FXPQuantCNN

KO | EN

gitlite — search

#javascript #hacktoberfest #llm #golang #deep-learning #macos #nodejs #java #awesome #cli #kubernetes #vue #ios #open-source

(★ 13)

Code and Results for Master Thesis Project on Fixed-point Quantization of Convolutional Neural Networks for Quantized Inference on Embedded Platforms

FXPQuantCNN 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Thesis Project on Convolutional Neural Network Quantization

Code used for the Master Thesis project on "Fixed-point Quantization of Convolutional Neural Networks for Quantized Inference on Embedded Platforms". Paper can be found at https://arxiv.org/abs/2102.02147

Path names are hard coded and so need to be altered before using this code. Usage of algorithms can be found in the jupyter notebooks under the analysis/ folder.

As for quantizing a CNN model, I first load an fxp_model from model_gen to quantize activations, then pass that model to functions in fxp_quantize.py to quantize weights and biases. The function evaluate() for class QuantizationEvaluator under optimized_search.py shows how this is performed.

Following is an overview of content in the respective directories

- algorithms/
    - brute_force.py - Brute force BW, F. Also found in jupyter notebooks in analysis/Brute Force Analysis/
    - fixed_bitwidth.py - Fixed 8-bit implementation
    - ind_optimized_search.py - Layer and parameter independent optimized search
    - optimized_search.py - Final working method of Optimized Search algorithm.
- analysis/ -  Main work on analyzing effects of quantization on various models for weights and activations
    - Brute Force Analysis/ - Notebooks on Brute Force Analysis for weights and activations
    - Final results (Linear Optimized Search)/ - Final tests comparing fixed bitwidth and optimized search
    - Fixed Bitwidth/ - Notebook to vary bitwidth approach
    - Layer Dependent Optimized Search/ - Parameter independent, Layer dependent optimized search
    - Layer Independent Optimized Search/ - Parameter and layer independent optimized search
    - Misc Experiments with Weights/
    - Weight Distributions/ - Weight distributions plotted for DF networks
    - Plotting.ipynb - Plotting misc results for report
- model_gen/ - scripts to generate the model structures in Keras. All classes have 2 functions. get_fxp_model() is used to load a model with a lambda layer to quantize activations. get_float_model() is used to load the original full precision model.
    - inception_cnn.py - Build the inception cnn model with pre-trained weights
    - keras_cnn.py - Two classes to build two different sequential CNN models with pre-trained weights
- model_archive/ - saved NN models
- model_training - Training of models on datasets
- outputs/ - data from experimentation. Code for storing and loading of the data can be found in the respective notebooks under analysis/
    - Brute_Force_Analysis/ - Brute force analysis results
    - Comp_Dependent_Optimized_Search - Final results after running optimized_search.py
    - Dependent_Optimized_Search - Layer-dependent but parameter independent quantization results
    - Independent_Optimized_Search - Layer and parameter independent optimized search
- test_models/ - NN models used for testing. Includes saved pre-trained CNN models and file for pre-trained weights. Also includes scripts to get the respective dataset in its appropriate form for (training and) inference.
- datasets/ - Fashion-MNIST and SVHN datasets that had to be downloaded. Scripts to get the data can be found under the respective models in test_models/

- convert_float_fixed.py - Class for quantizing an array of numbers
- evaluation_metrics.py - Functions to calculate memory consumption and cost of multiplications
- fxp_quantize.py - Using convert_float_fixed.py to fix quantization of weights and biases. Use get_fxp_model() in model_gen/ with appropriate quantization parameters to load a Keras model with quantized activations. Then pass that model to functions in fxp_quantize to quantize weights or biases. Examples shown in jupyter notebooks under analysis/
- model_data.py - functions to get properties of a model. Includes a class called Model to put together properties of a CNN model.
- requirements.txt - Python packages that were used for all the work

FXPQuantCNN

FXPQuantCNN 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

FXPQuantCNN

FXPQuantCNN 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README