scott306lr / aaml_final

Want a faster ML processor? Do it yourself! -- A framework for playing with custom opcodes to accelerate TensorFlow Lite for Microcontrollers (TFLM). . . . . . Online tutorial: https://google.github.io/CFU-Playground/ For reference docs, see the link below.

CFU Playground - AAML Final Project - Group 4

Setup Guide

1. Prepare a supported board (We use Nexys A7-100T) and install required toolchains. See Setup Guide for more details

2. Clone the CFU-Playground Repository from the github

git clone https://github.com/scott306lr/AAML_Final.git

3. Run the setup script

cd AAML_Final
./scripts/setup

4. Build Program and Start System

# Automized script for building the project.
# Equivalent to "make build && make load", while fixing multiple definition of a non-constant variable.

# Enter the project directory
cd proj/AAML_final_proj

# Building project with the default model
bash run.sh

# Building project with our custom model
bash run.sh -m "model_compression/final_0.875_qat_model.tflite"

# Verbose mode, for debugging
bash run.sh -v

After the build process is finished, press enter and type 'reboot' to reboot and start the system.

5. Run Golden Test

After the program started, type 11g to run test.

6. Run Evaluation Test

After the program started, press ctrl+c to leave litex-term, then run the following command:

python evaluation.py

Final Results

Latency: 2sec -> 0.75sec (x2.67 speed up)