Mmdetection: How to set up configs for fine tuning (transfer learning)

[wrkgm]

Hi, I am having trouble doing transfer learning with mmdetection. I have two datasets:
A - Synthetic dataset. 100k images, 1 class. Used for pretraining
B - Real world dataset. 3k images, 1 class. Used for transfer learning.

I've previously been using TF object detection API. My procedure has been: pretrain on A, then transfer learning on B. This has worked well on TF object detection API. When trying the same thing with mmdetection, I get good convergence for pretraining, but when I do transfer learning, the loss does not converge at all, it begins at 0.2 and rises to around 0.35.

My question: is there something I am forgetting to do in order to switch to transfer learning? Here are the modifications I made to the config for transfer learning:

• Switch data config to point at new dataset
• Switch "img_scale" argument to account for size of new images
• Switch "img_norm_cfg" to have the image stats of the new images
• Switch "frozen_stages" from -1 to 4
• Add "pretrained" arugment to point at epoch_30.pth from pretraining.

I've also tried setting frozen_stages to everything from 1-4, with no change. I've tried using learning rates 0.02, 0.002, and 0.0002, with little change, the loss just changes more slowly, but still fails to ever drop.

I feel that I am missing something obvious here in configs, since the same datasets with the same network (Faster RCNN with resnet 101) have worked well using TF api.

Any suggestions? Thanks!! :)

[hellock]

You may just follow what you did with TF, but if you have any specific questions on the apis or configs of mmdetection, we can provide some help.
frozen_stages means freezing the stages of ResNe(X)ts.
pretrained only affects the backbone.

[wrkgm]

Thank you! When doing transfer learning, is there anything else I should do besides these? I wonder if I'm missing a step and that is why I see poor performance. TF does all these things for you itself when transfer learning.