Need help with SGAE?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

yangxuntu
143 Stars 39 Forks 27 Commits 30 Opened issues

Services available

!
?

Need anything else?

Contributors list

No Data

SGAE/ pytorch 0.4.0

Auto-Encoding Scene Graphs for Image Captioning, CVPR 2019

Acknowledgement

This code is implemented based on Ruotian Luo's implementation of image captioning in https://github.com/ruotianluo/self-critical.pytorch.

And we use the visual features provided by paper Bottom-up and top-down attention for image captioning and visual question answering in https://github.com/peteanderson80/bottom-up-attention.

If you like this code, please consider to cite their corresponding papers and my CVPR paper.

Installation anaconda and the environment

I provide the anaconda environment for running my code in https://drive.google.com/drive/folders/1GvwpchUnfqUjvlpWTYbmEvhvkJTIWWRb?usp=sharing. You should download the file ''environmentyx1.yml'' from this link and set up the environment as follows. 1.Download the anaconda from the website https://www.anaconda.com/ and install it. 2.Go to website https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html?highlight=environment to learn how to learn how to 'creating an environment from an environment.yml file'. ``` conda env create -f environmentyx1.yml

3.After installing anaconda and setting up the environment, run the following code to get into the environment.
source activate yx1
If you want to exit from this environment, you can run the following code to exit.
source deactivate ```

Downloading meta data, e.g., image captions, visual features, image scene graphs, sentence scene graphs.

You can get more details from https://github.com/ruotianluo/self-critical.pytorch.

1.Download preprocessed coco captions from link from Karpathy's homepage. Extract dataset_coco.json from the zip file and copy it in to data/.

This file provides preprocessed captions and also standard train-val-test splits. The do:

python scripts/prepro_labels.py --input_json data/dataset_coco.json --output_json data/cocotalk.json --output_h5 data/cocotalk
preprolabels.py will map all words that occur <= 5 times to a special UNK token, and create a vocabulary for all the remaining words. The image information and vocabulary are dumped into data/cocotalk.json and discretized caption data are dumped into data/cocotalklabel.h5.

Also, you can download the file 'cocobu2.json' and 'cocobu2_label.h5' from https://drive.google.com/drive/folders/1GvwpchUnfqUjvlpWTYbmEvhvkJTIWWRb?usp=sharing and put them into the folder 'data' (if you do not have this folder, just create one), which are processed by myself for facilitating the usage of this code. I also release two well-trained models based on these two files which are modelid740072 and modelid640075.

2.Download Bottom-up features.

Download pre-extracted feature from https://github.com/peteanderson80/bottom-up-attention. You can either download adaptive one or fixed one. We use the ''10 to 100 features per image (adaptive)'' in our experiments. For example:

mkdir data/bu_data; cd data/bu_data
wget https://storage.googleapis.com/bottom-up-attention/trainval.zip
unzip trainval.zip
Then :
python script/make_bu_data.py --output_dir data/cocobu
This will create data/cocobufc, data/cocobuatt and data/cocobu_box. If you want to use bottom-up feature, you can just follow the following steps and replace all cocotalk with cocobu.

3.Download the extracted image scene graph and sentence scene graph.

Download the files 'cocopredsg.zip' and 'cocospicesg2.zip' from https://drive.google.com/drive/folders/1GvwpchUnfqUjvlpWTYbmEvhvkJTIWWRb?usp=sharing and put them into the folder 'data' and then unzip them. The file 'cocopredsg.zip' contains all the image scene graphs and 'cocospicesg2.zip' contains all the sentence scene graphs.

Training the model

1.After downloading the codes and meta data, you can train the model by using the following code:

python train_mem.py --id id66 --caption_model lstm_mem4 --input_json data/cocobu2.json --input_fc_dir data/cocobu_fc --input_att_dir data/cocobu_att --input_rela_dir data/cocobu_sg_img --input_ssg_dir data/coco_spice_sg2 --input_label_h5 data/cocobu2_label.h5 --sg_dict_path data/spice_sg_dict2.npz --batch_size 50 --accumulate_number 2 --learning_rate 5e-4 --learning_rate_decay_start 0 --learning_rate_decay_every 5 --scheduled_sampling_start 0 --checkpoint_path id66 --save_checkpoint_every 5000 --val_images_use 5000 --max_epochs 150 --rnn_size 1000 --input_encoding_size 1000 --att_feat_size 2048 --att_hid_size 512 --self_critical_after 40 --train_split train --memory_size 10000 --memory_index c --step2_train_after 10 --step3_train_after 20 --use_rela 0 --gpu 5
Important notes: I reorganized and optimized the code recently and found that even without MGCN, this code can achieve 127.8 CIDEr-D score. But you need to have a 16G gpu like DGX. If your memory is not enough, you should change --batchsize from 50 to 25, and --accumulatenumber from 2 to 4, which can make the batch size be 100. But I found that these two different settings will lead to different performances.

