This repository contains notebooks that show the usage of TensorFlow Lite for quantizing deep neural networks.
Includes lightweight MobileNetV2 backend-based and heavyweight InceptionV2 backend-based segmentation models.
Boundless paper - https://arxiv.org/pdf/1908.07007.pdf.
Quantized using the COCO-text dataset.
100 images randomly sampled from the COCO-text dataset for integer quantizing the EAST model.
Contains TFLite models generated from the MobileDet checkpoints.
The tar file contains 100 images from the train2014 split of the COCO dataset. It's useful to generate a representative dataset required for integer quantization in TFLite.
Thanks to @khanhlvg for helping out with the metadata.
This release contains TFLite models in different quantization variants for the CartoonGAN model. All the models have been populated with metadata. Thanks to @margaretmz for helping out regarding that.
Contains TF Lite variants of the EAST model proposed in An Efficient and Accurate Scene Text Detector. The original model (frozen_east_text_detection.pb) file was provided in this blog post OpenCV Text Detection (EAST text detector).
This release contains TF Lite models that are based on an InceptionV3 backbone producing higher quality images. The higher quality comes at the expense of increased latency, though. These models also support dynamic shapes as input. A brief overview of the structure of the models is available here.
The checkpoints were obtained using the code that comes from Magenta's arbitrary image stylization work.
Note: These TF Lite models are populated with required metadata that would make it super easy to import them in Android Studio. Know more about metadata generation for TF Lite models from here.