GRiT Region Caption Evaluation

Modified from https://github.com/JialianW/GRiT/.

Inference

Env

VG Dataset

Download images from official website
Download grit pre-processed annotations: test.json

download weights

mkdir models && cd models
wget https://datarelease.blob.core.windows.net/grit/models/grit_b_densecap_objectdet.pth
cd ..

Eval

python eval_grit.py

Performance

Method	CIDEr	METEOR	SPICE	BLEU@1	BLEU@2	BLEU@3	BLEU@4	ROUGE
GRiT	145.0	17.3	30.9	36.4	22.5	15.6	11.5	34.8

GRiT: A Generative Region-to-text Transformer for Object Understanding

GRiT is a general and open-set object understanding framework that localizes objects and describes them with any style of free-form texts it was trained with, e.g., class names, descriptive sentences (including object attributes, actions, counts and many more).

GRiT: A Generative Region-to-text Transformer for Object Understanding
Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, Lijuan Wang
¹State University of New York at Buffalo, ²Microsoft
arXiv technical report (PDF)

Installation

Please follow Installation instructions.

ChatGPT with GRiT

We give ChatGPT GRiT's dense captioning outputs (object location and description) to have it describe the scene and even write poetry. ChatGPT can generate amazing scene descriptions given our dense captioning outputs. An example is shown below: 🤩🤩🤩

Object Understanding Demo - One Model Two tasks

Download the GRiT model or use the following commend to download:

mkdir models && cd models
wget https://datarelease.blob.core.windows.net/grit/models/grit_b_densecap_objectdet.pth && cd ..

The downloaded GRiT model was jointly trained on dense captioning task and object detection task. With the same trained model, it can output both rich descriptive sentences and short class names by varying the flag --test-task. Play it as follows! 🤩

Output for Dense Captioning (rich descriptive sentences)

python demo.py --test-task DenseCap --config-file configs/GRiT_B_DenseCap_ObjectDet.yaml  --input demo_images --output visualization --opts MODEL.WEIGHTS models/grit_b_densecap_objectdet.pth

Output for Object Detection (short class names)

python demo.py --test-task ObjectDet --config-file configs/GRiT_B_DenseCap_ObjectDet.yaml  --input demo_images --output visualization --opts MODEL.WEIGHTS models/grit_b_densecap_objectdet.pth

Output images will be saved under the visualization folder, which looks like:

You can also try the Colab demo provided by the TWC team:

Benchmark Inference and Evaluation

Please follow dataset preparation instructions to download datasets.

Download our trained models and put them to models/ for evaluation.

Object Detection on COCO 2017 Dataset

Model	val AP	test-dev AP	Download
GRiT (ViT-B)	53.7	53.8	model
GRiT (ViT-L)	56.4	56.6	model
GRiT (ViT-H)	60.4	60.4	model

To evaluate the trained GRiT on coco 2017 val, run:

# GRiT (ViT-B)
python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_B_ObjectDet.yaml --output-dir-name ./output/grit_b_objectdet --eval-only MODEL.WEIGHTS models/grit_b_objectdet.pth
# GRiT (ViT-L)
python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_L_ObjectDet.yaml --output-dir-name ./output/grit_l_objectdet --eval-only MODEL.WEIGHTS models/grit_l_objectdet.pth
# GRiT (ViT-H)
python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_H_ObjectDet.yaml --output-dir-name ./output/grit_h_objectdet --eval-only MODEL.WEIGHTS models/grit_h_objectdet.pth

Dense Captioning on VG Dataset

Model	mAP	Download
GRiT (ViT-B)	15.5	model

To test on VG test set, run:

python train_net.py --num-gpus-per-machine 8 --config-file configs/GRiT_B_DenseCap.yaml --output-dir-name ./output/grit_b_densecap --eval-only MODEL.WEIGHTS models/grit_b_densecap.pth

It will save the inference results to output/grit_b_densecap/vg_instances_results.json. We use the VG dense captioning official evaluation codebase to report the results. We didn't integrate the evaluation code into our project as it was written in Lua. To evaluate on VG, please follow the original codebase's instructions and test based upon it. We're happy to discuss in our issue section about the issues you may encounter when using their code.

Training

To save training memory, we use DeepSpeed for training which can work well for activation checkpointing in distributed training.

To train on single machine node, run:

python train_deepspeed.py --num-gpus-per-machine 8 --config-file configs/GRiT_B_ObjectDet.yaml --output-dir-name ./output/grit_b_objectdet

To train on multiple machine nodes, run:

python train_deepspeed.py --num-machines 4 --num-gpus-per-machine 8 --config-file configs/GRiT_B_ObjectDet.yaml --output-dir-name ./output/grit_b_objectdet

Train without deepspeed

Fix the problem of "AttributeError: module 'distutils' has no attribute 'version'" in torch.utils.tensorboard by pip install setuptools==59.5.0

pytorch/pytorch#69894 (comment)

python train_net.py --num-gpus-per-machine 1 --config-file configs/GRiT_B_DenseCap.yaml --output-dir-name ./output/reimp-grit_b_densecap \
DATALOADER.DATASET_BS 1 \
DATALOADER.NUM_WORKERS 0
# Original DATALOADER.DATASET_BS 2, the GPU memory takes 24G
# from PIL import Image
# Image.fromarray(data[0]["image"].permute(1,2,0).numpy()).save("haha.jpg")
# [batched_inputs[0]["image"].shape] + [i.shape for i in features.values()]
# instances.gt_boxes.nonempty()

Acknowledgement

Our code is in part based on Detic, CenterNet2, detectron2, GIT, and transformers. We thank the authors and appreciate their great works!

Citation

If you find our work interesting and would like to cite it, please use the following BibTeX entry.

@article{wu2022grit,
  title={GRiT: A Generative Region-to-text Transformer for Object Understanding},
  author={Wu, Jialian and Wang, Jianfeng and Yang, Zhengyuan and Gan, Zhe and Liu, Zicheng and Yuan, Junsong and Wang, Lijuan},
  journal={arXiv preprint arXiv:2212.00280},
  year={2022}
}

nicehuster / GRiT-eval