Commit 4c2ee0fa authored by Cristina Vasconcelos's avatar Cristina Vasconcelos Committed by A. Unique TensorFlower
Browse files

Internal change

PiperOrigin-RevId: 451273837
parent 76c365f0
# Proper Reuse of Image Classification Features Improves Object Detection
Coming soon
1. CVPR 2022 paper
2. Table of results
\ No newline at end of file
This project brings the backbone freezing training approach into the Mask-RCNN
architecture. Please see the paper for more details
\([arxiv](https://arxiv.org/abs/2204.00484) - selected for oral presentation at
CVPR 2022\).
### Training Mask-Rcnn Models with backbone frozen.
#### Freezing Resnet-RS-101 checkpoint (ImageNet pretrained).
1. Download the ResNet-RS-101 pretrained checkpoint from
[TF-Vision Model Garden](https://github.com/tensorflow/models/tree/master/official/vision#resnet-rs-models-trained-with-various-settings),
\([checkpoint](https://storage.cloud.google.com/tf_model_garden/vision/resnet-rs/resnet-rs-101-i192.tar.gz)\)
2. Config files used in our Resnet-101 ablations are included in the
[configs folder](https://github.com/tensorflow/models/tree/master/official/projects/backbone_reuse/configs/experiments/faster_rcnn).
Select one according to the target architecture (FPN, NASFPN, NASFPN +
Cascades) and training schedule preference (shorter--72 epochs, or longer
--600 epochs).
3. Change the config flag `init_checkpoint` to point to the downloaded file.
You are all set. Follow the standard TFVision Mask-Rcnn training pipeline to
complete the training.
#### How does it work?
The config files set the task's flag `freeze_backbone: true`. This flag prevents
the pretrained backbone weights from being updated during the downstream model
training.
## Citation
```
@inproceedings{vasconcelos2022backbonefreeze,
title = {Proper Reuse of Image Classification Features Improves Object Detection},
author = {Cristina Vasconcelos and Vighnesh Birodkar and Vincent Dumoulin},
booktitle={CVPR}
year={2022},
```
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment