A library (W.I.P) that automatically quantizes a Keras Model in Qm.n format. Configurable with Hydra.
pip install -r requirements.txt
python main.py
Given a model M with L layers and a quatization bits space B, the number of total configurations is B^L. The goal of this tool is to find for M the best configuration among the B^L possible ones.
AutoQ follows the following pipeline:
- Compute, for each layer l, the best values m and n in order to minimize the quantization loss.
- Sample a subset of the B^L configurations. This will be the training set for the regressor (see following)
- Train a regressor r that receives in input a configuration q and predicts the performances of the model M quantized following q.
- Execute an optimization algorithm that explore the configuration space (using r), in order to find the best configuration q for M that preserves the performances while minimizes the number of bits used.\
In config the default paremeters are defined:
- qmnAnalyzer defines which class is responsible for extracting the informations for each layer
- model defines the model M
- dataset defines the dataset on which M will be evaluated
- explorer defines the class that will be responsible of creating the training set for the regressor r
- regressor defines the regressor r
- optimizer defines the optimizer algorithm
- bitSpace defines the bit space B
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent