Code

ViT UNETR SSL#10

Code

Draft

valosekj wants to merge 100 commits intomainivadomed/model-seg-dcm:mainfrom jv/vit_unetr_sslivadomed/model-seg-dcm:jv/vit_unetr_sslCopy head branch name to clipboard

+3,146

Commits

Commits on Mar 8, 2024

• Add notebook debugging SSL pretraining transformations

  valosekj

  committed
  75a5f85

  Copy full SHA for 75a5f85

• Fix typo

  valosekj

  committed
  e37ff59

  Copy full SHA for e37ff59

• Add script for Self-Supervised Pre-training using Vision Transformer (ViT)

  valosekj

  committed
  80fd7db

  Copy full SHA for 80fd7db

• Fix typo

  valosekj

  committed
  4dc7e81

  Copy full SHA for 4dc7e81

• Do sanity dimension check after transformations

  valosekj

  committed
  3a9baef

  Copy full SHA for 3a9baef

• Use patch_size=(16, 16, 16) and batch_size = 2

  Show description for 4517f7a
  valosekj

  committed
  4517f7a

  Copy full SHA for 4517f7a

Commits on Mar 9, 2024

• rename 'crop_size' to 'spatial_size'

  valosekj

  committed
  a621a9a

  Copy full SHA for a621a9a

• use 'contrast-agnostic' data augmentation for Training of the fine-tuned model

  valosekj

  committed
  516f1ec

  Copy full SHA for 516f1ec

• Add script to finetune the 3D Single-Class Spinal Cord Lesion Segmentation Model Using SSL Pre-trained Weights

  valosekj

  committed
  fd8f943

  Copy full SHA for fd8f943

• update description

  valosekj

  committed
  d1b9c3d

  Copy full SHA for d1b9c3d

• use 'logger.info' instead of 'print'

  valosekj

  committed
  45ef357

  Copy full SHA for 45ef357

• make '--data' arg notrequired

  valosekj

  committed
  0955689

  Copy full SHA for 0955689

• explicitly specify CUDA GPU number

  valosekj

  committed
  9ea8bca

  Copy full SHA for 9ea8bca

• change sliding_window_inference roi_size and batch_size

  valosekj

  committed
  3bc999a

  Copy full SHA for 3bc999a

• remove 'to_onehot_y=True' and 'softmax=True' because we work with single channel prediction

  valosekj

  committed
  8e8a977

  Copy full SHA for 8e8a977

Commits on Mar 12, 2024

• add notebook to debug transformations when using SC masks to crop patches

  valosekj

  committed
  42394e8

  Copy full SHA for 42394e8

• rerun the notebook to show different slices

  valosekj

  committed
  622c3f6

  Copy full SHA for 622c3f6

• Add 'keys' arg

  valosekj

  committed
  59f4db0

  Copy full SHA for 59f4db0

• Add 'keys' arg

  valosekj

  committed
  a3ae050

  Copy full SHA for a3ae050

• Update 'RandCoarseDropoutd' params

  valosekj

  committed
  e1cfad1

  Copy full SHA for e1cfad1

• Update comments

  valosekj

  committed
  d4e9995

  Copy full SHA for d4e9995

• Fix imports

  valosekj

  committed
  4c35d42

  Copy full SHA for 4c35d42

• Change batch_size and NUM_WORKERS to 4

  valosekj

  committed
  05ab59a

  Copy full SHA for 05ab59a

• Make '--data' non required

  valosekj

  committed
  0b12c10

  Copy full SHA for 0b12c10

• change 'num_workers' to 0 to prevent 'RuntimeError: received 0 items of ancdata'

  valosekj

  committed
  f5881aa

  Copy full SHA for f5881aa

Commits on Mar 13, 2024

• Update input arg description

  valosekj

  committed
  7aa9858

  Copy full SHA for 7aa9858

• Rerun the notebook

  valosekj

  committed
  16691d7

  Copy full SHA for 16691d7

• Print hyper-parameters into the log file

  valosekj

  committed
  56ca187

  Copy full SHA for 56ca187

• Add '--cuda' input arg

  valosekj

  committed
  ce7315a

  Copy full SHA for ce7315a

Commits on Mar 15, 2024

• track and save epoch time

  valosekj

  committed
  1ebf833

  Copy full SHA for 1ebf833

• Plot and save input and output validation images to see how the model is learning

  valosekj

  committed
  208b2c1

  Copy full SHA for 208b2c1

• Add comment

  valosekj

  committed
  5a9e69a

  Copy full SHA for 5a9e69a

• Do not plot 'outputs_v2' as it is a hidden representation

  valosekj

  committed
  eca6616

  Copy full SHA for eca6616

• Include the epoch number as master title

  valosekj

  committed
  8e55985

  Copy full SHA for 8e55985

• Add 'torch.multiprocessing.set_sharing_strategy('file_system')' to solve number of workers issue

