Merge branch 'master' into adds-codeformat

.github/workflows/chatops.yml

@@ -11,6 +11,7 @@ jobs:
       - name: dispatch
         uses: peter-evans/[email protected]
         with:
+          token: ${{ secrets.PR_MAINTAIN_WYLI }}
           # reaction-token: ${{ secrets.GITHUB_TOKEN }}
           reactions: false
           config: >

.pre-commit-config.yaml

@@ -0,0 +1,63 @@
+default_language_version:
+  python: python3.8
+
+ci:
+  autofix_prs: true
+  autoupdate_commit_msg: '[pre-commit.ci] pre-commit suggestions'
+  autoupdate_schedule: quarterly
+  # submodules: true
+
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.0.1
+    hooks:
+      - id: end-of-file-fixer
+      - id: trailing-whitespace
+      - id: check-yaml
+      - id: check-docstring-first
+      - id: check-executables-have-shebangs
+      - id: check-toml
+      - id: check-case-conflict
+      - id: check-added-large-files
+        args: ['--maxkb=1024']
+      - id: detect-private-key
+
+  #- repo: https://github.com/asottile/pyupgrade
+  #  rev: v2.23.2
+  #  hooks:
+  #    - id: pyupgrade
+  #      args: [--py36-plus]
+  #      name: Upgrade code
+
+  #- repo: https://github.com/asottile/yesqa
+  #  rev: v1.2.3
+  #  hooks:
+  #    - id: yesqa
+  #      name: Unused noqa
+
+  #- repo: https://github.com/PyCQA/isort
+  #  rev: 5.9.3
+  #  hooks:
+  #    - id: isort
+  #      name: Format imports
+
+  # - repo: https://github.com/psf/black
+  #   rev: 21.7b0
+  #   hooks:
+  #     - id: black
+  #       name: Format code
+
+  #- repo: https://github.com/executablebooks/mdformat
+  #  rev: 0.7.8
+  #  hooks:
+  #    - id: mdformat
+  #      additional_dependencies:
+  #        - mdformat-gfm
+  #        - mdformat_frontmatter
+  #      exclude: CHANGELOG.md
+
+  # - repo: https://github.com/PyCQA/flake8
+  #   rev: 3.9.2
+  #   hooks:
+  #     - id: flake8
+  #       name: Check PEP8

