Tiny ML classifier - Glasses or No Glasses

• Create infrastructure for Git
• Quick R&D
• Free data
• Transfer
• Augmentation data
• Pytorch or TF (as tflite)
• free skeleton model for init weights (I think it will be EfficientNet-Lite* models, MobileNetV2 or ResNet50)
• Training Loop (Use API for convertin to Tiny models - for example .tflite)
• Create inference demo
• Optimised
• Bonus --->
  • Deployment for example RPi (with minimal memory - like RPi 3 with 1GB memory and without threading CPU)
  • Documentation on GitHub
  • Video demonstration realtime classification from Camera

git clone https://github.com/rvjenya/tiny-ml-classifier.git
or use ssh - [email protected]:rvjenya/tiny-ml-classifier.git
cd tiny-ml-classifier

Create env with Python3.7

python3.7 -m venv venv

source venv/bin/activate

./install.sh

• Go to (kaggle.com) and download Free dataset from - Here

• Tools for Augmentation data and exporting many formats (roboflow.com)

• From Scratch Augmentation - use Albumentations

• Open on your PC with GPU or Colab (I've attached colab version but you can export it to your GPU env)

https://colab.research.google.com/drive/1hThmbqVvYiMUD5AOORX_TXjLq2XiH-DO?usp=sharing

Edit > Notebook settings or Runtime > Change runtime type and select GPU as Hardware accelerator

• Done Models:
  • mobilenet_v2 (will be faster)
  • Inception_v3

For all models I've made FP16 version tflite - you can use it on GPU.

Final models .pb and .tflite here You can test my another tflite models with number calibration = 45 / 100 and 200 (If you want, you can try it parameters in TFLiteConverter step)

My result of training by 10 Epoch:

Testing classification:

If you need specific architecture you can use these models:

• FP32 (Full)
• FP16 (optimisation for GPU)

(venv) python cam-demo.py

python3 classify-cam-rpi.py \
  --model /model/tflite/optimise_to_3Mb/MobileNetV2/model.tflite \
  --labels /model/tflite/optimise_to_3Mb/MobileNetV2/labels.txt