https://github.com/dyollb/segmantic
Semantic segmentation and image-to-image translation based on AI
https://github.com/dyollb/segmantic
image-processing image-segmentation machine-learning medical-image-analysis monai pytorch simpleitk
Last synced: about 1 month ago
JSON representation
Semantic segmentation and image-to-image translation based on AI
- Host: GitHub
- URL: https://github.com/dyollb/segmantic
- Owner: dyollb
- License: mit
- Created: 2021-09-13T07:12:45.000Z (over 3 years ago)
- Default Branch: main
- Last Pushed: 2024-11-07T15:13:11.000Z (7 months ago)
- Last Synced: 2025-03-27T10:21:30.554Z (about 2 months ago)
- Topics: image-processing, image-segmentation, machine-learning, medical-image-analysis, monai, pytorch, simpleitk
- Language: Python
- Homepage: https://dyollb.github.io/segmantic/
- Size: 8.11 MB
- Stars: 1
- Watchers: 1
- Forks: 3
- Open Issues: 11
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Semantic segmentation
[](https://github.com/dyollb/segmantic/actions)
[](https://github.com/dyollb/segmantic/actions)
[](https://https://opensource.org/licenses/MIT)
[](https://codecov.io/gh/dyollb/segmantic)
Semantic segmentation and image-to-image translation based on AI. This repo collects methods for pre-process data, style transfer and semantic segmentation.
## Installation
Dependencies for specific versions of CUDA are listed in `requirements/cu113.txt`, `requirements/cu111.txt`, etc.. It is advisable to install the package in a virtual environment, e.g. using `venv`
```sh
cd /your/working/directory
python -m venv .venv
```
Activate it using e.g. `source .venv/bin/activate` on Linux/Mac and `.venv\Scripts\activate.bat` on Windows.
To install this repo (this will install all dependencies):
```sh
pip install git+https://github.com/dyollb/segmantic.git#egg=segmantic
```
Or in edit/dev mode
```sh
pip install -e git+https://github.com/dyollb/segmantic.git#egg=segmantic[dev]
```
On Windows installing torch with GPU support is slightly more involved. Make sure to first install torch matching the installed CUDA version first or use the requirements files, e.g. for CUDA 11.3
```
--find-links https://download.pytorch.org/whl/cu113/torch_stable.html
torch==1.11.0+cu113
torchvision==0.12.0+cu113
```
## Example scripts
Run training:
```sh
segmantic-unet train --help
segmantic-unet train -i work/inputs/images -l work/inputs/labels -t work/inputs/labels.txt -r work/outputs
```
Or with a config file - first create empty config file (yml or json format):
```sh
segmantic-unet train-config -c config.yml --print-defaults
```
Edit `config.yml` e.g. to
```yaml
image_dir: work/inputs/images
labels_dir: work/inputs/labels
tissue_list: work/inputs/labels.txt
output_dir: work/outputs
checkpoint_file: null
num_channels: 1
spatial_dims: 3
max_epochs: 500
augment_intensity: true
augment_spatial: false
mixed_precision: true
cache_rate: 1.0
gpu_ids:
- 0
```
Now run training:
```sh
segmantic-unet train-config -c config.yml
```
## What is this tisse_list?
The example above included a tissue_list option. This is a path to a text file specifying the labels contained in a segmented image. By convention the 'label=0' is the background and is omitted from the the format. A segmentation with three tissues 'Bone'=1, 'Fat'=2, and 'Skin'=3 would be specified as follows:
```
V7
N3
C0.00 0.00 1.00 0.50 Bone
C0.00 1.00 0.00 0.50 Fat
C1.00 0.00 0.00 0.50 Skin
```
## Specifying a dataset via the 'dataset' option
Instead of providing the 'image_dir'/'labels_dir' pair, the training data can also be described by one or multiple json files. Example config that globs data from multiple json files:
```json
{
"datalist": ["/dataA/dataset.json", "/dataB/dataset.json"],
"output_dir": "",
"Etc": "etc"
}
```
The `dataset.json` loosely follows the convention used for the [Medical Segmentation Decathlon](http://medicaldecathlon.com/) datasets, and popular codes e.g. [nnUNet](https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/dataset_conversion.md).