DiNTS/README.md

@@ -0,0 +1,115 @@
+# DiNTS
+
+## Overview
+
+Repository for Paper "DiNTS: Differentiable Neural Network Topology Search for 3D Medical Image Segmentation" (CVPR'21)
+
+## Installation
+
+The code was tested with Anaconda and Python 3.7. After installing the Anaconda environment:
+
+Clone the repo:
+
+Install dependencies:
+For PyTorch dependency, see pytorch.org for more details.
+Install custom dependencies, horovod and openmpi e.t.c:
+
+Use docker files:
+Create a Dockerfile with
+```
+FROM nvcr.io/nvidian/pytorch:20.03-py3
+RUN HOROVOD_NCCL_LINK=SHARED HOROVOD_NCCL_LIB=/usr/lib/x86_64-linux-gnu HOROVOD_NCCL_INCLUDE=/usr/include HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL HOROVOD_WITH_PYTORCH=1 pip install --no-cache-dir --upgrade horovod
+RUN apt install graphviz
+RUN pip install graphviz
+RUN pip install torchviz
+RUN pip install tensorboardx
+RUN pip install monai
+```
+
+Then
+```
+docker build -f Dockerfile -t nvcr.io/nvidian/pytorch:20.03-py3-horovod .
+```
+
+Run into Docker:
+```
+sudo docker run -it --gpus all --pid=host --shm-size 8G -v /home/yufan/Projects/:/workspace/Projects/ nvcr.io/nvidian/pytorch:20.03-py3-horovod
+```
+
+## Commands
+
+All the scripts are in `./code`, `main.sh` will call other
+scripts to perform search (`search.sh`), retrain (`retrain.sh`), test on validation results (`test_val.sh`), inference on test data (`test.sh`) and fuse (`fuse.sh`) all 5 fold results. Need to change
+variable values within submit_all.sh to run your specific experiments.
+The general pipeline is:
+
+search the model on each task independently and results will be saved in
+```
+/workspace/Projects/darts-oneshot/search-bilevel-ent-mem${MEM}-${TASK}
+```
+retrain the model given `${TASK}`, `${MODEL}`, `${FOLD}`, results will be saved in the save folder as the searched model folder
+Record the best retrained model manually and change the numbers (BESTTask01 e.t.c) in `main.sh`
+Test on validation results on test results
+The test results will be saved in each retraining folder
+Fuse the five fold results and the results will be saved in designated folder
+(Needs to be combined with step 6) The fused results are upsampled results, needs to resample to the original nii data size. I downloaded the fused results to my local workstation and resampled it (using fuse.py). The path to the original msd dataset needs to be provided.
+
+The above operations are performed by change the value of N in `main.sh` and run
+bash `main.sh`
+Also you need to change the values of TASKS, FOLDS e.t.c to run specific experiments
+
+### Argument Explanation
+The arguments can also be found in config.py. Here are some quick references:
+
+- iters is the total searching or retraining iteration. warmup_iters in searching is the iterations where architecture is not updated (the lr_scheduler is working, the temperature is not annealed).
+- EF is the temperature annealing scheme. 0 is no annealing, 2 is linear decay
+- EP=1, EF=0, topology=1 means no temperature annealing, use topology loss and entropy loss (not degrading binarized model validation result).
+- base_lr * gpu number will be the final network weights lr
+- opset is the operation set. opset 0, and setting cellop=4 will be the same with C2FNAS. opset1 includes 5 operations, 3D conv, 3 psudo 3d. Details can be found in operation.py
+- path_iters is the number of iteration to print out architecture informations. Also in sampling based searching strategy, this is the training iteration for each sampled sub-model
+- path_num is the number of pathes per block during sampling based training
+- use_max will use one-shot supernet training, not sampling based method (should always add this one if training the whole supernet)
+- there are several other arguments like finetune, retrain (deprecated), ef_end. those are related to binarize the model in searching stage and then finetune it. We are not using it since we double the channel number during retraining.
+
+## Utilities
+
+To further fuse the output segmetnation masks, user needs to create a `.json` file describing shape of each image for resizing/re-sampling. An example is shown in `./code/msd_size.json`.
+```
+{
+    "Task07_Pancreas": {
+        "pancreas_044.nii.gz": [
+            512,
+            512,
+            87
+        ],
+        "pancreas_144.nii.gz": [
+            512,
+            512,
+            48
+        ],
+        "pancreas_039.nii.gz": [
+            512,
+            512,
+            87
+        ]
+    }
+}
+```
+
+## Bibtex
+```
+@inproceedings{he2021dints,
+  title={DiNTS: Differentiable Neural Network Topology Search for 3D Medical Image Segmentation},
+  author={He, Yufan and Yang, Dong and Roth, Holger and Zhao, Can and Xu, Daguang},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+  pages={5841--5850},
+  year={2021}
+}
+@inproceedings{yu2020c2fnas,
+  title={C2fnas: Coarse-to-fine neural architecture search for 3d medical image segmentation},
+  author={Yu, Qihang and Yang, Dong and Roth, Holger and Bai, Yutong and Zhang, Yixiao and Yuille, Alan L and Xu, Daguang},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+  pages={4126--4135},
+  year={2020}
+}
+```

