Refactor code and add new functionalities

.github/workflows/pre-commit.yml

@@ -13,9 +13,9 @@ jobs:
         python-version: ["3.9"]
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
           check-latest: true

.github/workflows/unit-test.yml

@@ -16,9 +16,9 @@ jobs:
 
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           check-latest: true
           python-version: ${{ matrix.python-version }}
@@ -28,9 +28,9 @@ jobs:
             docker/environment_mac_m1.yml
             docker/Dockerfile
             docker/requirements.txt
+            pyproject.toml
       - name: Install dependencies
         run: |
-          python -m pip install --upgrade pip
           pip install tensorflow-cpu==2.12.0
           pip install jax==0.4.20
           pip install jaxlib==0.4.20

.pre-commit-config.yaml

@@ -24,17 +24,17 @@ repos:
         files: "docker/requirements.txt"
       - id: trailing-whitespace
   - repo: https://github.com/PyCQA/isort
-    rev: 5.12.0
+    rev: 5.13.2
     hooks:
       - id: isort
   - repo: https://github.com/psf/black
-    rev: 23.11.0
+    rev: 23.12.1
     hooks:
       - id: black
         args:
           - --line-length=100
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.7.1
+    rev: v1.8.0
     hooks: # https://github.com/python/mypy/issues/4008#issuecomment-582458665
       - id: mypy
         name: mypy
@@ -54,15 +54,15 @@ repos:
             --warn-unreachable,
           ]
   - repo: https://github.com/pre-commit/mirrors-prettier
-    rev: v3.1.0
+    rev: v4.0.0-alpha.8
     hooks:
       - id: prettier
         args:
           - --print-width=100
           - --prose-wrap=always
           - --tab-width=2
   - repo: https://github.com/charliermarsh/ruff-pre-commit
-    rev: "v0.1.6"
+    rev: "v0.1.9"
     hooks:
       - id: ruff
   - repo: https://github.com/pre-commit/mirrors-pylint

Makefile

@@ -3,10 +3,9 @@ pip:
 
 test:
 	pytest --cov=imgx -n 4 imgx
-	pytest --cov=imgx_datasets -n 4 imgx_datasets
 
 build_dataset:
-	tfds build imgx_datasets/male_pelvic_mr
-	tfds build imgx_datasets/amos_ct
-	tfds build imgx_datasets/muscle_us
-	tfds build imgx_datasets/brats2021_mr
+	tfds build imgx/datasets/male_pelvic_mr
+	tfds build imgx/datasets/amos_ct
+	tfds build imgx/datasets/muscle_us
+	tfds build imgx/datasets/brats2021_mr

README.md

@@ -8,21 +8,8 @@ This repository includes the implementation of the following work
 - [Importance of Aligning Training Strategy with Evaluation for Diffusion Models in 3D Multiclass Segmentation](https://arxiv.org/abs/2303.06040)
 
 :construction: **The codebase is still under active development for more enhancements and
-applications.** :construction:
-
-- November 2023:
-  - :warning: Upgraded to JAX to 0.4.20.
-  - :warning: Removed Haiku-specific modification to convolutional layers. This may impact model
-    performance.
-  - :smiley: Added example notebooks for inference on a single image without TFDS.
-  - Added integration tests for training, validation, and testing.
-  - Refactored config.
-    - Added `patch_size` and `scale_factor` to data config.
-    - Moved loss config from the main config to task config.
-  - Refactored code, including defining `imgx/task` submodule.
-- October 2023:
-  - :blush: Migrated from [Haiku](https://github.com/google-deepmind/dm-haiku) to
-    [Flax](https://github.com/google/flax) following Google DeepMind's recommendation.
+applications. Please check [release notes](https://github.com/mathpluscode/ImgX-DiffSeg/releases)
+for more information. ** :construction:
 
 :mailbox: Please feel free to
 [create an issue](https://github.com/mathpluscode/ImgX-DiffSeg/issues/new/choose) to request
@@ -38,7 +25,7 @@ Current supported functionalities are summarized as follows.
 
 **Data sets**
 
-See the [readme](imgx_datasets/README.md) for further details.
+See the [readme](imgx/datasets/README.md) for further details.
 
