zsl-gcn-pth

KO | EN

gitlite — search

#javascript #python #react #ai #llm #android #deep-learning #docker #awesome #awesome-list #pytorch #security #openai #open-source

(★ 80)

zero-shot-gcn in pytorch

File Explorer

Download Latest Version (.zip)

- - git-gcn-teaser.png
  - test+train_fc_c.png
  - test+train_wv_c.png
  - train_fc_animal.png
  - train_fc_c.png
  - train_fc_ccarnivore.png
  - train_fc_ccontainer.png
  - train_fc_inst.png
  - train_wv_animal.png
  - train_wv_c.png
  - train_wv_ccarnivore.png
  - train_wv_ccontainer.png
  - train_wv_inst.png
- - 1k-2-hops.txt
  - 1k-3-hops.txt
  - 1k-all.txt
  - 1k.txt
  - 2-hops.txt
  - 3-hops.txt
  - all.txt
  - corresp-2-hops.json
  - corresp-3-hops.json
  - corresp-all.json
  - invdict_wordn.json
  - invdict_wordntext.json
  - words.pkl
  - words.txt
- imagenet_graph.pkl
- word_embedding_model
- - convert.py
  - layers.py
  - metrics.py
  - models.py
  - train_gcn.py
  - utils.py
- - convert_to_gcn_data.py
  - download_image.py
  - extract_pool5.py
  - obtain_word_embedding.py
  - prepare_list.py
  - tmp_test.py
- preprocess_data.sh
- test_imagenet.py
- test_imagenet_conse.py
- test_imagenet_pll.py
- train_deploy.sh
.gitignore
DATASET.md
feats
images
README.md

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Zero-shot GCN in pytorch

## Introduction
This repository replicates the result in [zero-shot-gcn](https://github.com/JudyYe/zero-shot-gcn/tree/master/src). The code is largely migrated from

## Requirement
+pytorch (tensorflow is still needed for tf.FLAGS)

## Difference to the original version

- Fix the wrong word embedding
- Fix the wrong dropout setting
- Change a few things as type in [adgpm](https://github.com/cyvius96/adgpm).

**None of these actually change the result much**.

## Before training

**Extract features**

Download the ImageNet22k images in your own ways.

Download resnet model from [site](https://drive.google.com/open?id=0B7fNdx_jAqhtbVYzOURMdDNHSGM).

(You can also use the torchvision pretrained model (which is used by [adgpm](https://github.com/cyvius96/adgpm)). That actually gives better result.)

```
python main.py xxx/ImageNet22k/images/ --arch resnet50 --pretrained --evaluate --batch-size 750 --workers 24
```

Then create a symbolic link for features to `../feats`.

**Preprocess**
```
python tools/obtain_word_embedding.py
python convert_to_gcn_data.py --fc res50 --wv glove
```

## Training
```
python gcn/train_gcn.py --dataset ../data/glove_res50/
```

(To run the gpm model in adgpm paper, use following command
```
python gcn/train_gcn.py --dataset ../data/glove_res50/ --save_path log_gpm --hiddens 2048d,d --adj_norm_type in --feat_norm_type l2
```
)

## Test

Evaluate images in parallel (Faster than the original tensorflow version)
```
python test_imagenet_pll.py --model xxxx/feat__300 --feat ../feats
```

To evaluate conse result:
```
python test_imagenet_conse.py --model xxxx/feat__300 --feat ../feats
```

# Original README

This code is a re-implementation of the zero-shot classification in ImageNet in the paper [Zero-shot Recognition via Semantic Embeddings and Knowledge Graphs](https://arxiv.org/abs/1803.08035). The code is developed based on the [TensorFlow framework](https://www.tensorflow.org/) and the Graph Convolutional Network (GCN) [repo](https://github.com/tkipf/gcn/tree/master/gcn).

![](https://raw.githubusercontent.com/ruotianluo/zsl-gcn-pth/master/data/docs/git-gcn-teaser.png)

Our pipeline consists of two parts: CNN and GCN.
- **CNN**: **Input** an image and **output** deep features for the image.
- **GCN**: **Input** the word embedding for every object class, and **output** the visual classifier for every object class. Each visual classifier (1-D weight vector) can be applied on the deep features for classification.

## Citation
If you use our code in your research or wish to refer to the benchmark results, please use the following BibTeX entry.
```
@article{wang2018zero,
  title={Zero-shot Recognition via Semantic Embeddings and Knowledge Graphs},
  author={Wang, Xiaolong and Ye, Yufei and Gupta, Abhinav},
  journal={CVPR},
  year={2018}
}
```

## Using Our Code

```bash
git clone git@github.com:JudyYe/zero-shot-gcn.git
cd zero-shot-gcn/src
```
Without further specification, we default the root directory to `zero-shot-gcn/src`.

