|
--- |
|
license: apache-2.0 |
|
tags: |
|
- ENOT-AutoDL |
|
- SpeechEnhancement |
|
--- |
|
|
|
# MP-SENet optimization on VoiceBank+DEMAND dataset with ENOT-AutoDL. |
|
|
|
This repository contains the optimized version of [MP-SENet](https://github.com/yxlu-0102/MP-SENet) model. |
|
Number of multiplication and addition operations (MACs) was used for computational complexity measurement. PESQ score was used as a quality metric. |
|
|
|
## Optimization results |
|
|
|
We use MACs as a latency measure because this metric is device-agnostic and implementation independent. |
|
There is also a possibility to optimize a model by target device latency using ENOT neural architecture selection algorithm. |
|
Please, keep in mind that acceleration by device latency differs from acceleration by MACs. |
|
|
|
| **Model** | **MACs** | **Acceleration (MACs)** | PESQ score (the higher the better) | |
|
|----------------|:--------:|:-----------------------:|:----------------------------------:| |
|
| baseline | 302.39 B | 1.0 | 3.381 | |
|
| ENOT optimized | 120.95 B | 2.5 | 3.386 | |
|
|
|
You can use `Baseline_model.pth` or `ENOT_optimized_model.pth` in the original repo by loading a model as generator in the following way: |
|
```python |
|
generator = torch.load("ENOT_optimized_model.pth") |
|
``` |
|
|
|
Each of these two files contain a model object, saved by `torch.save`, so you can load them only from the original repository root because of imports. |
|
|
|
If you want to book a demo, please contact us: enot@enot.ai . |
