Releases: sayakpaul/Adventures-in-TensorFlow-Lite
Metadata populated DeepLabV3-based segmentation models
Boundless (quarter) TFLite models
Boundless paper - https://arxiv.org/pdf/1908.07007.pdf.
Modified int8 variant of EAST
Quantized using the COCO-text dataset.
Sample images for COCO-text
100 images randomly sampled from the COCO-text dataset for integer quantizing the EAST model.
MobileDet TFLite models
Contains TFLite models generated from the MobileDet checkpoints.
100 COCO (train2014) images
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.
Metadata populated DeepLabV3-based segmentation models
Thanks to @khanhlvg for helping out with the metadata.
CartoonGAN TFLite models
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.
EAST model in TF Lite for text detection
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).
Arbitrary style transfer models in TF Lite with InceptionV3 backbone (with support for dynamic shapes)
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.