sushizixin / CLIP4IDC

CLIP4IDC: CLIP for Image Difference Captioning (AACL 2022)

CLIP4IDC: CLIP for Image Difference Captioning

The implementation of paper CLIP4IDC: CLIP for Image Difference Captioning.

CLIP4IDC is a image difference captioning model based on CLIP (ViT-B). The model achieve competitive results on CLEVR-Change, Spot-the-Diff and Image Editing Request.

Requirement

# From CLIP
conda install --yes -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0
pip install ftfy regex tqdm
pip install opencv-python boto3 requests pandas
pip install "git+https://github.com/salaniz/pycocoevalcap.git"

Data Preparing

For CLEVR-Change

The official data can be found here: google drive link provided by Robust Change Captioning (ICCV19).

Extracting this file will create data directory.

tar -xzvf clevr_change.tar.gz

For Spot-the-Diff

Resized images can be download from Learning to Describe Differences Between Pairs of Similar Images (EMNLP18). Raw captions can be download from link.

Collecting the captions belonging to the same image pair by running

python preprocess/reformat_dataset.py

For the convenience, you can also download the three json files from link.

You would get

your_data_path
|–– clevr_change/
|   |–– data/
|   |   |–– images/
|   |   |–– nsc_images/
|   |   |–– sc_images/
|   |   |–– sc_images/
|   |   |–– change_captions.json
|   |   |–– no_change_captions.json
|   |   |–– splits.json
|   |   |–– type_mapping.json
|–– spot-the-diff/
|   |–– images/
|   |–– data/
|   |–– train.json
|   |–– val.json
|   |–– test.json
|   |–– reformat_train.json
|   |–– reformat_val.json
|   |–– reformat_test.json

Prepare for Evaluation

For CLEVR-Change

To evaluate captions, we need to first reformat the caption annotations into COCO eval tool format. Please run the command python utils/eval_utils.py according to the instructions given in Evaluation in Robust Change Captioning (ICCV19).

Renaming the output file as clevr_total_change_captions_reformat.json.

For Spot-the-Diff

Running the command python gt/eval_utils.py, renaming the output file as spot_total_change_captions_reformat.json. You would get

gt
|–– clevr_total_change_captions_reformat.json
|–– spot_total_change_captions_reformat.json

Pretrained Weight

cd ckpts
mkdir pretrained
mkdir trained

You can download the Pretrained Weights from the IDC Adaptation and the Trained Weights from the IDC Finetuning. You would get

ckpts
|–– pretrained/
|   |–– pytorch_model.bin.clevr
|   |–– pytorch_model.bin.spot
|–– trained/
|   |–– pytorch_model.bin.clevr
|   |–– pytorch_model.bin.spot

The pretrained weights are the output of adaptation (retrieval) stage. The trained weights are the output of the finetuning (captioning) stage.

How to Run

--features_path is the data root path

--pretrained_clip_name can be set with ViT-B/32

--resume_model can be used to reload the saved optimizer state to continuely train the model, Note: need to set the corresponding chechpoint via --init_model simultaneously.

Download CLIP (ViT-B/32) weight,

wget -P ./modules https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt

Adapation

Experiments are conducted on two NVIDIA V100. Time required for each task is less than 24h.

CLEVR_Change

DATA_PATH=[Your CLEVR_Change data path]
python -m torch.distributed.launch --nproc_per_node=2 main_task_retrieval.py \
--do_train \
--num_thread_reader=4 \    # Please don't change this value when reproducing the results
--epochs=12 \
--batch_size=128 \
--n_display=50 \
--data_path ${DATA_PATH} \
--features_path ${DATA_PATH} \
--output_dir ckpts/ckpt_clevr_retrieval \
--lr 1e-4 \
--max_words 32 \
--batch_size_val 128 \
--datatype clevr \
--coef_lr 1e-3 \
--freeze_layer_num 0 \
--linear_patch 2d \
--pretrained_clip_name ViT-B/32 

The Text-to-Image-Pair retrieval results are close to

R@1: 26.8 - R@5: 58.7 - R@10: 70.0

The Image-Pair-to-Text retrieval results are close to

R@1: 46.4 - R@5: 83.0 - R@10: 86.6

Spot-the-Diff

DATA_PATH=[Your Spot-the-Diff data path]
python -m torch.distributed.launch --nproc_per_node=2 main_task_retrieval.py \
--do_train \
--num_thread_reader=4 \   
--epochs=20 \
--batch_size=128 \
--n_display=50 \
--data_path ${DATA_PATH} \
--features_path ${DATA_PATH}/images \
--output_dir ckpts/ckpt_spot_retrieval \
--lr 1e-4 \
--max_words 32 \
--batch_size_val 128 \
--datatype spot \
--coef_lr 1e-3 \
--freeze_layer_num 0 \
--linear_patch 2d \
--pretrained_clip_name ViT-B/32

Finetuning

Time required for each task is less than 24h.

