4-bit-quantization-with-tensorflow-1.15.2

Provides a custom quantization_utils.h and quantization_utils.cc file with modifications to the file from [https://github.com/tensorflow/tensorflow/releases/tag/v1.15.3] source code at location tensorflow-1.15.2/tensorflow/lite/tools/optimize/

Refer the publication at the link [https://arxiv.org/abs/1810.05723] for all the concepts behind modifications made.

1. Per-channel bit allocation\
2. Bias-Correction

Requirements

1. bazel version 0.24 or 0.25\
2. Tensorflow 1.15.x source code

Steps

1.Download both the files.
2.Replace original files in the tensorflow source code with custom ones.
3.Build tensorflow from source, directions are provided here [https://www.tensorflow.org/install/source].
4.Perform Full-integer (DEFAULT) quantization by providing representative dataset.\

Full integer quantization

Frozen.pb is the representation of the trained DNN/CNN model (Programme to create frozen.pb file from checkpoints is provided in [https://github.com/AkashB23/Quantization-of-DNNs-with-Tensorflow/blob/master/create_model.py]) path to representation dataset with atleast 100 images Output will be the .TFlite file with weights and activations in INT8 datatype but with precision of INT4. Run python IntegerQuantization.py >> file.txt to get layer and channel wise characteristics along with allocated Bit value for each channel.