astra-vision / densemtl Goto Github PK

I'm interesting in it

congratulations for great job
for the easy apply to many other experiments on multitask learning, can you provide the independent code of xTAM module?
for example: given x (B, C, H, W), y (B, C, H, W) are features of semantic and depth respectively, how I can build xTAM(x, y)?

Use ECA or SE attention?

Thanks for your inspiring work.
I find that you construct the Cross-task Attention module in ours_xatt.py. It seems that this module is composed of SxTAM and CxTAM. In the SxTAM module, the cross-attention between x and y is computed. But there is no self-attn or sigmoid function in the SxTAM class (from the xTAM module of the paper) and I can not find the CxTAM function in your code base.

**{f'CxTAM_{s}': CxTAM() if self.use_CxTAM else nn.Identity() for s in others(t)},

Thank you for your great work. May I ask how to use images for simple inference？It seems that only evaluation and training can be done now.

when I run

python main.py \
  --base=configs/da/sy2cs/fs_bs2.yml \
  --config configs/da/sy2cs/resnet101_ours_SD.yml

Traceback (most recent call last):
  File "main.py", line 2, in <module>
    from utils.logs import cpprint
ModuleNotFoundError: No module named 'utils.logs'

I think utils miss logs module.