## Dataset Preparation
Please read [`DATASET.md`](DATASET.md) for downloading images and extracting image features.

## Testing Demo
With extracted feature and semantic embeddings, at this point, we can perform zero-shot classification with the [model](https://www.dropbox.com/sh/q9mid4wjj5vy0si/AADg8_NobfxkDot3VM7tE8Fua?dl=0) we provide.
```Shell
wget -O ../data/wordnet_resnet_glove_feat_2048_1024_512_300 https://www.dropbox.com/s/e7jg00nx0h2gbte/wordnet_resnet_glove_feat_2048_1024_512_300?dl=0
python test_imagenet.py --model ../data/wordnet_resnet_glove_feat_2048_1024_512_300
```
The above line defaults to `res50` + `2-hops` combination and test under two settings: unseen classes with or without seen classes. (see the paper for further explaination.)

We also provide other configurations. Please refer to the code for details.

## Main Results
We report the results with the above testing demo code (using ResNet-50 visual features and GloVe word embeddings). All experiments are conducted with the ImageNet dataset.

We first report the results on testing with only unseen classes. We compare our method with the state-of-the-art method [`SYNC`](https://arxiv.org/abs/1603.00550) in this benchmark.

<center>
 
ImageNet Subset |Method | top 1 | top 2 |top 5 |top 10| top 20 |
---|---|---|---|---|---|---|
2-hops | SYNC GCNZ (Ours) | 10.5 **21.0** |	17.7 **33.7** | 28.6 **52.7** |	40.1 **64.8** |	52.0 **74.3**
3-hops | SYNC GCNZ (Ours) | 2.9 **4.3** | 	4.9 **7.7** |9.2 **14.2**	| 14.2 **20.4** |	20.9 **27.6** |
All | SYNC GCNZ (Ours) | 1.4 **1.9**	|2.4 **3.4** |4.5 **6.4**	|7.1 **9.3**	|10.9 **12.7** |

</center>

We then report the results under the generalized zero-shot setting, i.e. testing with both unseen and seen classes. We compare our method with the state-of-the-art method [`ConSE`](https://arxiv.org/abs/1312.5650) in this benchmark.

<center>
 
ImageNet Subset |Method | top 1 | top 2 |top 5 |top 10| top 20 |
---|---|---|---|---|---|---|
2-hops (+1K) | ConSE GCNZ (Ours) | 0.1 **10.2** |	11.2 **21.2** | 24.3 **42.1** |	29.1 **56.2** |	32.7 **67.5**
3-hops (+1K) | ConSE GCNZ (Ours) | 0.2 **2.4** | 	3.2 **5.3** |7.3 **12.0**	| 10.0 **18.2** |12.2 **25.4** |
All (+1K) | ConSE GCNZ (Ours) | 0.1 **1.1**	|1.5 **2.4** |3.5 **5.4**	|4.9 **8.3**	|6.2 **11.7** |

</center>

We also visualize the t-SNE plots of GCN inputs and outputs for two subtrees of WordNet as followings.

synset word|t-SNE of input word embeddings | t-SNE of output visual classifiers
---|:-------------------------:|:-------------------------:
instrumentality instrumentation|![](https://raw.githubusercontent.com/ruotianluo/zsl-gcn-pth/master/data/docs/train_wv_inst.png) | ![](https://raw.githubusercontent.com/ruotianluo/zsl-gcn-pth/master/data/docs/train_fc_inst.png)
animal, animate being, beast, brute, creature, fauna | ![](https://raw.githubusercontent.com/ruotianluo/zsl-gcn-pth/master/data/docs/train_wv_animal.png) | ![](https://raw.githubusercontent.com/ruotianluo/zsl-gcn-pth/master/data/docs/train_fc_animal.png)

## Training
As [`DATASET.md`](DATASET.md) illustrates, `convert_to_gcn_data.py` prepares data to train GCN. It supports two CNN network `fc` = `res50` or `inception`, and three semantic embedding `wv` = `glove` or `google` or `fasttext`.
The output will be saved to ../data/$wv_$fc/
```Shell
python convert_to_gcn_data.py --fc res50 --wv glove
```

After preparing the data, we can start training by using:
```Shell
python gcn/train_gcn.py --gpu $GPU_ID 	--dataset ../data/glove_res50/ --save_path $SAVE_PATH
```

zsl-gcn-pth

File Explorer

Today's picks

Was this content helpful?

Recent Feedback

zsl-gcn-pth

File Explorer

Today's picks

Was this content helpful?

Recent Feedback

Download README