Model Archive File

The YZLITE uses an archive file (.yzlite.zip) to store the relevant model information.

Overview

The model archive file is automatically created after running the train command and is updated after running the evaluate, quantize, and update_params commands.

The model archive file uses the standard Zip File Format
and its name has the format: <model name>.yzlite.zip where <model name> is the name of the YZLITE model.

The model archive file is useful as it allows for grouping the various training and evaluation files into a single, distributable file.
This file can also be directly loaded by many YZLITE commands and Python APIs, e.g.:

yzlite profile ~/my_model.yzlite.zip

Contents

The model archive file stores a given model’s:

• Model specification Python script

• Trained model files (.tflite, .h5)

• Training logs

• Evaluation logs

Directory Structure

Assume we have the following model archive file ~/workspace/my_model.yzlite.zip. The contents of this archive would have the following contents:

/my_model.py                   - The model specification script
/my_model.tflite               - The quantized model which can programmed onto an embedded device
/my_model.h5                   - The trained, non-quantized, Keras model
/my_model.h5.summary.txt       - A text summary of the .h5 model
/my_model.tflite.summary.txt   - A text summary of the .tflite model
/train/log.txt                 - Log file generated during training
/train/training-history.png    - Training history diagram
/train/training-history.json   - Training history in JSON format
/eval/h5/                      - Evaluation results from the .h5 (i.e. non-quantized) model
/eval/tflite/                  - Evaluation results from the .tflite (i.e. quantized) model