CNN-LSTM

Implementation of the CNN-LSTM model for classification of plasma confinement states.

NF publication 2020: https://iopscience.iop.org/article/10.1088/1741-4326/ab6c7a

The model architecture presented in the following code is not the same as the one shown in the publication. It was highly pruned, reducing #params by a factor of 36, which led to an improvement of +1%,+8%,+3% for L,D and H modes respectively.The current architecture is the one used in the real-time PCS of TCV.

Installation

# Installation of Miniconda from scratch

• Get and install Miniconda:
  1. cd your_project/ (Miniconda packages might require a significant space ~Gbs)
  2. wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
  3. bash Miniconda3-latest-Linux-x86_64.sh
  4. export PATH="/home/user/your_project/miniconda3/bin:$PATH" (or where you have decided to install miniconda3)

# Create an environment

• An environment file is provided for version compatibility.
  conda create --name my_gpu_env --file environment_gpu.txt
pip install tensorflow-gpu==2.3
  Please use this tf version since older ones, as tf 2.2, had some time consuming warnings when doing model predictions
  as explained here tensorflow/tensorflow#41347

Preparation of Experiments

Obtain dataset

Dataset

• https://www.dropbox.com/sh/iw5sj788wdlz40y/AABGIa8lp4cMMrFj8yY0VeaDa?dl=0

• The most up-to-date validated files are in apau_and_marceca/ folder

• Filenames must have the following format: "Machine_ShotNumber_Labeler.csv", e.g: "TCV_26386_apau_and_marceca_labeled.csv"

Visualize data inputs from generator

python lstm_data_generator.py

Run an experiment

(python lstm_train.py #experiment #machine #labeler)
python lstm_train.py 2 TCV apau_and_marceca

Compute predictions

(python lstm_predict.py #experiment #epoch #test_machine)
python lstm_predict.py 2 400 JET

Evaluate Kappa scores

(python lstm_scores.py #experiment #epoch #test_machine)
python lstm_scores.py 2 400 JET

Predict for a given shot and save in DIS_tool format

python algorithms/ConvLSTM/evaluate_model_from_detected_signals.py #experiment #epoch #shot #target_machine #normalization_method #git_test
python algorithms/ConvLSTM/evaluate_model_from_detected_signals.py baseline_16042021_exp9 400 32195 TCV minmax False