DiNTS/configs/config_Task01_BrainTumour.yaml

@@ -0,0 +1,32 @@
+augmentation_monai:
+  aug_0: RandRotated|keys~image,label|range_x~0.3|range_y~0.3|range_z~0.3|mode~bilinear,nearest|prob~0.2
+  aug_1: RandZoomd|keys~image,label|min_zoom~0.8|max_zoom~1.2|mode~trilinear,nearest|align_corners~True,None|prob~0.16
+  aug_2: RandGaussianSmoothd|keys~image|sigma_x~0.5,1.15|sigma_y~0.5,1.15|sigma_z~0.5,1.15|prob~0.15
+  aug_3: RandScaleIntensityd|keys~image|factors~0.3|prob~0.5
+  aug_4: RandShiftIntensityd|keys~image|offsets~0.1|prob~0.5
+  aug_5: RandGaussianNoised|keys~image|std~0.01|prob~0.15
+  aug_6: RandFlipd|keys~image,label|spatial_axis~0|prob~0.5
+  aug_7: RandFlipd|keys~image,label|spatial_axis~1|prob~0.5
+  aug_8: RandFlipd|keys~image,label|spatial_axis~2|prob~0.5
+core:
+  amp: true
+  deterministic: false
+  foreground_crop_margin: 0
+  infer_num_sw_batch_size: 6
+  infer_num_tta: 1
+  infer_overlap_ratio: 0.625
+  infer_patch_size: 96,96,96
+  input_channels: 4
+  intensity_norm: NormalizeIntensityd|keys~image|nonzero~True|channel_wise~True
+  interpolation: Identityd|keys~image,label
+  learning_rate: 0.025
+  learning_rate_milestones: 0.2,0.4,0.6,0.8
+  num_epochs: 1695
+  num_epochs_per_validation: 80
+  num_images_per_batch: 2
+  num_patches_per_image: 1
+  num_sw_batch_size: 6
+  output_classes: 4
+  overlap_ratio: 0.625
+  patch_size: 96,96,96
+  random_seed: 888

DiNTS/configs/config_Task02_Heart.yaml

@@ -0,0 +1,32 @@
+augmentation_monai:
+  aug_0: RandRotated|keys~image,label|range_x~0.3|range_y~0.3|range_z~0.3|mode~bilinear,nearest|prob~0.2
+  aug_1: RandZoomd|keys~image,label|min_zoom~0.8|max_zoom~1.2|mode~trilinear,nearest|align_corners~True,None|prob~0.16
+  aug_2: RandGaussianSmoothd|keys~image|sigma_x~0.5,1.15|sigma_y~0.5,1.15|sigma_z~0.5,1.15|prob~0.15
+  aug_3: RandScaleIntensityd|keys~image|factors~0.3|prob~0.5
+  aug_4: RandShiftIntensityd|keys~image|offsets~0.1|prob~0.5
+  aug_5: RandGaussianNoised|keys~image|std~0.01|prob~0.15
+  aug_6: RandFlipd|keys~image,label|spatial_axis~0|prob~0.5
+  aug_7: RandFlipd|keys~image,label|spatial_axis~1|prob~0.5
+  aug_8: RandFlipd|keys~image,label|spatial_axis~2|prob~0.5
+core:
+  amp: true
+  deterministic: false
+  foreground_crop_margin: 0
+  infer_num_sw_batch_size: 6
+  infer_num_tta: 1
+  infer_overlap_ratio: 0.625
+  infer_patch_size: 96,96,96
+  input_channels: 1
+  intensity_norm: NormalizeIntensityd|keys~image|nonzero~True|channel_wise~True
+  interpolation: Spacingd|keys~image,label|pixdim~1.0,1.0,1.0|mode~bilinear,nearest|align_corners~True,True
+  learning_rate: 0.025
+  learning_rate_milestones: 0.2,0.4,0.6,0.8
+  num_epochs: 20800 #4GPU
+  num_epochs_per_validation: 1000 #4GPU
+  num_images_per_batch: 2
+  num_patches_per_image: 1
+  num_sw_batch_size: 6
+  output_classes: 2
+  overlap_ratio: 0.625
+  patch_size: 96,96,96
+  random_seed: 888