2.You can also go to my google drive to download two well-trained models which are modelid740072 and modelid640075 for getting about 128.3 CIDEr-D scores, and these two models are trained by using the above scripts.

3.The details of parameters:

--id: the id of your model, which is usually set as the same as check point, which is helpful for you to train from the check point.

--batchsize, --accumulatenumber: these two parameters are set for users who do not have large gpu, if you want to set batch size as 100, you can set batchsize as 50, and set accumulatenumber as 2, also you can set batchsize as 20 and accumulatenumber as 5. Importantly, they are not totally equal to set batchsize as 100 and accumulatenumber as 1, the bigger the bathc_size is, the higher the performance.

--selfcriticalafter: when reinforcement leanring begins, if this value is set as 40, it means that after training 40 epoches, the reinforcement loss is used. Generally, if you want to have a good CIDEr-D score, you should use cross entropy loss first and then use reinforcement loss.

--step2trainafter: when the dictionary is learned, for example, if this value is set as 10, then before 10 epochs, only the decoder is trained by sentence scene graphs and the dictionary is not learned.

--step3trainafter: when image captioning encoder-decoder is learned, for example, if this value is set as 20, then before 20 epochs, only sentence scene graphs are used to learn the dictionary, and after 20 epochs, the sentence scene graphs are no longer used and the image encoder-decoder is trained.

--use_rela: whether use image scene graph

4.Tranining from checkpoints. The codes provide the ability of training from checkpoints. For example, if you want to train the model from one checkpoint, say, 22, you can use the following code to continute:

python train_mem.py --id id66 --caption_model lstm_mem4 --input_json data/cocobu2.json --input_fc_dir data/cocobu_fc --input_att_dir data/cocobu_att --input_rela_dir data/cocobu_sg_img --input_ssg_dir data/coco_spice_sg2 --input_label_h5 data/cocobu2_label.h5 --sg_dict_path data/spice_sg_dict2.npz --batch_size 50 --accumulate_number 2 --learning_rate 5e-4 --learning_rate_decay_start 0 --learning_rate_decay_every 5 --scheduled_sampling_start 0 --checkpoint_path id66 --save_checkpoint_every 5000 --val_images_use 5000 --max_epochs 150 --rnn_size 1000 --input_encoding_size 1000 --att_feat_size 2048 --att_hid_size 512 --self_critical_after 40 --train_split train --memory_size 10000 --memory_index c --step2_train_after 10 --step3_train_after 20 --use_rela 0 --gpu 5 --start_from 22 --memory_cell_path id66/memory_cellid660022.npz
the parameter startfrom and memorycell_path are used for training from checkpoints.

Evaluating the model

1.After training the model or downloading the well-trained model, you can evaluate them by using the following code:

python eval_mem.py --dump_images 0 --num_images 5000 --model id66/modelid660066.pth --infos_path id66/infos_id660066.pkl --language_eval 1 --beam_size 5 --split test --index_eval 1 --use_rela 0 --training_mode 2 --memory_cell_path id66/memory_cellid660066.npz --sg_dict_path data/spice_sg_dict2.npz --input_ssg_dir data/coco_spice_sg2 --batch_size 50
what you need to do is to switch the model id with your id, like 66 to 01, and change the number like 0066 to your trained model, like 0066 to 0001.

Generating Scene graphs:

1.For sentence scene graph, you can directly download the revised code in spice-1.0.jar and createcocosg.py, put spice-1.0.jar in /coco-caption/pycocoevalcap/spice, then you should set cocouse as cocotrain or cocoval in file createcoco_sg.py, then run this code and the sentence scene graphs are generated in /coco-caption/pycocoevalcap/spice/sg.json.

2.Then use processspicesg.py to process sg.json.

3.For image scene graph, you can directly download the code provided by https://github.com/rowanz/neural-motifs for generating image scene graphs.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.