README.md

Procedure to Execute Experiments

Note

Only the procedures to train, batch-wise sample, and test the embedding network are included below. The remainder of the execution procedure is yet to be documented, but follows a similar workflow and can be retraced through the modularly structured pluggable portions and the commented lines in run.py.

Clone the repository and set up the Python environment

• This implementation used anaconda for managing libraries, and the environment can be reproduced using the dependency file: dependencies.yml.
• For instance, to clone the dependencies into a new environment, use,

  conda env create -f dependencies.yml
  

• Alternatively, use pip manager to install the dependencies from the dependencies.txt](./dependencies.txt) file.

  pip install -r dependencies.txt
  

Further References

• A detailed guide for managing anaconda environbments can be found at anaconda's official page.
• A helpful reference for managing pip packages can be found here.

Train the Embedding Network

Keeping all parameters fixed, and using the default architecture for the underlying network, the following steps will train the embedding network.

Note

• By default, the model states after each epoch will be saved into Experiments/logs/prototypical.
• To train on a different dataset, the dataset interfacing classes need to be defined according to the data files structure. See the class for ISIC 2018 data, for instance.

• Activate the conda environment set up in the previous stage.

  conda activate skin_fsl
  

• Navigate to Skin-FSL/prototypical/Experiments/src.

• Edit run.py and uncomment the line to call the trainer. The line is a function call like so: trainer.train().

  • To edit the file on a CLI interface, nano editor is a convenient option.

    nano run.py    (to open the file in nano editor)
    ctrl+o         (to save changes)
    ctrl+x         (to close the editor)
    

• Execute run.py (not context-sensitive).

  python run.py
  

• The following is a sample output when running the training experiment.

Test the Embedding Network

Note

• The model state for evaluation is set by default. To change this, change the value assigned to variable model_path in src/prototypical/tester_exhaustive.py.
• To test on a different dataset, the dataset interfacing classes need to be defined according to the data files structure. See the class for ISIC 2018 data, for instance.

• Activate the conda environment set up in the previous stage.

  conda activate skin_fsl
  

• Navigate to Skin-FSL/prototypical/Experiments/src.

• Edit run.py and uncomment the line to call the tester. The line is a function call like so: tester_exhaustive.test().

  • To edit the file on a CLI interface, nano editor is a convenient option.

    nano run.py    (to open the file in nano editor)
    ctrl+o         (to save changes)
    ctrl+x         (to close the editor)
    

• Execute run.py (not context-sensitive).

  python run.py
  

• The following is a sample output when running a test on the trained network.