README.md

metadata

license: apache-2.0
language:
  - en

UForm

Pocket-Sized Multimodal AI
For Content Understanding and Generation

Description

UForm-Gen is a small generative vision-language model primarily designed for Image Captioning and Visual Question Answering. The model consists of two parts:

1. UForm Vision Encoder
2. Sheared-LLaMA-1.3B manually tuned on the instruction dataset

The model was pre-trained on: MSCOCO, SBU Captions, Visual Genome, VQAv2, GQA and a few internal datasets. UForm-Gen-Chat is SFT version of UForm-Gen for multimodal chat.

Usage

pip install uform

from uform.gen_model import VLMForCausalLM, VLMProcessor

model = VLMForCausalLM.from_pretrained("unum-cloud/uform-gen-chat")
processor = VLMProcessor.from_pretrained("unum-cloud/uform-gen-chat")

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "<image> {Your message}"}
]

image = processor.image_processor(Image.open("zebra.jpg")).unsqueeze(0)

input_ids = processor.tokenizer.apply_chat_template(
    messages, return_tensors="pt", add_generation_prompt=True
)

attention_mask = torch.ones(1, input_ids.shape[1] + processor.num_image_latents - 1)

inputs = {
    "input_ids": input_ids,
    "attention_mask": attention_mask,
    "images": image,
}

outputs = model.generate(
    **inputs,
    do_sample=False,
    use_cache=True,
    max_new_tokens=1024,
    eos_token_id=32001,
    pad_token_id=processor.tokenizer.pad_token_id,
)

message = processor.batch_decode(outputs[:, inputs["input_ids"].shape[1]:-1])

Evaluation

For captioning evaluation we measure CLIPScore and RefCLIPScore¹.

Model	Size	Caption Length	CLIPScore	RefCLIPScore
llava-hf/llava-1.5-7b-hf	7B	Long	0.878	0.529
llava-hf/llava-1.5-7b-hf	7B	Short	0.886	0.531
Salesforce/instructblip-vicuna-7b	7B	Long	0.902	0.534
Salesforce/instructblip-vicuna-7b	7B	Short	0.848	0.523
unum-cloud/uform-gen-chat	1.5B	Long	0.860	0.525
unum-cloud/uform-gen-chat	1.5B	Short	0.858	0.525

¹ We used apple/DFN5B-CLIP-ViT-H-14-378 CLIP model.