Repository files navigation

为了方便3D医学图像相关的任务, 实现了主要的计算机视觉算子,例如RoIAlign, NMS 等,并且基于这些算子实现了多种数据增强的方法.

Computer vision operators in medical image, such as RoIAlign, DCNv1, DCNv2 and NMS for both 2/3D images. And based on these operators, a variety of data augmentation methods have been implemented.

see medvision/csrc and medvision/ops

Torch :implemented with torch functions.

see medvision/aug_cuda and medvision/aug_cuda_batch

or medvision/aug_cpu for numpy version

All of these will support forward and backward (need test).

Methods	Cuda
RandomAffine
RandomScale
RandomShift
RandomRotate
RandomFlip
CropRandom Series
RandomElasticDeformation
Resize
Pad
RandomBlur
RandomNoise

# run and check you cuda and torch# make sure your torch.version.cuda == cuda_home_versionpython -m torch.utils.collect_env# if neededexport PATH="path_to/gcc-5.4+/bin/:$PATH"export CUDA_HOME="/path_to/cuda-9.0+"git clone https://github.com/TimothyZero/MedVisioncd MedVisionpython setup.py develop# recommended# orpip install -e.# -e : editable, sometimes may cause cpu 100%

*: kaggle T4

1. AT_CHECK was not declared in this scope

For torch 1.5+,AT_CHECK is replaced withTORCH_CHECK, so if your torch version > 1.5 ,

#ifndef AT_CHECK#defineAT_CHECK TORCH_CHECK#endif

at the beginning of the .cu code.

2. debug in CUDA

#include<stdio.h>printf("Hello from block %d, thread %d\n",a,b);

3. .contiguous() is very import in roi align!

4. CUDA Too many resources requested for launch

   "Too Many Resources Requested for Launch - This error means that the number of registers available on the multiprocessor is being exceeded. Reduce the number of threads per block to solve the problem."

This framework is published under the Apache License Version 2.0.

https://github.com/XinyiYing/D3Dnet

https://github.com/open-mmlab/mmdetection

https://github.com/MIC-DKFZ/medicaldetectiontoolkit