  Show description for e7575c5
  valosekj

  committed
  e7575c5

  Copy full SHA for e7575c5

• Create validation_figures directory if it does not exist

  valosekj

  committed
  3b6df14

  Copy full SHA for 3b6df14

• Use 3 leading zeros for the epoch number in the figures fname

  valosekj

  committed
  59b061c

  Copy full SHA for 59b061c

• Link issue

  valosekj

  committed
  f2d648b

  Copy full SHA for f2d648b

• Add note for 'RandCropByPosNegLabeld'

  Show description for a6bc6ec
  valosekj

  committed
  a6bc6ec

  Copy full SHA for a6bc6ec

• Add 'number_of_holes' arg to specify the number of holes to be used for the RandCoarseDropoutd and RandCoarseShuffled transforms

  valosekj

  committed
  c1553d6

  Copy full SHA for c1553d6

Commits on Mar 16, 2024

• typo

  valosekj

  committed
  0e802df

  Copy full SHA for 0e802df

• batch_size = 8

  valosekj

  committed
  9255855

  Copy full SHA for 9255855

• NUM_WORKERS = batch_size

  valosekj

  committed
  8296913

  Copy full SHA for 8296913

• number_of_holes=5

  Show description for 93c5415
  valosekj

  committed
  93c5415

  Copy full SHA for 93c5415

• Update transforms for training of the fine-tuned model

  Show description for a4182b5
  

  valosekj

  and

  naga-karthik

  authored
  a4182b5

  Copy full SHA for a4182b5

• update comment, remove unused imports

  valosekj

  committed
  2573578

  Copy full SHA for 2573578

• Add notebook with RandCoarseDropoutd transform debug

  valosekj

  committed
  cfb8f12

  Copy full SHA for cfb8f12

Commits on Mar 17, 2024

• Fix 'dropout_holes=True' and 'dropout_holes=False' comments

  Show description for 7f27340
  valosekj

  committed
  7f27340

  Copy full SHA for 7f27340

• remove unused 'max_spatial_size' arg

  valosekj

  committed
  74b2f2c

  Copy full SHA for 74b2f2c

• use 'fill_value=0' for 'RandCoarseDropoutd'

  valosekj

  committed
  bdb2ac1

  Copy full SHA for bdb2ac1

• Plot also RandCoarseDropoutd dropout_holes=False fill_value=0

  valosekj

  committed
  f7f3689

  Copy full SHA for f7f3689

• Add note that the batch size is actually doubled (8*2=16), because we are using two augmented samples per 3D Volume

  valosekj

  committed
  af9c6d0

  Copy full SHA for af9c6d0

Commits on Mar 18, 2024

• Add '--cuda' input arg

  valosekj

  committed
  b416135

  Copy full SHA for b416135

• Remove 'AsDiscrete'

  valosekj

  committed
  0303776

  Copy full SHA for 0303776

• Remove 'AsDiscrete'

  valosekj

  committed
  4534c57

  Copy full SHA for 4534c57

• Add 'CUDA_NUM=args.cuda'

  valosekj

  committed
  94c8611

  Copy full SHA for 94c8611

• Add TODO to increase batch_size to 16

  valosekj

  committed
  5fb325a

  Copy full SHA for 5fb325a

• Use 'roi_size' for 'define_finetune_train_transforms'

  valosekj

  committed
  d1a03c8

  Copy full SHA for d1a03c8

• Use 'label_sc' to crop samples around the SC

  Show description for e3d8086
  valosekj

  committed
  e3d8086

  Copy full SHA for e3d8086

• batch_size = 8

  valosekj

  committed
  7e628af

  Copy full SHA for 7e628af

• NUM_WORKERS = batch_size

  valosekj

  committed
  f832c94

  Copy full SHA for f832c94

• Add 'import torch.multiprocessing'

  valosekj

  committed
  c68d65d

  Copy full SHA for c68d65d

• Fix shape logging

  valosekj

  committed
  a7c7c77

  Copy full SHA for a7c7c77

• Change 'img_size' to 'ROI_SIZE'

  valosekj

  committed
  1c6109c

  Copy full SHA for 1c6109c

• 'batch["label"]' --> 'batch["label_lesion"]'

