applezy8866

When using auto_fp16(), the inference result is different

3

Hello, When using auto_fp16() to wrap the model, with self.fp16_enabled = False during training, the inference results are different, I don't know why, because it seems auto_fp16() won't work when...

Questions about TemporalSelfAttention

2

I found when doing temporal alignment in TemporalSelfAttention, you rotate the previous bev feature and translate the reference points. Why not rotate and translate the previous bev feature or rotate...

Why use additional one frame when training?

1

Dear, I found that in dataset.py: " if not self.test_mode: self.num_frames_per_sample += 1" and in tracker.py: "num_frame = img.size(1) - 1" so what‘s the usage of the additional frame?