Hi!

Thank you for creating this repo, this was exactly what I was looking for!

I am trying to convert a caffe model to onnx, however, it seems that when I test the model with the same input, the output of the onnx model is different to the original caffe model.

I have attached a link to the .caffemodel and .prototxt that I am working with (as well as the produced onnx model) - github would not allow me to upload a zip file containing the models :( --> https://1drv.ms/u/s!AsfDc4tZ90mEmHlnImwWfaC1asOM
When passing an input image (the input_name is "data_input") of size (1,3,500,500), the caffe model produces the correct output with size (1,21,500, 500) (this is the shape of net.blobs['score'].data.shape), whereas the onnx model produces an output with size (1,3, 500,500). I am using the latest version of caffe2onnx.

Any assistance would be highly appreciated.

Thanks!

Hi @asiryan
I am trying to convert a caffe_xilinx yolov3 model to onnx.I got this error during conversion

google.protobuf.text_format.ParseError: 2027:3 : Message type "caffe.LayerParameter" has no field named "deephi_resize_param"

I think this deephi_resize_param would be a custom layer built in caffe-xilinx.Can u add support for custom layers ?

thanks a lot!!!

Hi,
Addition of SSD models support would be useful. Do you have any plans to add support for the same?
Example of caffe SSD model can be found here
https://github.com/chuanqi305/MobileNet-SSD

Also Argmax layer support is missing, this would be useful to support segmentation models