Instructions to use JosephZ/R1-SGG-7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use JosephZ/R1-SGG-7B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="JosephZ/R1-SGG-7B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("JosephZ/R1-SGG-7B")
model = AutoModelForImageTextToText.from_pretrained("JosephZ/R1-SGG-7B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use JosephZ/R1-SGG-7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "JosephZ/R1-SGG-7B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "JosephZ/R1-SGG-7B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/JosephZ/R1-SGG-7B

SGLang

How to use JosephZ/R1-SGG-7B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "JosephZ/R1-SGG-7B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "JosephZ/R1-SGG-7B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "JosephZ/R1-SGG-7B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "JosephZ/R1-SGG-7B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use JosephZ/R1-SGG-7B with Docker Model Runner:
```
docker model run hf.co/JosephZ/R1-SGG-7B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Model Description

An end-to-end multimodal LLM for Scene Graph Generation (SGG), which was introduced in Compile Scene Graphs with Reinforcement Learning

R1-SGG: Compile Scene Graphs with Reinforcement Learning

Structured Visual Reasoning with Multimodal LLMs and Reinforcement Learning

![Paper](https://img.shields.io/badge/arXiv-2504.13617-b31b1b.svg)

🚀 Update

✅ R1-SGG-7B, R1-SGG-Zero-7B
✅ Support PSG dataset (bbox format only, not Panoptic)
✅ Updated loss implementation
✅ Always use custom_per_device_train_batch_size instead of per_device_train_batch_size for faster sampling under gradient accumulation
⚠️ Current loss implementation might still be affected by gradient accumulation: trl issue #3021

🛠️ Setup Environment

bash install.sh

Main dependencies:

- torch == 2.5.0 or 2.5.1 (cu124, optional)
- transformers (supports Qwen2VL, Qwen2.5VL)
- trl
- vLLM

📚 Dataset

Load preprocessed datasets via:

from datasets import load_dataset

db_train = load_dataset("JosephZ/vg150_train_sgg_prompt")["train"]
db_val = load_dataset("JosephZ/vg150_val_sgg_prompt")["train"]

or for PSG:

db_train = load_dataset("JosephZ/psg_train_sg")["train"]  # keys: image_id, image, objects, relationships
db_val = load_dataset("JosephZ/psg_test_sg")["train"]

We transformed VG150 into HuggingFace Datasets format with keys:

image_id
image
prompt_open
prompt_close
objects
relationships

🔥 Supported Models

Qwen/Qwen2-VL-2B-Instruct
Qwen/Qwen2-VL-7B-Instruct
Qwen/Qwen2.5-VL-3B-Instruct
Qwen/Qwen2.5-VL-7B-Instruct

🏋️‍♂️ Training

Training with Supervised Fine-Tuning (SFT)

For SLURM users:

sbatch scripts/sft/7B_sgg.sh

For local machines:

bash scripts/sft_local/7B_sgg.sh

⏱️ Approximate training time:

2B models: ~4 hours (4×A100 SXM4 GPUs)
7B models: ~10 hours (4×A100 SXM4 GPUs)

Training with Reinforcement Learning (GRPO)

** Update (11/05/2025): to use "Hard Recall"**:

--reward_funcs format_reward edge_hard_reward

For A100 GPUs:

sbatch scripts/grpo/train_a100_2B.sh

(12 hours on 16×A100 GPUs)

For GH200 GPUs:

sbatch scripts/grpo/train_gh200.sh

(16 hours on 16×GH200 GPUs)

For clusters with many RTX_3090/4090 GPUs:

sbatch scripts/grpo/train_fused.sh

Training 7B models on 24GB cards is possible with Zero3, but slow due to communication bottlenecks.
(Fun fact: training with 120×RTX_4090 is crazy but severely limited by communication latency.)

💡 Recommended learning rate: 6e-7.

🧪 Inference and Evaluation

Inference with SFT-trained models:

bash scripts/inference/run_sgg_inference.sh $DATASET $MODEL_NAME $OUTPUT_DIR

For models trained with predefined categories, add true:

bash scripts/inference/run_sgg_inference.sh $DATASET $MODEL_NAME $OUTPUT_DIR true

Inference with GRPO-trained models:

bash scripts/inference/run_sgg_inference.sh $DATASET $MODEL_NAME $OUTPUT_DIR false/true true

Evaluation:

DATASET_TYPE=vg # or psg
python src/sgg_gather_preds.py $DATASET_TYPE $OUTPUT_DIR sgg_pred_results.json
python src/vg150_eval.py $DATASET sgg_pred_results.json

🤝 Acknowledgement

The GRPOTrainer used in this project is based on trl's GRPOTrainer, extended to support multimodal inputs.

📖 Citation

If you find this work helpful, please cite:

@article{chen2025compile,
  title={Compile Scene Graphs with Reinforcement Learning},
  author={Chen, Zuyao and Wu, Jinlin and Lei, Zhen and Pollefeys, Marc and Chen, Chang Wen},
  journal={arXiv preprint arXiv:2504.13617},
  year={2025}
}