DiNTS/configs/config_Task03_Liver.yaml

@@ -0,0 +1,32 @@
+augmentation_monai:
+  aug_0: RandRotated|keys~image,label|range_x~0.3|range_y~0.3|range_z~0.3|mode~bilinear,nearest|prob~0.2
+  aug_1: RandZoomd|keys~image,label|min_zoom~0.8|max_zoom~1.2|mode~trilinear,nearest|align_corners~True,None|prob~0.16
+  aug_2: RandGaussianSmoothd|keys~image|sigma_x~0.5,1.15|sigma_y~0.5,1.15|sigma_z~0.5,1.15|prob~0.15
+  aug_3: RandScaleIntensityd|keys~image|factors~0.3|prob~0.5
+  aug_4: RandShiftIntensityd|keys~image|offsets~0.1|prob~0.5
+  aug_5: RandGaussianNoised|keys~image|std~0.01|prob~0.15
+  aug_6: RandFlipd|keys~image,label|spatial_axis~0|prob~0.5
+  aug_7: RandFlipd|keys~image,label|spatial_axis~1|prob~0.5
+  aug_8: RandFlipd|keys~image,label|spatial_axis~2|prob~0.5
+core:
+  amp: true
+  deterministic: false
+  foreground_crop_margin: 0
+  infer_num_sw_batch_size: 6
+  infer_num_tta: 1
+  infer_overlap_ratio: 0.625
+  infer_patch_size: 96,96,96
+  input_channels: 1
+  intensity_norm: ScaleIntensityRanged|keys~image|a_min~-21.0|a_max~189.0|b_min~0.0|b_max~1.0|clip~True
+  interpolation: Spacingd|keys~image,label|pixdim~1.0,1.0,1.0|mode~bilinear,nearest|align_corners~True,True
+  learning_rate: 0.025
+  learning_rate_milestones: 0.2,0.4,0.6,0.8
+  num_epochs: 6400
+  num_epochs_per_validation: 305
+  num_images_per_batch: 2
+  num_patches_per_image: 1
+  num_sw_batch_size: 6
+  output_classes: 3
+  overlap_ratio: 0.625
+  patch_size: 96,96,96
+  random_seed: 888

DiNTS/configs/config_Task04_Hippocampus.yaml

@@ -0,0 +1,32 @@
+augmentation_monai:
+  aug_0: RandRotated|keys~image,label|range_x~0.3|range_y~0.3|range_z~0.3|mode~bilinear,nearest|prob~0.2
+  aug_1: RandZoomd|keys~image,label|min_zoom~0.8|max_zoom~1.2|mode~trilinear,nearest|align_corners~True,None|prob~0.16
+  aug_2: RandGaussianSmoothd|keys~image|sigma_x~0.5,1.15|sigma_y~0.5,1.15|sigma_z~0.5,1.15|prob~0.15
+  aug_3: RandScaleIntensityd|keys~image|factors~0.3|prob~0.5
+  aug_4: RandShiftIntensityd|keys~image|offsets~0.1|prob~0.5
+  aug_5: RandGaussianNoised|keys~image|std~0.01|prob~0.15
+  aug_6: RandFlipd|keys~image,label|spatial_axis~0|prob~0.5
+  aug_7: RandFlipd|keys~image,label|spatial_axis~1|prob~0.5
+  aug_8: RandFlipd|keys~image,label|spatial_axis~2|prob~0.5
+core:
+  amp: true
+  deterministic: false
+  foreground_crop_margin: 0
+  infer_num_sw_batch_size: 6
+  infer_num_tta: 1
+  infer_overlap_ratio: 0.625
+  infer_patch_size: 32,32,32
+  input_channels: 1
+  intensity_norm: NormalizeIntensityd|keys~image|nonzero~True|channel_wise~True
+  interpolation: Identityd|keys~image,label
+  learning_rate: 0.025
+  learning_rate_milestones: 0.2,0.4,0.6,0.8
+  num_epochs: 3200
+  num_epochs_per_validation: 150
+  num_images_per_batch: 2
+  num_patches_per_image: 1
+  num_sw_batch_size: 6
+  output_classes: 3
+  overlap_ratio: 0.625
+  patch_size: 32,32,32
+  random_seed: 888

