BUGs in training

[romeosgood]

Hello, thanks for this amazing work.

However, I met several problems on reproduction and training.

1. When I try to reproduce this work, the best result of 3D detection is [24.5, 18.4, 15.7] which is far behind your result.
2. When I try to replace dla34 with other backbone, such as hrnet, the loss is not convergent.
3. When I try to train with large backbone, I have to reduce the ’samples_per_gpu‘, I meet a bug called 'assert pred.size() == target.size() and target.numel() > 0', for example when ’samples_per_gpu‘ equals 2 or 3 or 4.
4. 'assert pred.size() == target.size() and target.numel() > 0' is a common bug when I do reproduction and training. I check the params and find that target.size() = [0,2], and have no idea what happened.

plz help me with this issues. Looking forward your reply. Thank you.

[sverdoot]

Actually you can either ignore the lines from 3rd and 4th points or replace target.numel() > 0 with target.numel() >= 0.

[romeosgood]

Actually you can either ignore the lines from 3rd and 4th points or replace target.numel() > 0 with target.numel() >= 0.

Thx for your advice! I'll try a second time.