Multiplex image model

Installation

Simply install the packages in provided requirements file:

pip install -r requirements.txt

Data preparation

To use the model, you need to have your data in the following directories structure:

.
└── data
    ├── test
    │   ├── dataset1
    │   │   └── imgs
    │   ├── dataset2
    │   │   └── imgs
    │   └── dataset3
    │       └── imgs
    └── train
        ├── dataset1
        │   └── imgs
        ├── dataset2
        │   └── imgs
        └── dataset3
            └── imgs

Where train and test are splits and dataset... is a dataset/panel with images. Provide full paths to your splits directories in congifs/all_panels_config.yaml under paths, as in provided example config. Then list all the datasets subdirectories you want to use from your split paths under datasets config key. Ultimately, for each dataset used, provide a list of markers that are represented in consequtive image channels for that dataset, in the markers config section.

If you are working on rudy, use the provided split_data.ipynb notebook to do the dataset split (only the section Images preparation).

Training a model

Training utilizes the Neptune logging. You can make free account with academic license (https://neptune.ai/research), then insert your project and api_token into train_masked_config.yaml under Neptune configuration.

To train a model, simply execute training script with its config file:

python3 train_masked_model.py train_masked_config.yaml

The model checkpoint will be saved in the directory specified in training config (by default ./checkpoints).