DiNTS/configs/config_Task05_Prostate.yaml

@@ -0,0 +1,32 @@
+augmentation_monai:
+  aug_0: RandRotated|keys~image,label|range_x~0.3|range_y~0.3|range_z~0.3|mode~bilinear,nearest|prob~0.2
+  aug_1: RandZoomd|keys~image,label|min_zoom~0.8|max_zoom~1.2|mode~trilinear,nearest|align_corners~True,None|prob~0.16
+  aug_2: RandGaussianSmoothd|keys~image|sigma_x~0.5,1.15|sigma_y~0.5,1.15|sigma_z~0.5,1.15|prob~0.15
+  aug_3: RandScaleIntensityd|keys~image|factors~0.3|prob~0.5
+  aug_4: RandShiftIntensityd|keys~image|offsets~0.1|prob~0.5
+  aug_5: RandGaussianNoised|keys~image|std~0.01|prob~0.15
+  aug_6: RandFlipd|keys~image,label|spatial_axis~0|prob~0.5
+  aug_7: RandFlipd|keys~image,label|spatial_axis~1|prob~0.5
+  aug_8: RandFlipd|keys~image,label|spatial_axis~2|prob~0.5
+core:
+  amp: true
+  deterministic: false
+  foreground_crop_margin: 0
+  infer_num_sw_batch_size: 6
+  infer_num_tta: 1
+  infer_overlap_ratio: 0.625
+  infer_patch_size: 96,96,32
+  input_channels: 2
+  intensity_norm: NormalizeIntensityd|keys~image|nonzero~True|channel_wise~True
+  interpolation: Spacingd|keys~image,label|pixdim~1.0,1.0,1.0|mode~bilinear,nearest|align_corners~True,True
+  learning_rate: 0.025
+  learning_rate_milestones: 0.2,0.4,0.6,0.8
+  num_epochs: 12800 # 4GPU
+  num_epochs_per_validation: 625 # 4GPU
+  num_images_per_batch: 2
+  num_patches_per_image: 1
+  num_sw_batch_size: 6
+  output_classes: 3
+  overlap_ratio: 0.625
+  patch_size: 96,96,32
+  random_seed: 888

DiNTS/configs/config_Task06_Lung.yaml

@@ -0,0 +1,32 @@
+augmentation_monai:
+  aug_0: RandRotated|keys~image,label|range_x~0.3|range_y~0.3|range_z~0.3|mode~bilinear,nearest|prob~0.2
+  aug_1: RandZoomd|keys~image,label|min_zoom~0.8|max_zoom~1.2|mode~trilinear,nearest|align_corners~True,None|prob~0.16
+  aug_2: RandGaussianSmoothd|keys~image|sigma_x~0.5,1.15|sigma_y~0.5,1.15|sigma_z~0.5,1.15|prob~0.15
+  aug_3: RandScaleIntensityd|keys~image|factors~0.3|prob~0.5
+  aug_4: RandShiftIntensityd|keys~image|offsets~0.1|prob~0.5
+  aug_5: RandGaussianNoised|keys~image|std~0.01|prob~0.15
+  aug_6: RandFlipd|keys~image,label|spatial_axis~0|prob~0.5
+  aug_7: RandFlipd|keys~image,label|spatial_axis~1|prob~0.5
+  aug_8: RandFlipd|keys~image,label|spatial_axis~2|prob~0.5
+core:
+  amp: true
+  deterministic: false
+  foreground_crop_margin: 0
+  infer_num_sw_batch_size: 6
+  infer_num_tta: 1
+  infer_overlap_ratio: 0.625
+  infer_patch_size: 96,96,96
+  input_channels: 1
+  intensity_norm: ScaleIntensityRanged|keys~image|a_min~-1000.0|a_max~1000.0|b_min~0.0|b_max~1.0|clip~True
+  interpolation: Spacingd|keys~image,label|pixdim~1.0,1.0,1.0|mode~bilinear,nearest|align_corners~True,True
+  learning_rate: 0.025
+  learning_rate_milestones: 0.2,0.4,0.6,0.8
+  num_epochs: 12800
+  num_epochs_per_validation: 625
+  num_images_per_batch: 2
+  num_patches_per_image: 1
+  num_sw_batch_size: 6
+  output_classes: 2
+  overlap_ratio: 0.625
+  patch_size: 96,96,96
+  random_seed: 888