Update the codes to latest mmdetection.

[freshn]

@liuzili97 @Turoad

Motivation
Update current codes into the latest mmdetection is helpful.

Additional context
By simply copying the model and config files and modifying import settings, my implementation fails to converge during the training. The mAP is in fact near 0. Could you give some suggestions that transfer your codes into the latest version?

[Turoad]

Hi freshn,

Thanks for your attention. I doesn't follow the latest version of mmdetection currently.
If you want to update codes, I suggest you can run the validate first. You'd better download the trained models and convert it into latest version of mmdetection. You can reproduce the result first, then try training the model.

[freshn]

@Turoad Thank you for your suggestions. I tried to evaluate the model. However, I found that the result was not good. The checkpoint I used is the one provided in this repository, specifically, ttfnet18_1x-fe6884.pth. The result is shown below. Do you have any suggestions?
@liuzili97

[freshn]

I got reasonable results by modifying the names in the state_dict. Based on that, the zero mAP may comes from the different names in the latest bbox_head.

[HappyPeanuts]

@freshn @Turoad Have you meet sucn an error: ImportError: cannot import name 'deform_conv_cuda' from 'ops.dcn.src' (unknown location)？I try to use TTFNet and TTFHead in my work,but meet such an error, i want to know how to resolve this problem. thanks a lot, looking forward to yor reply.

[goldentimecoolk]

Hi freshn,

Thanks for your attention. I doesn't follow the latest version of mmdetection currently. If you want to update codes, I suggest you can run the validate first. You'd better download the trained models and convert it into latest version of mmdetection. You can reproduce the result first, then try training the model.

@freshn @Turoad Hi, could you please share the trained model? The public links are unavailable. By the way, I convert ttfnet.py and ttf_head.py to mmdet 2.20.0, but the mAP is close to zero. When I trained other models defined in mmdet, it is normal. I'm trying to find the bug.