FlowFormer-Official
FlowFormer-Official copied to clipboard
drinkingcoder
Just made these changes: 1. written DDP code for training and validation for faster training and validation. 2. resolved bug in `gma.py` and `encoders.py` (these modules were not used while...
Hi @drinkingcoder, First of all, thank you for your wonderful research. I am interested in training the FlowFormer model in custom data. Currently, it seems that training codes are only...
Thank you for releasing your model code! I am really appreciating it. I am trying to train FlowFormer on FlyingThings3D dataset, but even with batch size reduced to 1 and...
Hi, Thanks for your awesome work. What is the reason of setting the num_workers=128 (so high) ?? Is it a typo?? [at this line](https://github.com/drinkingcoder/FlowFormer-Official/blob/6ba7ea82b45394d3a7a25808399695427c9febd8/core/datasets.py#L231) @drinkingcoder
Hi, thanks for this remarkable work. When I try to generate the Sintel submission results with evaluate_FlowFormer_tile.py, an error occurred: module 'datasets' has no attribute 'MpiSintel_submission'. I checked the datasets.py...
Thank you for sharing the code. I wonder where I can find the checkpoint for the model stated as C+T+S+K+H in the paper.
Does chairs.pth mean trained on FC with prerained weights of Imagenet or w/o things.pth === chairs+things sintel.pth === chairs+things+sintel kitti ==== chairs+things+sintel+kitti+hd1k things_kitti ==== chairs+things+kitti flow_former_small ==== small version of...
just need it for some comparisons. Thank you so much for sharing this great project
I am using 4xV100x32GB GPU to train with the "small_things_eval.py" config file, and using the "C+T+K+S" datasets strategy. The batch size is max set to 4, and I have also...
Hello, in the `MemoryEncoder Class`, it seems that there should be added the definition of `self.layers`. In the `if self.cfg.feat_cross_attn` part of `forward()` function, `self.layers` is used without definition.
