ML model for fish speed and trajectories predictions

Focal fish ID-0

Environment Installation

create and activate conda environment:

conda env create -n tf-fish -f tf-environment.yml
conda activate tf-fish 

ML experiments

• prepare_data.py: create dataset with coordinates and velocity of the focal fish and the relative position and velocity of the other fishes.

  • --f: select experiment based on number of fishes (i.e. 60, 80, or 100)
  • --e: select experiment 1, 2, 3

• data_loader.py: scale and divide dataset in 1) naive temporal splitting, i.e., train-val-test, 2) train-test for assessing final performance, and 3) k folders for temporal cross-validation.

  • --tr: percentage training (default 80%)
  • --val: percentage validation (default 10%)
  • --fld: number of folders for temporal cross-validation (default 5)

• train.py: create model architecture from models.py and train it on the training-validation sets from scaled_train_val_test.pkl. Current model choices are LSTM and GRU encoder-decoder.

  • --i: input sequence (past observations)
  • --o: output sequence (future predictions)
  • --b: batch size
  • --s: set seed for reproducibility
  • --ut: number of hidd units temporal layers
  • --ufc: number of hidd units fully connected
  • --lr: learning rate
  • --dr: drop rate
  • --ep: number of epochs
  • --l1: l1 regularizer
  • --l2: l2 regularizer
  • --save: name trained model to save

• infer.py: load trained model from train.py and evaluate it on test set from scaled_train_val_test.pkl. Visualize predictions (to be scaled back) both iteratively or with matplotlib.animation.

  • --i: input sequence
  • --o: output sequence
  • --it: iteratively visualize model predictions
  • --n: file name trained model

Example

1. prepare dataset based on focal fish id-0 considering experiment 1 with 100 fishes and save it to dataset.npy

 python prepare_data.py --f 100 --e 1

2. load dataset dataset.npy, scale and split it into:

• fold_{1,5}.pkl: 5 folders datasets for future work (hyperparameter tuning and model cross-validation
• scaled_train_test.pkl: dataset for final training and testing (80-10) of model (once validated) for future work
• scaled_train_val_test.pkl: dataset for naive training-validation-testing (70-10-20) model prediction

python data_loader.py --tr 80 --val 10 --fld 5

3. load scaled_train_val_test.pkl, consider traing and validation datasets for training, create LSTM model and save trained model to model_4i_2o

python train.py --i 4 --o 2 --ep 50 --b 64 --ut 100 --ufc 100 --dr 0.01 --l1 0.0001 --lr 0.0001 --save 'model_4i_2o'

4. load trained model model_4i_2o and perform predictions on test set from scaled_train_val_test.pkl

python infer.py --i 4 --o 2 --n 'model_4i_2o'