Pytorch2Caffe

This tool is for converting PyTorch CNN models to Caffe network configuration files (prototxt) and parameter files (caffemodel).

Install

To install pytorch2caffe, go to its folder, then:

$ pip install .

If you would like to install in develop mode:

$ pip install -e .

API

from pytorch2caffe.converter import pytorch2caffe_converter

def your_function(model):
    # set up input layer information
    input = Variable(torch.ones([1, 3, 416, 416]))
    source = '/path/to/image_list.txt'
    root_foler = 'path/to/image_folder'
    batch_size = 1
    new_height = 416
    new_width = 416

    # initialize a pytorch2caffe converter object
    torch2caffe = pytorch2caffe_converter(model)

    # set input
    torch2caffe.set_input(input, source, root_folder, batch_size, new_height, new_width)

    # translate 
    torch2caffe.trans_net('resnet50')

    # save results
    torch2caffe.save_prototxt('./resnet50.prototxt')
    torch2caffe.save_caffemodel('./resnet50.caffemodel')
    # optional
    torch2caffe.save_torch2caffe_names_json('./torch2caffe_names.json')

Done

• Added None blob checking

• Fixed input layer issues: Vitis requires input blob is named data, and it must be an input layer, not input field.

• Add input data source and loss layer

• Tensor operations are supported

• Remove Reduction layer, and check x.mean(3).mean(2) which is supposed to be a global average pooling layer in caffe

TODO

• Supported layer checking

• Global average pooling support: torch.nn.AdaptiveAvgPooling2d(1,1)

Note

1. Issues with name correspondence

We have a dict() in class Translog to record the torch function names and their corresponding caffe layer names. But we don't have that information for tensor operations.

For example, += can convert into Eltwise in Caffe, but it does not have torch name since it is not a function.

2. DPU IP does not support Global Average Pooling

Global average pooling is moved to CPU, resulting in multiple kernels.

It's actually supported. A working example:

layer {
  name: "global_avg_pool"
  type: "Pooling"
  bottom: "batch_norm_blob49"
  top: "mean_blob2"
  pooling_param {
    pool: AVE
    global_pooling: true
  }
}

3. The input output channel restriction for depthwise convolution

The Xilinx DPU Compiler actually has restrictions for depthwise conv's i/o channel number. If the number exceeds the maximum value, the compiler would trigger error:

[VAI_C-BACKEND][Check Failed: kernel_param * input_channel_group <= weights_buf_depth][/usr/local/anaconda3/conda-bld/vaic_1584632209107/work/dnnc/submodules/asicv2com/src/Operator/OperatorDptConv.cpp:31][DATA_OUTRANGE][Data value is out of range!] 

We don't know the actual value of weights_buf_depth because the C code is not open-sourced. So I found out the maximum value with binary search:

• for 7x7 depth conv, the maixmum value is 656,
• for 5x5 depth conv, the maximum value is 1296,
• for 3x3 depth conv, the maximum value is 2000+, we don't have to worry about it.