Here is the problem statement:
Openvino version :: 2020.3.194
Model :: mobilenetv1 trained in BVLC caffe.
Input shape :: [1, 3, 224, 224]
Scale used for normalization :: 58.8235.
Mean values per channel are :: [63.5797, 63.5797, 63.5797]
We ran this caffe model using BVLC caffe. Processing of the input image (I) is
[(I - mean values) / scale] goes for the forward pass. Result RCaffe (total number of correct cases).
Now this model mobilenetv1 is converted to openvino IR using the command mentioned below:
"python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_caffe.py --input_model mobilenetv1.caffemodel --input_proto mobilenetv1.prototxt --scale_values [58.8235,58.8235,58.8235] --output_dir openvino_models_mobilenetv1/ --mean_values [63.5797,63.5797,63.5797] --data_type FP32 --input_shape [1,3,224,224] --input data --output prob"
We ran this IR .xml and .bin files using openvino inference engine. Result ROpenvino (total number of correct cases).
Now we are getting different RCaffe and ROpenvino.
We have verified it using C++ and python code scripts. In both cases we are getting discrepancies in results from Caffe and Openvino IRs.
So, please let me know if I am doing something wrong or it is normal to get different results when we use our model with BVLC caffe compared to Openvino IRs of that model.
Sorry I have made a mistake in Heading/Subject of this question. Correct subject is ::
"Re: Difference in results for mobilenet v1 in BVLC caffe and Alexnet converted IR in openvino 2020.3.194."
It is recommended for you to use the latest OpenVINO toolkit version (2021.1) since it has the latest updates.
You may refer here for the guide.
Please get back to us after you had tested this with the latest toolkit version, whether the result is better or the same problem persists.
Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question.
Thanks a lot for your response. I was bit busy with other activities related to intel openvino. We are applying deployment solutions using openvino, so I got involved into some other things related to that.
As suggested by you, I will try to implement same experiments using latest version of openvino and will inform you about the results before 19 December 2020.
Till then, it will be really helpful for me if you continue this thread .
Software Developer, Vehant Technologies Pvt Ltd.