master
Move the research models into a research subfolder (#2430)
Add TPU SavedModel exporter and refactor OD code (#6737)
Merged commit includes the following changes: 185215255 by Zhichao Lu:
update protos.
Allow setting pointwise initializer and regularizer when use_depthwise is true.
Release MobileNet V3 models and SSDLite models with MobileNet V3 backbone. (#7678)
Implement feature based consistency in DeepMAC.
Fixes three bugs in online OpenImageChallenge evaluation: 1. label offset is not propagated for image-level labels 2. adds an option to treat no image-level labels as a sign that no labels were verified for a given image 3. updates image ids in object detection evaluation on merging multiple states.
Context R-CNN Updates: Added capabilities to use multi-headed or multi-layered attention, to place the attention heads pre- or post-second stage feature extraction, and to work with embedded features in the context feature back from pre- or post-second stage feature extraction. Added an option for RPN feature map crops to be piped through to model outputs.
Add option to use tf.compat.v1.image.resize_nearest_neighbor (instead of PACK and RESHAPE) for upsampling in efficientdet bifpn.
Object detection changes: (#7208)
Enable name based definition of keras initializers in hyperparams.
Open source MnasFPN and minor fixes to OD API (#8484)
Updated the tf_example_decoder such that it supports combining multiple datasets which contain different subset of keypoints.
Implement embedding based similarity mask head for DeepMAC.
Merged commit includes the following changes: 196161788 by Zhichao Lu:
Refactor tests for Object Detection API. (#8688)
Merged commit includes the following changes: 187187978 by Zhichao Lu:
Release MobileDet code and model, and require tf_slim installation for OD API. (#8562)
Make NMS running on CPU configurable.
Improves example in documentation comment explaining keypoint_flip_permutation for RandomHorizontalFlip and RandomVerticalFlip.
Refactor object detection box predictors and fix some issues with model_main. (#4965)
Add an option to use hardswish (instead of swish) for efficientnet. This is mainly to improve latency on edge device.
Add support in label map proto for LVIS specific fields.
Standardize fine tune checkpoints across all TF2 models.