Since the pytorch models we use are basically dynamic graph structures, the problem with dynamic graphs is that the graph structure cannot be determined once the forward is incomplete, and caffe is a static graph framework, which will cause the model conversion from pytorch to caffe to encounter many problems, and the pytorch version iteration is very fast,so this repo will no longer be recommended.If you want to convert pytorch to caffe,we suggest using pytorch->onnx->caffe by this repo https://github.com/xxradon/ONNXToCaffe.
Providing a tool for neural network frameworks for pytorch and caffe.
The nn_tools is released under the MIT License (refer to the LICENSE file for details).
Noting: pytorch version 1.1 is not supported now
The analyser can analyse all the model layers' [input_size, output_size, multiplication ops, addition ops, comparation ops, tot ops, weight size and so on] given a input tensor size, which is convenint for model deploy analyse.
Before you analyse your network, Netscope is recommended to visiualize your network.
Command：python caffe_analyser.py [-h] prototxt outdir shape
,), in caffe image shape should be:
batch_size, channel, image_height, image_width.For example python caffe_analyser.py resnet_18_deploy.prototxt analys_result.csv 1,3,224,224
Supporting analyse the inheritors of torch.nn.Moudule class.
Command：pytorch_analyser.py [-h] [--out OUT] [--class_args ARGS] path name shape
,), in pytorch image shape should be:
batch_size, channel, image_height, image_width.For example python pytorch_analyser.py example/resnet_pytorch_analysis_example.py resnet18 1,3,224,224
The new version of pytorch_to_caffe supporting the newest version(from 0.2.0 to 1.2.0) of pytorch. NOTICE: The transfer output will be somewhat different with the original model, caused by implementation difference.
Supporting layers types:
conv2d -> Convolution,
_conv_transpose2d -> Deconvolution,
_linear -> InnerProduct,
_split -> Slice,
max_pool2d,_avg_pool2d -> Pooling,
_max -> Eltwise,
_cat -> Concat,
dropout -> Dropout,
relu -> ReLU,
prelu -> PReLU,
_leaky_relu -> ReLU,
_tanh -> TanH,
threshold(only value=0) -> Threshold,ReLU,
softmax -> Softmax,
batch_norm -> BatchNorm,Scale,
instance_norm -> BatchNorm,Scale,
_interpolate -> Upsample
_hardtanh -> ReLU6
_permute -> Permute
_l2Norm -> Normalize
Supporting operations: torch.split, torch.max, torch.cat ,torch.sigmoid, torch.div
Supporting tensor Variable operations: var.view, + (add), += (iadd), -(sub), -=(isub) * (mul) = (imul), torch.Tensor.contiguous(_contiguous), torch.Tensor.pow(_pow), * torch.Tensor.sum(_sum), torch.Tensor.sqrt(_sqrt), torch.Tensor.unsqueeze(_unsqueeze) \ torch.Tensor.expand_as(_expand_as),
Need to be added for caffe in the future:
The supported above can transfer many kinds of nets, such as AlexNet(tested), VGG(tested), ResNet(fixed the bug in origin repo which mainly caused by ReLu layer function.), Inception_V3(tested).
The supported layers concluded the most popular layers and operations. The other layer types will be added soon, you can ask me to add them in issues.
Example: please see file example/alexnet_pytorch_to_caffe.py. Just Run python3 example/alexnet_pytorch_to_caffe.py.
Attention: the main difference from convert model is the BN layer,you should pay more attention to the BN parameters like momentum=0.1, eps=1e-5.
After Converter,we should use verify_deploy.py to verify the output of pytorch model and the convertted caffe model. If you want to verify the outputs of caffe and pytorch,you should make caffe and pytorch install in the same environment,anaconda is recommended. using following script,we can install caffe-gpu(master branch).
conda install caffe-gpu pytorch cudatoolkit=9.0 -c pytorch
other way,we can use docker,and in https://github.com/ufoym/deepo,for cuda9
docker pull ufoym/deepo:all-py36-cu90for cuda10
docker pull ufoym/deepo:all-py36-cu100please see file example/verify_deploy.py,it can verify the output of pytorch model and the convertted caffe model in the same input.