Skip to content

Latest commit

 

History

History
67 lines (46 loc) · 2.25 KB

README.md

File metadata and controls

67 lines (46 loc) · 2.25 KB
Raw

Models

This directory contains the pre-trained model weights and metadata. See instructions below about their origins.

Some of the tools use Python. Install their dependencies:

python3 -m venv venv
. venv/bin/activate

pip install -r llama2.c/requirements.txt
pip install -r yolov5/requirements.txt

Llama

The model used by llama2.c is based on Llama 2 and the TinyStories dataset. The model weights are from the tinyllamas repository. We are using the OG version (with 15M parameters) and quantize it. This model runs on the Arm Cortex-M7 CPU.

LLAMA_MODEL_NAME=stories15M
LLAMA_MODEL_DIR=llama2

wget -P models/${LLAMA_MODEL_DIR} \
    https://huggingface.co/karpathy/tinyllamas/resolve/main/${LLAMA_MODEL_NAME}.pt

python llama2.c/export.py \
    models/${LLAMA_MODEL_DIR}/${LLAMA_MODEL_NAME}_q80.bin \
    --version 2 \
    --checkpoint models/${LLAMA_MODEL_DIR}/${LLAMA_MODEL_NAME}.pt

The tokenizer comes from the llama2.c repository.

cp llama2.c/tokenizer.bin models/${LLAMA_MODEL_DIR}/

Vision

Object detection (with labels used for prompting Llama) is based on YOLOv5, specifially the smallest version "n" at a 224x224 resolution. This model runs on the Coral Edge TPU, which requires an additional compilation step (handled by the exporter).

git clone https://github.com/ultralytics/yolov5.git

YOLO_RESOLUTION=224
YOLO_VERSION=n
VISION_MODEL_NAME=yolov5${YOLO_VERSION}-int8_edgetpu
YOLO_MODEL_DIR=yolov5

python yolov5/export.py \
    --weights yolov5${YOLO_VERSION}.pt \
    --include edgetpu \
    --int8 \
    --img ${YOLO_RESOLUTION} \
    --data yolov5/data/coco128.yaml

mkdir models/${YOLO_MODEL_DIR}/
cp yolov5/${VISION_MODEL_NAME}.tflite models/${YOLO_MODEL_DIR}/

The labels are from the COCO dataset. Convert them to an easily readable format.

python models/export_coco_labels.py