ly19965 / EasyFace

Apache License 2.0
73 stars 13 forks source link

Unable to replicate stated training performance , Full Model training sample settings for WiderFace #3

Open Anguliachao opened 1 year ago

Anguliachao commented 1 year ago

Dear Author @ly19965 , I've tried with ur sample code in the repository, the WiderFace-mini works well with ur sample code , however I found when it comes to full-WiderFace set training, either I face loss goes nan or the final mAP is only 0.18. Thus I'd like to ask a few questions:

1. what exactly the parameter settings when it comes to full-widerface training(not the 1 epoch widerface-mini), while stated lr=0.02 in modelscope website but written  lr=0.01 in the Damo_lms.py?
2.  would u plan to give a full sample for that full-dataset training? If not, would u mind guiding me roughly of the full-dataset training setting? 
3. what exactly is mmcv_tiny.py in the modelscope website?

Many thanks in advance and look forward to your reply