:372: in function 'main'
cnnmrf.lua:92: in main chunk
Guess caused by some underlying error in converting caffe model to torch model through loadcaffe. The net to be converted was VGG19 from the files VGG_ILSVRC_19_layers.cafflemodel and VGG_ILSVRC_19_layers_deploy.prototxt. I have checked these files by checksum, they were OK. But the converted one, that is, VGG_ILSVRC_19_layers_deploy.prototxt.lua was something like:
require 'cudnn'
local model = {}
table.insert(model, {'conv1_1', cudnn.SpatialConvolution(0, -1118555060, 4, 1, 0, 0, -47878672, 4, 64)})
table.insert(model, {'relu1_1', cudnn.ReLU(true)})
table.insert(model, {'conv1_2', cudnn.SpatialConvolution(0, 1041742427, 4, 1, 0, 0, -76411872, 4, 64)})
table.insert(model, {'relu1_2', cudnn.ReLU(true)})
table.insert(model, {'pool1', cudnn.SpatialMaxPooling(2, 2, 2, 2, 0, 0):ceil()})
table.insert(model, {'conv2_1', cudnn.SpatialConvolution(0, -1172344618, 4, 1, 0, 0, -76404000, 4, 128)})
table.insert(model, {'relu2_1', cudnn.ReLU(true)})
table.insert(model, {'conv2_2', cudnn.SpatialConvolution(0, -1134797354, 4, 1, 0, 0, -76401184, 4, 128)})
table.insert(model, {'relu2_2', cudnn.ReLU(true)})
table.insert(model, {'pool2', cudnn.SpatialMaxPooling(2, 2, 2, 2, 0, 0):ceil()})
table.insert(model, {'conv3_1', cudnn.SpatialConvolution(0, 32609, 4, 1, 0, 0, -76397968, 4, 256)})
table.insert(model, {'relu3_1', cudnn.ReLU(true)})
table.insert(model, {'conv3_2', cudnn.SpatialConvolution(0, 976600020, 4, 1, 0, 0, -47647424, 4, 256)})
table.insert(model, {'relu3_2', cudnn.ReLU(true)})
table.insert(model, {'conv3_3', cudnn.SpatialConvolution(0, -1193143920, 4, 1, 0, 0, -47644208, 4, 256)})
table.insert(model, {'relu3_3', cudnn.ReLU(true)})
table.insert(model, {'conv3_4', cudnn.SpatialConvolution(0, 965660278, 4, 1, 0, 0, -47640880, 4, 256)})
table.insert(model, {'relu3_4', cudnn.ReLU(true)})
table.insert(model, {'pool3', cudnn.SpatialMaxPooling(2, 2, 2, 2, 0, 0):ceil()})
table.insert(model, {'conv4_1', cudnn.SpatialConvolution(0, 994497290, 4, 1, 0, 0, -47635600, 4, 512)})
table.insert(model, {'relu4_1', cudnn.ReLU(true)})
table.insert(model, {'conv4_2', cudnn.SpatialConvolution(0, 994239873, 4, 1, 0, 0, -47631248, 4, 512)})
table.insert(model, {'relu4_2', cudnn.ReLU(true)})
table.insert(model, {'conv4_3', cudnn.SpatialConvolution(0, -1162151849, 4, 1, 0, 0, -47626896, 4, 512)})
table.insert(model, {'relu4_3', cudnn.ReLU(true)})
table.insert(model, {'conv4_4', cudnn.SpatialConvolution(0, -1161155618, 4, 1, 0, 0, -47622544, 4, 512)})
table.insert(model, {'relu4_4', cudnn.ReLU(true)})
table.insert(model, {'pool4', cudnn.SpatialMaxPooling(2, 2, 2, 2, 0, 0):ceil()})
table.insert(model, {'conv5_1', cudnn.SpatialConvolution(0, 990007184, 4, 1, 0, 0, -47617264, 4, 512)})
table.insert(model, {'relu5_1', cudnn.ReLU(true)})
table.insert(model, {'conv5_2', cudnn.SpatialConvolution(0, -1165327502, 4, 1, 0, 0, -47612912, 4, 512)})
table.insert(model, {'relu5_2', cudnn.ReLU(true)})
table.insert(model, {'conv5_3', cudnn.SpatialConvolution(0, -1173363995, 4, 1, 0, 0, -338687408, 4, 512)})
table.insert(model, {'relu5_3', cudnn.ReLU(true)})
table.insert(model, {'conv5_4', cudnn.SpatialConvolution(0, 980886043, 4, 1, 0, 0, -47608144, 4, 512)})
table.insert(model, {'relu5_4', cudnn.ReLU(true)})
table.insert(model, {'pool5', cudnn.SpatialMaxPooling(2, 2, 2, 2, 0, 0):ceil()})
table.insert(model, {'torch_view', nn.View(-1):setNumInputDims(3)})
table.insert(model, {'fc6', nn.Linear(0, 0)})
table.insert(model, {'relu6', cudnn.ReLU(true)})
table.insert(model, {'drop6', nn.Dropout(0.500000)})
table.insert(model, {'fc7', nn.Linear(0, 0)})
table.insert(model, {'relu7', cudnn.ReLU(true)})
table.insert(model, {'drop7', nn.Dropout(0.500000)})
table.insert(model, {'fc8', nn.Linear(0, 0)})
table.insert(model, {'prob', cudnn.SoftMax()})
return model
Found it was in loadcaffe.lua:
C.convertProtoToLua(handle, lua_name, backend)
then in ffi.lua:
ffi.cdef[[
...
void convertProtoToLua(void** handle, const char* lua_name, const char* cuda_package);
...
]]
Does this mean the loadcaffe installed in luarocks was damaged before install? What could be the cause? Thanks in advance!