Once for All on ECG Dataset

Also refer to README from here for more informations in the repo

Prepare the dataset

Using the ecg-tinyml repo, generate the three .csv files, or use the prepared mitbih.zip and extract it under $HOME/dataset/mitbih as below

├── dataset
    ├──mitbih
        ├── mitbih_test.csv
        ├── mitbih_train.csv
        ├── mitbih_val.csv

If you want to use the dataset placed elsewhere, please change the directory specified at ofa/imagenet_classification/data_providers/mitbih.py

Prepare the environment

Use the same environment as from ecg-tinyml repo

To train OFA Networks

conda activate tinyml-ecg

# The following commands can be run separately as it takes significant time for each stage 
python train_ofa_net.py --task kernel
python train_ofa_net.py --task depth --phase 1
python train_ofa_net.py --task depth --phase 2
python train_ofa_net.py --task expand --phase 1
python train_ofa_net.py --task expand --phase 2

Sample the subnetworks

After the last stage training, there will be a .tar file under exp/kernel_depth2kernel_depth_width/phase2/checkpoint/checkpoint.pth.tar. Use this file to perform sampling the subnetworks.

To get random networks and benchmark its performance on validation set, run python sample_randomly.py script

After picking the best networks from the accuracy-complexity tradeoff graph, update the config file at configs/sample_subnetwork.yaml and run python sample_with_config.py to sample a particular subnetworks. The script will export .pth and .onnx models. After that, use the scripts in ecg-tinyml repo to further convert it to TensorFlow format