ActonMartin
commented
4 years ago It's a strange thing. After I run the clone code of your repo. The model size is about 20M. SAME net , input,But different size My code. And this time, the model produced by your repo could be opened by a compression software.(The download file from Baidu_yun could load by torch,but can't open,I don't know the reason.) When I open the file produced by your repo, the numbers in archive\data\ is different from mine. Maybe we also have some difference in some places of the code. And maybe could explain why so smaller the paper's issue.
I am surprised that the model size of yolo nano's network is so small. I did not find the official implementation code on the preprint website:yolo_nano. In your repository, I saw that your saved network is about 20M. I also saved the state_dict information of the network model, and then I checked the keys and values in my model and yours, and found that they are the same. But there are more than 50 M saved in mine. What is even more puzzled is that the model file I saved can be opened with a compression software, but your model file cannot be opened with a compression software. So my question is what trick did you use to save this model? Finally, I am still very interested in only 4M model files in the paper:yolo_nano.