The aim of this study is explain the different types of quantization methods of TensorFlow Lite and make a comparison with Neuton.AI platform. I would like to thank Neuton.AI for their help during the project.
The source of data is here: occupancy_detection_dataset
The problem is: Recognize whether someone is in the room or not based on measurements of temperature, humidity, light, and CO2. Ground-truth occupancy was obtained from time stamped pictures that were taken every minute.
Features and Target
- Temperature Humidity
- Light
- CO2
- Humidity Ratio
- Occupancy – 0 - not occupied; 1 - occupied status.
There are two colab notebooks:
- Occupancy_TFLite_Models.ipynb: includes all types of quantization methods of TensorFlow Lite.
- Occupancy_TFLite_vs_NeutonAI.ipynb: includes the comparison with Neuton.AI
You can find the results below:
| Model | Parameters | Accuracy | Size_kb |
|---|---|---|---|
| TensorFlow Model | 57.0 | 0.974729 | 25.750000 |
| TensorFlow Model_2 | 3553.0 | 0.990975 | 79.601562 |
| TFLite_no_quantization | - | 0.974729 | 1.632812 |
| Dynamic_range_quantization | - | 0.974729 | 1.734375 |
| Representative_Dataset_Float_Fallback | - | 0.974729 | 2.085938 |
| Neuton_AI | 60.0 | 0.981279 | 0.140000 |