  valosekj

  committed
  b92e7d5

  Copy full SHA for b92e7d5

• Plot and save input and output validation images to see how the model is learning

  valosekj

  committed
  f174fb9

  Copy full SHA for f174fb9

• Fix ROI_SIZE for sliding_window_inference

  valosekj

  committed
  532e24b

  Copy full SHA for 532e24b

• Crop samples of 64x64x64 also for Validation of the fine-tuned model

  valosekj

  committed
  5c10890

  Copy full SHA for 5c10890

• update docstring

  valosekj

  committed
  d33abfa

  Copy full SHA for d33abfa

• log validation samples shapes

  valosekj

  committed
  efeccf9

  Copy full SHA for efeccf9

• Save validation figure only if it contains a lesion

  valosekj

  committed
  f34ece1

  Copy full SHA for f34ece1

Commits on Mar 19, 2024

• Plot GT together with image

  valosekj

  committed
  e0afafe

  Copy full SHA for e0afafe

• print unique values in the slice to see if it is binary

  valosekj

  committed
  e588a22

  Copy full SHA for e588a22

• update output fig fname

  valosekj

  committed
  2d0b530

  Copy full SHA for 2d0b530

• Add 'AsDiscreted' for Training and Validation of the fine-tuned model

  valosekj

  committed
  988680e

  Copy full SHA for 988680e

• threshold val_labels_list and val_outputs_list by 0.5 threshold before computing validation dice and plotting

  valosekj

  committed
  f675611

  Copy full SHA for f675611

• add normalized relu normalization

  naga-karthik

  committed
  cc78896

  Copy full SHA for cc78896

• fix binarization bug

  naga-karthik

  committed
  92334ea

  Copy full SHA for 92334ea

• remove 'logger.info(np.unique(output.detach().cpu().numpy()))'

  valosekj

  committed
  2c49ca2

  Copy full SHA for 2c49ca2

• overlay prediction over input image

  valosekj

  committed
  fbe3dd7

  Copy full SHA for fbe3dd7

Commits on Mar 20, 2024

• Fix variable when getting probabilities from logits

  valosekj

  committed
  ba39d28

  Copy full SHA for ba39d28

• Add debug lines

  valosekj

  committed
  c8d9a5d

  Copy full SHA for c8d9a5d

• PEP8

  valosekj

  committed
  ac040a5

  Copy full SHA for ac040a5

• Set validation batch_size to 1

  valosekj

  committed
  1e2b61d

  Copy full SHA for 1e2b61d

• improve comments

  valosekj

  committed
  8f49545

  Copy full SHA for 8f49545

• fix figure title

  valosekj

  committed
  3a74e28

  Copy full SHA for 3a74e28

• comment 'AsDiscreted' transforms

  valosekj

  committed
  dea121c

  Copy full SHA for dea121c

Commits on Mar 21, 2024

• Make '--pretrained-model' non required to allow training from the scratch

  valosekj

  committed
  013ca65

  Copy full SHA for 013ca65

Commits on Mar 22, 2024

• Add script to create spine-generic MSD dataset

  valosekj

  committed
  f1aae1b

  Copy full SHA for f1aae1b

• run notebook again

  valosekj

  committed
  e4c87be

  Copy full SHA for e4c87be

Commits on Mar 23, 2024

• Make 'create_msd_data.py' compatible with other BIDS datasets

  valosekj

  committed
  4448356

  Copy full SHA for 4448356

• Add note that no testing set is created

  valosekj

  committed
  a354a21

  Copy full SHA for a354a21

Commits on Mar 24, 2024

• Add README with instructions on how to download T2w images from multiple datasets

  valosekj

  committed
  141c96b

  Copy full SHA for 141c96b

• fix typo

  valosekj

  committed
  7f8e6bb

  Copy full SHA for 7f8e6bb

• remove unused imports

  valosekj

  committed
  46f1bd9

  Copy full SHA for 46f1bd9

• fix sc suffix for sci-paris

  valosekj

  committed
  afeb7c5

  Copy full SHA for afeb7c5

• Use os.path.abspath for 'args.path_data' and 'args.path_out'

  valosekj

  committed
  da1c131

  Copy full SHA for da1c131

• Update logging message

  valosekj

  committed
  ed42c68

  Copy full SHA for ed42c68

• Fix '.replace' to prevent '//' in the output string

  valosekj

  committed
  7f28fed

  Copy full SHA for 7f28fed

• Add 'create_msd_data.py' commands to create MSD-style JSON datalists

  valosekj

  committed
  99a936b

  Copy full SHA for 99a936b