Flame
Flame
Do you know how to allow amp when doing quantization?
I am trying to do int8 quantization use quanto library, but when inference I will get error. https://github.com/huggingface/optimum-quanto
You could use infer.sh, and modify the parameter as you want. vp means visualization path you want to output, tdp mean the image you want to used for segmenting. python...
Which model setting are you using? Can you show me your model setting in buildsam3d? @noahrecher
Can you try to make line 196 "skip layer = 2" in buildsam3d.py ?
Can you try to use this building method temporarily? https://github.com/arcadelab/FastSAM3D_slicer/blob/main/fastsam%2Fsegment_anything%2Fbuild_sam3D.py I will update new checkpoints soon.
The flash attention part is just used for inference, not for distillation. You could feel free to use flash attention for your inference.
You are correct except one point: You could use our checkpoint to Inference, it supports flash attention.
> Thank you for your quick reply! I am not familiar to flash attention, so it is maybe a silly question: Based on your answer and the code I think...
> I have another question regarding to the distillation process: From `utils/prepare_nnunet.py` the images and labels from one dataset shall be stored under `label_name/dataset_name/imagesTr` and `label_name/dataset_name/labelsTr`, but `preparelabel.py` and `validation.py`...