Anytime Dense Prediction with Confidence Adaptivity

Official PyTorch implementation for the following paper:

Anytime Dense Prediction with Confidence Adaptivity. ICLR 2022.
Zhuang Liu, Zhiqiu Xu, Hung-ju Wang, Trevor Darrell and Evan Shelhamer
UC Berkeley, Adobe Research

Our implementation is based upon HRNet-Semantic-Segmentation.

---

Our full method, named Anytime Dense Prediction with Confidence (ADP-C), achieves the same level of final accuracy with HRNet-w48, and meanwhile significantly reduces total computation.

Main Results

Setting (HRNet-W48)	model	exit1	exit2	exit3	exit4	mean mIoU	exit1	exit2	exit3	exit4	mean GFLOPs
HRNet-W48	-	-	-	80.7	-	-	-	-	696.2	-
EE	model	34.3	59.0	76.9	80.4	62.7	521.6	717.9	914.2	1110.5	816.0
EE + RH	model	44.6	60.2	76.6	79.9	65.3	41.9	105.6	368.0	701.3	304.2
ADP-C: EE + RH + CA	model	44.3	60.1	76.8	81.3	65.7	41.9	93.9	259.3	387.1	195.6

Installation

Please check INSTALL.md for installation instructions.

Evaluation on pretrained models

Download our pretrained model from the table above and specify its location by TEST.MODEL_FILE

Early Exits (EE)

python tools/test_ee.py --cfg experiments/cityscapes/w48.yaml \
TEST.MODEL_FILE <PRETRAINED MODEL>.pth

This should give

34.33	59.01	76.90	80.43	62.67

Redesigned Heads (RH)

python tools/test_ee.py --cfg experiments/cityscapes/w48.yaml \
EXIT.TYPE 'flex' EXIT.INTER_CHANNEL 128 \
TEST.MODEL_FILE <PRETRAINED MODEL>.pth

This should give

44.61	60.19	76.64	79.89	65.33

ADP-C (EE + RH + CA)

python tools/test_ee.py \
--cfg experiments/cityscapes/w48.yaml MODEL.NAME model_anytime  \
EXIT.TYPE 'flex' EXIT.INTER_CHANNEL 128 \
MASK.USE True MASK.CONF_THRE 0.998 \
TEST.MODEL_FILE <PRETRAINED MODEL>.pth

This should give

44.34	60.13	76.82	81.31	65.65

ADP-C (EE + RH + CA) (w18) Pretrained w18 with ADP-C

python tools/test_ee.py \
--cfg experiments/cityscapes/w18.yaml MODEL.NAME model_anytime  \
EXIT.TYPE 'flex' EXIT.INTER_CHANNEL 64 \
MASK.USE True MASK.CONF_THRE 0.998 \
TEST.MODEL_FILE <PRETRAINED MODEL>.pth

This should give

40.83	48.19	68.26	77.02	58.57

Train

There are two configurations for the backbone HRnet model, w48.yaml and w18.yaml under experimens/cityscapes. Note that the following commands are for using HRNet-w48 as backbone. Please change EXIT.INTER_CHANNEL to 64 when using w18 as backbone.

Early Exits (EE)

python -m torch.distributed.launch tools/train_ee.py \
--cfg experiments/cityscapes/w48.yaml

Redesigned Heads (RH)

python -m torch.distributed.launch tools/train_ee.py \
--cfg experiments/cityscapes/w48.yaml \
EXIT.TYPE 'flex' EXIT.INTER_CHANNEL 128

Confidence Adatative (CA)

python -m torch.distributed.launch tools/train_ee.py \
--cfg experiments/cityscapes/w48.yaml \
MASK.USE True MASK.CONF_THRE 0.998

ADP-C (EE + RH + CA)

python -m torch.distributed.launch tools/train_ee.py \
--cfg experiments/cityscapes/w48.yaml \
EXIT.TYPE 'flex' EXIT.INTER_CHANNEL 128 \
MASK.USE True MASK.CONF_THRE 0.998

Evaulation results will be generated at the end of training.

• result.txt: contains mIOU for each exit and the average mIOU of the four exits.

• test_stats.json: contains FLOPs and number of parameters.

• final_state.pth: the trained model file.

• config.yaml: the configuration file.

Test

Evaluation

python tools/test_ee.py --cfg <Your output directoy>/config.yaml

Acknowledgement

This repository is built upon HRNet-Semantic-Segmentation.

License

This project is released under the MIT license. Please see the LICENSE file for more information.

Citation

If you find this repository helpful, please consider citing:

@Article{liu2022anytime,
  author  = {Zhuang Liu and Zhiqiu Xu and Hung-Ju Wang and Trevor Darrell and Evan Shelhamer},
  title   = {Anytime Dense Prediction with Confidence Adaptivity},
  journal = {International Conference on Learning Representations (ICLR)},
  year    = {2022},
}