How to convert Mxnet to Caffe? about mmdnn

[TobeyQin]

You should first convert mxnet pre-trained model to IR(see here), and then convert IR to caffe.

I will take resnet152 model as an example:

1. Convert mxnet to IR:
   python -m mmdnn.conversion._script.convertToIR -f mxnet -n mxnet/models/resnet-152-symbol.json -w mxnet/models/resnet-152-0000.params -d resnet152 --inputShape 3 224 224

2. Convert IR to caffe code:
   python -m mmdnn.conversion._script.IRToCode -f caffe -n resnet152.pb -w resnet152.npy -d caffe_resnet152.py -dw caffe_resnet152.npy

3. Generate caffe model from code:
   python caffe_resnet152.py -w caffe_resnet152.npy -p resnet152.prototxt -m resnet152.caffemodel

Hope this can help solve your question. Currently, from IR to caffe is under initial state, we haven't done enough test, if there is any issue when you use it, welcome to report.

[tzhang2014]

thx , I have run the above example, when I run the first command , Loadiing I met three warnings: 1.MXNet Parser has not supported operator num with name data. 2. convert the null operator with name [data] into tnput layer. 3. SoftmaxOutput attrs are not supported in IR. So, When I run the second command ,I met a error that printf("=======current_node=======\n{}".format(current_node.layer),file=f)
SyntaxError: invalid syntax .

[kitstar]

Hi @tzhang2014 ,
Fixed. Please check the new module, Thanks.

[gunasekaran7]

I am trying to convert from IR to Caffe. And I am getting the following error. Any help would be appreciated.
I am using python3.4 and IR is created from a Keras model.

[kitstar]

Hi @gunasekaran7 , looks like there are some problem on Caffe Environment. Could you check it?