 - Muscle ultrasound from [Marzola et al. 2021](https://data.mendeley.com/datasets/3jykz7wz8d/1).
 - Male pelvic MR from [Li et al. 2022](https://zenodo.org/record/7013610#.Y1U95-zMKrM).
@@ -68,6 +55,10 @@ See the [readme](imgx_datasets/README.md) for further details.
 **Training**
 
 - Patch-based training.
+- Data augmentation with anisotropic support, including
+  - Random affine: rotation, scaling, shearing, shifting.
+  - Random gamma adjustment.
+  - Random flip.
 - Multi-device training (one model per device) with
   [`pmap`](https://jax.readthedocs.io/en/latest/_autosummary/jax.pmap.html).
 - Mixed precision training.
@@ -284,7 +275,6 @@ Run the command below to test and get a coverage report. As JAX tests require tw
 
 ```bash
 pytest --cov=imgx -n 4 imgx -k "not integration"
-pytest --cov=imgx_datasets -n 4 imgx_datasets
 ```
 
 `-k "not integration"` excludes integration tests, which require downloading muscle ultrasound and

docker/requirements.txt

@@ -1,15 +1,16 @@
 SimpleITK==2.3.1
 chex==0.1.8
-coverage==7.3.2
+coverage==7.3.3
 flax==0.7.5
 hydra-core==1.3.2
 kaggle==1.5.16
+matplotlib==3.8.2
 nbmake==1.4.6
 numpy==1.26.2
 opencv-python==4.8.1.78
 optax==0.1.7
-pandas==2.1.3
-pre-commit==3.5.0
+pandas==2.1.4
+pre-commit==3.6.0
 protobuf==3.20.3 # https://github.com/tensorflow/datasets/issues/4858
 pytest-cov==4.1.0
 pytest-mock==3.12.0
@@ -19,8 +20,9 @@ pytest-xdist==3.5.0
 pytest==7.4.3
 rdkit-pypi==2022.9.5
 rich==13.7.0
-ruff==0.1.6
+ruff==0.1.8
 tensorflow-datasets==4.9.3
-torch==2.1.1 # for testing only
-wandb==0.16.0
+tomli==2.0.1
+torch==2.1.2 # for testing only
+wandb==0.16.1
 wily==1.25.0

examples/segmentation/config.yaml

@@ -9,14 +9,14 @@ data:
       - 0
       - 0
     data_augmentation:
-      max_rotation:
-        - 0.088
-      max_translation:
-        - 10
-        - 10
-      max_scaling:
-        - 0.15
-        - 0.15
+      max_rotation: 30
+      max_zoom: 0.2
+      max_shear: 30
+      max_shift: 0.3
+      max_log_gamma: 0.3
+      v_min: 0.0
+      v_max: 1.0
+      p: 0.5
   trainer:
     max_num_samples: 512000
     batch_size: 64
@@ -40,6 +40,7 @@ task:
     scale_factor:
       - 2
       - 2
+    num_res_blocks: 2
     num_channels:
       - 8
       - 16
@@ -49,10 +50,11 @@ task:
     num_heads: 8
     widening_factor: 4
     num_transform_layers: 1
+    dropout: 0.1
   loss:
     dice: 1.0
     cross_entropy: 0.0
-    focal: 20.0
+    focal: 1.0
   early_stopping:
     metric: mean_binary_dice_score_without_background
     mode: max

imgx/conf/data/amos_ct.yaml

@@ -5,9 +5,14 @@ loader:
   patch_shape: [128, 128, 128]
   patch_overlap: [64, 0, 0] # image shape is [192, 128, 128]
   data_augmentation:
-    max_rotation: [0.088, 0.088, 0.088] # roughly 5 degrees
-    max_translation: [10, 10, 10]
-    max_scaling: [0.15, 0.15, 0.15]
+    max_rotation: 30 # degrees
+    max_zoom: 0.2 # as a fraction of the image size
+    max_shear: 30 # degrees
+    max_shift: 0.3 # as a fraction of the image size
+    max_log_gamma: 0.3
+    v_min: 0.0 # minimum value for intensity
+    v_max: 1.0 # maximum value for intensity
+    p: 0.5 # probability of applying each augmentation
 
 trainer:
   max_num_samples: 100_000

imgx/conf/data/brats2021_mr.yaml

@@ -5,9 +5,14 @@ loader:
   patch_shape: [128, 128, 128]
   patch_overlap: [0, 0, 32] # image shape is [179, 219, 155]
   data_augmentation:
-    max_rotation: [0.088, 0.088, 0.088] # roughly 5 degrees
-    max_translation: [20, 20, 4]
-    max_scaling: [0.15, 0.15, 0.15]
+    max_rotation: 30 # degrees
+    max_zoom: 0.2 # as a fraction of the image size
+    max_shear: 30 # degrees
+    max_shift: 0.3 # as a fraction of the image size
+    max_log_gamma: 0.3
+    v_min: 0.0 # minimum value for intensity
+    v_max: 1.0 # maximum value for intensity
+    p: 0.5 # probability of applying each augmentation
 
