Pytorch->Caffe

[TonyTangYu]

Platform:ubuntu 16.04
Python version:python 3.5.2
Source framework with version (like Tensorflow 1.4.1 with GPU):
pytorch0.3.1 with GPU
Destination framework with version (like CNTK 2.3 with GPU):
caffe
Pre-trained model path (webpath or webdisk path):
https://github.com/clavichord93/FD-MobileNet/tree/master/saved_models
Running scripts:
In fact, FD-MobileNet is a new kind of deep learning model on the basis of MobileNet. I want to convert this kind of model of pytorch into that of caffe. But I don't know how to use MMdnn. The example esage doesn't give me a clear clue. I wonder how to do this and is there a doc about MMdnn showing all command of MMdnn?

[JiahaoYao]

Hi @TonyTangYu , the tutorial is available at https://github.com/Microsoft/MMdnn#support-frameworks.
However, for the simplity, currently pytorch converter only supports pth file containing the whole model. In the future, it will enable pth just containing the state_dict .

---

Here comes the approach for your problem. You have to make some modification in order to reach such effect.

First, git clone the MMdnn

git clone https://github.com/Microsoft/MMdnn

Commit pytorch_parser.py from Line 67 to Line 76, and add model = model_file_name at Line 77.

It might look like this.

    def __init__(self, model_file_name, input_shape):
        super(PytorchParser, self).__init__()
        # if not os.path.exists(model_file_name):
        #     print("Pytorch model file [{}] is not found.".format(model_file_name))
        #     assert False
        # # test

        # # cpu: https://github.com/pytorch/pytorch/issues/5286
        # try:
        #     model = torch.load(model_file_name)
        # except:
        #     model = torch.load(model_file_name, map_location='cpu')
        model = model_file_name
        self.weight_loaded = True

        # Build network graph
        self.pytorch_graph = PytorchGraph(model)
        self.input_shape = tuple([1] + input_shape)
        self.pytorch_graph.build(self.input_shape)
        self.state_dict = self.pytorch_graph.state_dict
        self.shape_dict = self.pytorch_graph.shape_dict

Then, you install this local mmdnn. You are supposed to move to the path of MMdnn

pip install -e . -U

Secondly, you get the 'FD-mobile' repo.

git clone https://github.com/clavichord93/FD-MobileNet

Also, you commit evaluate.py from Line45 to Line 56, and add this below.

    size = 224
    from mmdnn.conversion.pytorch.pytorch_parser import PytorchParser
    pytorchparser = PytorchParser(model, [3, size, size])
    IR_file = 'FD_mobile'
    pytorchparser.run(IR_file)

The main() function should look like this.

def main():
    global args, best_prec1, last_epoch
    args = parser.parse_args()
    with open(args.data_config, 'r') as json_file:
        data_config = json.load(json_file)
    with open(args.model_config, 'r') as json_file:
        model_config = json.load(json_file)
    if not os.path.exists(args.checkpoint):
        raise RuntimeError('checkpoint `{}` does not exist.'.format(args.checkpoint))

    # create model
    print('==> Creating model `{}`...'.format(model_config['name']))
    model = models.get_model(data_config['name'], model_config)
    checkpoint = torch.load(args.checkpoint, map_location='cpu')
    print('==> Checkpoint name is `{}`.'.format(checkpoint['name']))
    model.load_state_dict(checkpoint['state_dict'])
    size = 224
    from mmdnn.conversion.pytorch.pytorch_parser import PytorchParser
    pytorchparser = PytorchParser(model, [3, size, size])
    IR_file = 'FD_mobile'
    pytorchparser.run(IR_file)

You run the script, and then you will get the IR structure file.

python evaluate.py --data config/imagenet/data-config/imagenet-test.json --model config/imagenet/model-config/fd-mobilenet/1x-FDMobileNet-224.json --checkpoint saved_models/1x-FDMobileNet-224.pth.tar

The result is

IR network structure is saved as [FD_mobile.json].
IR network structure is saved as [FD_mobile.pb].
IR weights are saved as [FD_mobile.npy].

After that, you can use this line to convert IR to Caffe Code.

 mmtocode -f caffe -n FD_mobile.pb -w FD_mobile.npy -d caffe_converted.py -dw caffe_converted.npy

Parse file [FD_mobile.pb] with binary format successfully.
Target network code snippet is saved as [caffe_converted.py].
Target weights are saved as [caffe_converted.npy].

Finally, you can also get the caffe model like this.

 mmtomodel -f caffe -in caffe_converted.py -iw caffe_converted.npy -o caffe_target

You will get this result.

Caffe model files are saved as [caffe_target.prototxt] and [caffe_target.caffemodel], generated by [caffe_converted.py] and [caffe_converted.npy].