Jerry Jiarui XU
Jerry Jiarui XU
> hi @xvjiarui > have u tried `GN` instead of `Sync BN`? Sorry, I haven't tried `GN`. It would be very kind of you if you could try it and...
Since we also need to perform experiments on ImageNet and Kinetics classification task, the `context_block` is only inserted into backbone.
Hi, @nemonameless I am sorry I couldn't reproduce your result. May I have you BN or GN config file?
Sorry for the late reply. We have updated the results in table. Models will be released very soon. @walton-wang929 As mentioned, SyncBN has some stability issue. Your could try to...
Sorry to bring this up again sooo late. FYI, model w/ and w/o sync BN is provided 4 months ago.
Sorry for the late reply. Please find the instantiation of gc block [here](https://github.com/xvjiarui/GCNet/blob/master/mmdet/ops/ct/context_block.py). As for the whole network, it's simply ResNet structure with gc block after each residual block.
> HELLO，I only find 2D GCNet, but I can't find 3D GCNet, Only 2D GC block is provided in this repo. 3D GCNet could be easily extended from 2D version....
Hi @SuperAttitude We didn't have 1d code in this repo. But you could change the conv2d to conv1d and modify the shape.
Hi @Shiro-LK , Sorry for the late reply. Yes, your implementation is correct.
Sorry for the late reply The attention across time is relative hard to visualize. From the Table 1 in the paper, the attention on Kinetics seems to be a little...