megvii-research/AnchorDETR

Load pretrained model AnchorDETR-DC5.

Closed this issue · 2 comments

I'm trying to train with my custom data from your pretrained model.
I tested successfully with pre-trained model AnchorDETR-C5 with num_feature_levels = 1
.
But when I try with AnchorDETR-DC5 with num_feature_levels=3 . I get this error:
RuntimeError: Error(s) in loading state_dict for AnchorDETR: size mismatch for input_proj.0.0.weight: copying a param with shape torch.Size([256, 2048, 1, 1]) from checkpoint, the shape in current model is torch.Size([256, 512, 3, 3]).
When I check, it seems that pretrained model for AnchorDETR-DC5 just use num_feature_levels=1.
When I look at file size, they are the same. I suppose if model is trained with multi level feature, input_proj layer should be larger -> model size is larger.

So released AnchorDETR-DC5 is just for num_feature_levels=1.
Is this correct? If so, can you release pretrained for num_feature_levels=3 too?

Hi, AnchorDETR-DC5 only use the single ResNet-DC5 feature, so that you should not set the num_feature_levels to 3.
The code with multiple feature levels is just a simple example which should not be optimal, and we will not release the model with multiple level features.