Hugging Face's logo

5CD-AI
/
Vintern-4B-v1

Visual Question Answering
Transformers
Safetensors
Vietnamese
Chinese
English
internvl_chat
feature-extraction
custom_code
Model card Files Community
Edit model card

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

Vintern-4B-v1 ❄️ - The LLaVA 🌋 Challenger

What's new in Vintern-4B-v1!

  • We successfully reproduced the training process of InternVL from scratch.
  • The model is the result of integrating sail/Sailor-4B-Chat and InternViT-300M-448px through an MLP layer.
  • It promises better inference and image recognition results, text and OCR compared to the Vintern-1B version.
  • Additional experiments are currently being conducted.

Model Details

Model Name Vision Part Language Part
Vintern-4B-v1 InternViT-300M-448px Sailor-4B-Chat

Zalo VMLU Benchmark

The Vintern-4B-v1 achieved a score of 42.87 on the Zalo VMLU Benchmark.

generation_config = dict(max_new_tokens= 64, do_sample=False, num_beams = 1, repetition_penalty=3.5)
question = "Bạn là thầy giáo giải trắc nghiệm rất chính xác. Bạn biết chắc chắn đáp án đúng nhất. Chỉ đưa ra chữ cái đứng trước câu trả lời đúng của câu hