Welcome to ACTIS - Augmentation Consistency Training for Instance Segmentation.
Feel free to checkout our method: ACTIS paper
To install the solution, follow these steps:
- Install conda (https://docs.conda.io/en/latest/miniconda.html)
- Create a new conda environment
conda create -n actis python pip - Activate the environment
conda activate actis - Clone the repository
git clone https://github.com/Kainmueller-Lab/ACTIS.git - Install our method
pip install -e . - use the help command to see how to use the method
actis -h
The data used in our study can be found in compressed numpy format inside this repository: data
There are three datasets:
- Flywing
- Mouse
- DSB2018
Each dataset has the following structure: ├── test │ └── test_data.npz └── train └── train_data.npz
The parameter file is a toml file that contains all the parameters needed to run an experiment.
An example parameter file can be found in src/actis/utils/resources/parameters.toml
The following parameters are possible.
| Variable | Default Value | Explanation |
|---|---|---|
| base_dir | actis | The base directory where results should live. |
| data | Flywing | The dataset to train on. Can be an absolut path to the data or either of "Flywing", "Mouse", "DSB2018" if paired with setting base_dir to github root. |
| checkpoint_path | ./experiments/myexperiment/train/checkpoints/best_model.pth | Where to store your model. |
| experiment | myexperiment | Name of your experiment. |
| num_annotated | 19 | Number of annotated labels to use. |
| batch_size_labeled | 4 | Batch size of labeled data. Total btach size will be the sum of this and the batch_size_unlabeled. Make sure they fit on your GPU. |
| batch_size_unlabeled | 8 | Batch size of unlabeled data. Total btach size will be the sum of this and the batch_size_labeled. Make sure they fit on your GPU. |
| num_workers | 4 | Number of workers to use for data loading. |
| training_steps | 50000 | Number of total training steps to do. |
| in_channels | 1 | Input channels to your network architecture. Should be chosen based on your data. |
| num_fmaps_out | 3 | Number of classes to predict. |
| learning_rate | 0.01 | Learn rate to use. |
| seed | 1 | Seed used for sampling. |
| semi_loss_weight | 1.0 | Weighting factor for the QuantileLoss loss. |
| quantile | 0.2 | Used for quantile loss scheduling. |
| quantile_end | 0.1 | Used for quantile loss scheduling. |
| quantile_warmup_steps | 5000 | Used for quantile loss scheduling. |
| load_student_weights | true | Flag indicating to load student weights |
| warmup_steps | 0 | Warmup steps before momentum is used. |
| fast_update_slow_model | false | Whether to fast update teacher model. |
| size | 128 | Affects GaussianBlur kernel size for mono color augmentations. |
| encoder_weights | imagenet | Pre-trained weights used for encoder initialization |
| no_moco | false | Do not use a slow teacher model. |
| s | 1.0 | Affects mono color augmentations. |
Create your parameter.toml file based on your specific needs.
Then call the following command to run a supervised training experiment:
actis train_super <path_to_parameter.toml> --log-out=<path_to_log_folder>
To run a semi-supervised experiment, call the following command:
actis train_semi <path_to_parameter.toml> --log-out=<path_to_log_folder>
To evaluate any trained model call the following command:
actis evaluate <path_to_parameter.toml> --log-out=<path_to_log_folder>
Please use the citation method from the CITATION.cff file! You can copy it from the "ABOUT" section of your repository.