Issues with using pre-trained weights Resnet50.npy with RetinaNet

xmyqsh / RetinaNet

Focal loss for Dense Object Detection

MIT License

34 stars 5 forks source link

Issues with using pre-trained weights Resnet50.npy with RetinaNet #2

Open dunyasik opened 7 years ago

dunyasik commented 7 years ago

Hi, I have tried several pre-trained Resnet50.npy models to train the RetinaNet using the exact commands provided in the readme here. However, after 40 50 iterations the loss becomes NaN everytime.

Example pre-trained model I tried: https://lexiondebug.blob.core.windows.net/mlmodel/models/res50.npy

Could you provide the pre-train model you use to train your models, or please guide me what else could be the problem here? Thanks alot

xmyqsh commented 6 years ago

https://drive.google.com/file/d/0B_xFdh9onPagSWU1ZTAxUTZkZTQ/view

dunyasik commented 6 years ago

Thanks a lot!!! seems to be working fine. will provide further feedback when training finishes.