training.md

Training

Set up environment

cd Training
conda env create -f environment.yml
conda activate tinyml

Find hyper-parameters

In this section, we need to find the network structure and hyper-parameters for feature extraction module.
We use nni with Tree-structured Parzen Estimator tunner to search for our network architecture and feature extraction hyper-parameters. Run

python nni_nas.py

Then you can open the link output in the terminal to check best hyper-parameter.
We also provide one selected hyper-parameters.

Train more models with selected hyper-parameters and select the best one

Run

python train.py --param_path <hyper-parameter, store in json file>
# For example, you can run
# python train.py --param_path ./nni_params/features_extr_5v3.json

This will train 1000 models with selected hyper-parameters, create dirs and store results in:

• detection performance: train_result/model_best/{model_index}
• model weights in tflite: train_ckpt/model_best/{model_index}
• training logs: log/model_best/{model_index}

In train_ckpt/model_best, we have:

• FLOAT16.tflite model with float16 quantization in tflite format
• INT8.tflite model with quant aware training in int8 in tflite format
• NONE.tflite model without quantization in tflite format
• Original.h5 model without quantization in keras format

We also provide a simple analyse script, run

python analyse.py

This will output detection performance report and stored in train_result/model_best/model_best.csv.