 trainer:
   max_num_samples: 100_000

imgx/conf/data/male_pelvic_mr.yaml

@@ -5,15 +5,20 @@ loader:
   patch_shape: [256, 256, 32]
   patch_overlap: [0, 0, 16] # image shape is [256, 256, 48]
   data_augmentation:
-    max_rotation: [0.088, 0.088, 0.088] # roughly 5 degrees
-    max_translation: [20, 20, 4]
-    max_scaling: [0.15, 0.15, 0.15]
+    max_rotation: 30 # degrees
+    max_zoom: 0.2 # as a fraction of the image size
+    max_shear: 30 # degrees
+    max_shift: 0.3 # as a fraction of the image size
+    max_log_gamma: 0.3
+    v_min: 0.0 # minimum value for intensity
+    v_max: 1.0 # maximum value for intensity
+    p: 0.5 # probability of applying each augmentation
 
 trainer:
   max_num_samples: 100_000
   batch_size: 8 # all model replicas are updated every `batch_size` samples
   batch_size_per_replica: 1 # each model replicate takes `batch_size_per_replica` samples per step
   num_devices_per_replica: 1 # model is split into num_devices_per_replica shards/slices
 
-patch_size: [2, 2, 2]
+patch_size: [2, 2, 1] # do not downsample z axis
 scale_factor: [2, 2, 2]

imgx/conf/data/muscle_us.yaml

@@ -5,9 +5,14 @@ loader:
   patch_shape: [480, 512]
   patch_overlap: [0, 0]
   data_augmentation:
-    max_rotation: [0.088] # roughly 5 degrees
-    max_translation: [10, 10]
-    max_scaling: [0.15, 0.15]
+    max_rotation: 30 # degrees
+    max_zoom: 0.2 # as a fraction of the image size
+    max_shear: 30 # degrees
+    max_shift: 0.3 # as a fraction of the image size
+    max_log_gamma: 0.3
+    v_min: 0.0 # minimum value for intensity
+    v_max: 1.0 # maximum value for intensity
+    p: 0.5 # probability of applying each augmentation
 
 trainer:
   max_num_samples: 512_000

imgx/conf/task/gaussian_diff_seg.yaml

@@ -33,20 +33,22 @@ model:
   num_spatial_dims: 3
   patch_size: MISSING # data dependent, will be set after loading config
   scale_factor: MISSING # data dependent, will be set after loading config
+  num_res_blocks: 2
   num_channels: [32, 64, 128, 256]
   out_channels: MISSING # data dependent, will be set after loading config
   num_heads: 8
   widening_factor: 4
+  dropout: 0.1
 
 loss:
   dice: 1.0
   cross_entropy: 0.0
-  focal: 20.0
+  focal: 1.0
   mse: 0.0
   vlb: 0.0
 
 early_stopping: # used on validation set
   metric: "mean_binary_dice_score_without_background"
   mode: "max"
   min_delta: 0.0001
-  patience: 5
+  patience: 10

imgx/conf/task/seg.yaml

@@ -6,19 +6,21 @@ model:
   num_spatial_dims: 3
   patch_size: MISSING # data dependent, will be set after loading config
   scale_factor: MISSING # data dependent, will be set after loading config
+  num_res_blocks: 2
   num_channels: [32, 64, 128, 256]
   out_channels: MISSING # data dependent, will be set after loading config
   num_heads: 8
   widening_factor: 4
   num_transform_layers: 1
+  dropout: 0.1
 
 loss:
   dice: 1.0
   cross_entropy: 0.0
-  focal: 20.0
+  focal: 1.0
 
 early_stopping: # used on validation set
   metric: "mean_binary_dice_score_without_background"
   mode: "max"
   min_delta: 0.0001
-  patience: 5
+  patience: 10

imgx/data/__init__.py

@@ -1,9 +1 @@
 """Module to handle data."""
-from __future__ import annotations
-
-from typing import Callable
-
-import jax
-import jax.numpy as jnp
-
-AugmentationFn = Callable[[jax.Array, dict[str, jnp.ndarray]], dict[str, jnp.ndarray]]

imgx/data/augmentation.py → imgx/data/augmentation/__init__.py

@@ -1,12 +1,13 @@
-"""Image augmentation functions."""
+"""Data augmentation module."""
 from __future__ import annotations
 
 from collections.abc import Sequence
+from typing import Callable
 
 import jax
-import jax.numpy as jnp
+from jax import numpy as jnp
 
-from imgx.data import AugmentationFn
+AugmentationFn = Callable[[jax.Array, dict[str, jnp.ndarray]], dict[str, jnp.ndarray]]
 
 
 def chain_aug_fns(