CLEVR_Change

Reproducing the results on the single NVIDIA V100.

DATA_PATH=[Your CLEVR_Change data path]
python -m torch.distributed.launch --nproc_per_node=1 main_task_caption.py \
--do_train \
--num_thread_reader=4 \  # Please don't change this value when reproducing the results
--epochs=50 \
--batch_size=16 \
--n_display=50 \
--data_path ${DATA_PATH} \
--features_path ${DATA_PATH} \
--output_dir ckpts/ckpt_clevr_caption \
--lr 1e-4 \
--max_words 32 \
--batch_size_val 64 \
--datatype clevr \
--coef_lr 1e-3 \
--freeze_layer_num 0 \
--linear_patch 2d \
--pretrained_clip_name ViT-B/32 \
--init_model ckpts/pretrained/pytorch_model.bin.clevr \
--seed 2021 

The best results are obtained at epoch 19

BLEU_1: 0.8648, BLEU_2: 0.7797, BLEU_3: 0.6758, BLEU_4: 0.5687
METEOR: 0.3840, ROUGE_L: 0.7643, CIDEr: 1.5075

Reproducing the results on Two NVIDIA V100.

DATA_PATH=[Your CLEVR_Change data path]
python -m torch.distributed.launch --nproc_per_node=2 main_task_caption.py \
--do_train \
--num_thread_reader=4 \
--epochs=50 \
--batch_size=64 \
--n_display=50 \
--data_path ${DATA_PATH} \
--features_path ${DATA_PATH} \
--output_dir ckpts/ckpt_clevr_caption_multigpu \
--lr 1e-4 \
--max_words 32 \
--batch_size_val 64 \
--datatype clevr \
--coef_lr 1e-3 \
--freeze_layer_num 0 \
--linear_patch 2d \
--pretrained_clip_name ViT-B/32 \
--init_model ckpts/pretrained/pytorch_model.bin.clevr \
--seed 2021 

The best results are obtained at Epoch 26

BLEU_1: 0.8573, BLEU_2: 0.7761, BLEU_3: 0.6734, BLEU_4: 0.5663
METEOR: 0.3900, ROUGE_L: 0.7640, CIDEr: 1.5039

Effects of Seed

We find that the random seed has great effects on the results. The following results of other seeds are reproduced on Two NVIDIA V100.

Seed	BLEU_4	METEOR	ROUGE_L	CIDEr
42	53.50	40.02	72.12	133.52
44	54.76	40.59	72.24	131.23
100	53.21	37.93	75.78	146.69
225	52.85	37.84	74.92	141.77
800	56.52	38.16	76.00	146.29
1024	55.30	39.21	75.35	139.39
2048	54.60	38.49	76.05	147.50
6666	55.65	39.16	76.44	149.17
mean	~54.55	~38.93	~74.86	~141.95
sample var	~1.44	~1.01	~2.95	~45.26

(For Seed results) You could also replace

    caption = random.choice(caption)  # Line 193 in dataloader_clevr_caption.py
    
    no_caption = random.choice(no_caption)  # Line 204 in dataloader_clevr_caption.py

with

if self.subset == "train":
    caption = random.choice(caption)
else:
    caption = "none" 

and

if self.subset == "train":
    no_caption = random.choice(no_caption)
else:
    no_caption = "none" 

respectively.

Spot-the-Diff

DATA_PATH=[Your Spot-the-Diff data path]
python -m torch.distributed.launch --nproc_per_node=1 main_task_retrieval.py \
--do_train \
--num_thread_reader=4 \
--epochs=50 \
--batch_size=16 \
--n_display=50 \
--data_path ${DATA_PATH} \
--features_path ${DATA_PATH}/images \
--output_dir ckpts/ckpt_spot_caption \
--lr 1e-4 \
--max_words 32 \
--batch_size_val 32 \
--datatype spot \
--coef_lr 1e-3 \
--freeze_layer_num 0 \
--linear_patch 2d \
--pretrained_clip_name ViT-B/32 \
--init_model ckpts/pretrained/pytorch_model.bin.spot 

Take note of the commented out lines used to fix the random seed. The results of several runs should be comparable with

BLEU_1: 0.4145, BLEU_2: 0.2760, BLEU_3: 0.1782, BLEU_4: 0.1160
METEOR: 0.1419, ROUGE_L: 0.3502, CIDEr: 0.4735

Running scripts are provided in scripts.

Citation

If you find CLIP4IDC useful in your work, you can cite the following paper:

@article{guo2022clip4idc,
  title={CLIP4IDC: CLIP for Image Difference Captioning},
  author={Guo, Zixin and Wang, Tzu-Jui Julius and Laaksonen, Jorma},
  journal={arXiv preprint arXiv:2206.00629},
  year={2022}
}

Acknowledgments

Our code is largely borrowed from CLIP, UniVL and CLIP4Clip.

TODO

We are sorry that some lines of the code are redundant and some variables are named with "video".