search

mathmanu / caffe-jacinto

This repository has moved. The new link can be obtained from https://github.com/TexasInstruments/jacinto-ai-devkit
116 stars 35 forks source link

Incorrect results for SSD Mobilenet model trained on caffe (forked from ssd caffe by weiliu) #35

Open chandrakantkhandelwal opened 5 years ago

chandrakantkhandelwal commented 5 years ago

I tried to run SSD Mobilentv1 model as per the instructions mentioned in the following git repo: https://github.com/chuanqi305/MobileNet-SSD (just run the demo.py file, trained model and prototxt is given in the repo)

I am getting different (random and negative coordinates) output when I ran this code using caffe-jacinto. It works well with: https://github.com/weiliu89/caffe

Please let me know if I am doing something wrong!

mathmanu commented 5 years ago

Could you please try running with this repository:

https://github.com/NVIDIA/caffe

And let us know whether you are getting correct results or not?

chandrakantkhandelwal commented 5 years ago

@mathmanu checked with nvidia caffe, it's same output as https://github.com/weiliu89/caffe. Note: I had to uncomment engine: CAFFE line in the prototxt for it to run on nvidia as well as ssd caffe.

mathmanu commented 5 years ago

One question (for me to uderstand). When did you last pull the code from caffe-jacinto?

chandrakantkhandelwal commented 5 years ago

12th March 2019

chandrakantkhandelwal commented 5 years ago

@mathmanu I used the given trained model as starting point and then re-trained it on pascal voc dataset using caffe-jacinto for few epochs. Now it is working fine.

mathmanu commented 5 years ago

It probably means that your original training did use pretrained model correctly. You can probably compare the logs of your working training and not working to check this.