Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/googleinterns/wss

A simple consistency training framework for semi-supervised image semantic segmentation
https://github.com/googleinterns/wss

segmentation semi-supervised-learning

Last synced: about 16 hours ago
JSON representation

A simple consistency training framework for semi-supervised image semantic segmentation

Host: GitHub
URL: https://github.com/googleinterns/wss
Owner: googleinterns
License: apache-2.0
Created: 2020-08-14T21:28:54.000Z (over 4 years ago)
Default Branch: master
Last Pushed: 2024-07-30T20:50:14.000Z (5 months ago)
Last Synced: 2024-12-14T13:15:26.335Z (8 days ago)
Topics: segmentation, semi-supervised-learning
Language: Python
Homepage:
Size: 4.01 MB
Stars: 167
Watchers: 11
Forks: 24
Open Issues: 11
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE

Awesome Lists containing this project

Awesome-Semi-Supervised-Semantic-Segmentation - Code

README

        # [PseudoSeg: Designing Pseudo Labels for Semantic Segmentation](https://arxiv.org/pdf/2010.09713v1.pdf)

PseudoSeg is a simple consistency training framework for semi-supervised image

semantic segmentation, which has a simple and novel re-design of pseudo-labeling

to generate well-calibrated structured pseudo labels for training with unlabeled

or weakly-labeled data. It is implemented by [Yuliang Zou](https://yuliang.vision/) (research intern) in 2020 Summer.

__This is not an official Google product.__

## Instruction

### Installation

- Use a virtual environment

```bash

virtualenv -p python3 --system-site-packages env

source env/bin/activate

```

- Install packages

```bash

pip install -r requirements.txt

```

### Dataset

Create a `dataset` folder under the ROOT directory, then download the pre-created tfrecords for [voc12](https://filebox.ece.vt.edu/~ylzou/summer2020pseudoseg/pascal_voc_seg.tar) and [coco](https://filebox.ece.vt.edu/~ylzou/summer2020pseudoseg/coco.tar), and extract them in `dataset` folder. You may also want to check the filenames for each split under `data_splits` folder.

### Training

**NOTE:** 

- We train all our models using 16 V100 GPUs.

- The ImageNet pre-trained models can be download [here](https://github.com/tensorflow/models/blob/master/research/deeplab/g3doc/model_zoo.md#model-details-3).

- For VOC12, `${SPLIT}` can be `2_clean, 4_clean, 8_clean, 16_clean_3` (representing 1/2, 1/4, 1/8, and 1/16 splits), `NUM_ITERATIONS` should be set to 30000.

- For COCO, `${SPLIT}` can be `32_all, 64_all, 128_all, 256_all, 512_all` (representing 1/32, 1/64, 1/128, 1/256, and 1/512 splits), `NUM_ITERATIONS` should be set to 200000.

#### Supervised baseline

```bash

python train_sup.py \

  --logtostderr \

  --train_split="${SPLIT}" \

  --model_variant="xception_65" \

  --atrous_rates=6 \

  --atrous_rates=12 \

  --atrous_rates=18 \

  --output_stride=16 \

  --decoder_output_stride=4 \

  --train_crop_size="513,513" \

  --num_clones=16 \

  --train_batch_size=64 \

  --training_number_of_steps="${NUM_ITERATIONS}" \

  --fine_tune_batch_norm=true \

  --tf_initial_checkpoint="${INIT_FOLDER}/xception_65/model.ckpt" \

  --train_logdir="${TRAIN_LOGDIR}" \

  --dataset_dir="${DATASET}"

```

#### PseudoSeg (w/ unlabeled data)

```bash

python train_wss.py \

  --logtostderr \

  --train_split="${SPLIT}" \

  --train_split_cls="train_aug" \

  --model_variant="xception_65" \

  --atrous_rates=6 \

  --atrous_rates=12 \

  --atrous_rates=18 \

  --output_stride=16 \

  --decoder_output_stride=4 \

  --train_crop_size="513,513" \

  --num_clones=16 \

  --train_batch_size=64 \

  --training_number_of_steps="${NUM_ITERATIONS}" \

  --fine_tune_batch_norm=true \

  --tf_initial_checkpoint="${INIT_FOLDER}/xception_65/model.ckpt" \

  --train_logdir="${TRAIN_LOGDIR}" \

  --dataset_dir="${DATASET}"

```

#### PseudoSeg (w/ image-level labeled data)

```bash

python train_wss.py \

  --logtostderr \

  --train_split="${SPLIT}" \

  --train_split_cls="train_aug" \

  --model_variant="xception_65" \

  --atrous_rates=6 \

  --atrous_rates=12 \

  --atrous_rates=18 \

  --output_stride=16 \

  --decoder_output_stride=4 \

  --train_crop_size="513,513" \

  --num_clones=16 \

  --train_batch_size=64 \

  --training_number_of_steps="${NUM_ITERATIONS}" \

  --fine_tune_batch_norm=true \

  --tf_initial_checkpoint="${INIT_FOLDER}/xception_65/model.ckpt" \

  --train_logdir="${TRAIN_LOGDIR}" \

  --dataset_dir="${DATASET}" \

  --weakly=true

```

### Evaluation

**NOTE:** `${EVAL_CROP_SIZE}` should be `513,513` for VOC12, `641,641` for COCO.

```bash

python eval.py \

  --logtostderr \

  --eval_split="val" \

  --model_variant="xception_65" \

  --atrous_rates=6 \

  --atrous_rates=12 \

  --atrous_rates=18 \

  --output_stride=16 \

  --decoder_output_stride=4 \

  --eval_crop_size="${EVAL_CROP_SIZE}" \

  --checkpoint_dir="${TRAIN_LOGDIR}" \

  --eval_logdir="${EVAL_LOGDIR}" \

  --dataset_dir="${DATASET}" \

  --max_number_of_evaluations=1

```

### Visualization

**NOTE:** `${VIS_CROP_SIZE}` should be `513,513` for VOC12, `641,641` for COCO.

```bash

python vis.py \

  --logtostderr \

  --vis_split="val" \

  --model_variant="xception_65" \

  --atrous_rates=6 \

  --atrous_rates=12 \

  --atrous_rates=18 \

  --output_stride=16 \

  --decoder_output_stride=4 \

  --vis_crop_size="${VIS_CROP_SIZE}" \

  --checkpoint_dir="${CKPT}" \

  --vis_logdir="${VIS_LOGDIR}" \

  --dataset_dir="${PASCAL_DATASET}" \

  --also_save_raw_predictions=true

```

### Citation

If you use this work for your research, please cite our paper.

```

@article{zou2020pseudoseg,

  title={PseudoSeg: Designing Pseudo Labels for Semantic Segmentation},

  author={Zou, Yuliang and Zhang, Zizhao and Zhang, Han and Li, Chun-Liang and Bian, Xiao and Huang, Jia-Bin and Pfister, Tomas},

  journal={International Conference on Learning Representations (ICLR)},

  year={2021}

}

```