Quantization-of-DNNs-with-Tensorflow

Includes all necessary files to arrive at a TFlite starting with checkpoints

Requirements\

1.Tensorflow 1.15.2 installed 2.python > 2.7

Refer the link https://github.com/tensorflow/models/tree/master/research/slim to know all the necessary file citations mentioned here.

1.Need to clone the repository with above link.
2.Copy all the Files provided to the folder with slim clone.

Training any of the CNN models provided in the Tensorflow slim repository would provide you with Checkpoints that are source to use files in this repository.

1.create_model.py

Provide the path to checkpoints and other input details as mentioned in the comments provided in the code to get Frozen.pb protocolbuffer file of the trained CNN model.

2.WQuantization.py

Specify the input and output node along with path to previously generated Frozen.pb file, the output.tflite file is the Flatbuffer file which is 1/4th size of the Floating point model and can be used in Android apps and other embedded platforms supported by tensorflow (here quantization refers only to weight quantization).

3.visualize.py

Inorder to Graphically visualize the trained CNN/DNN architecture you can use this file and get a logs folder created as output later, with the tensorboard preinstalled in the device. tensorboard --logdir=#(path to logs folder) will land you in a the tensoeflow graph.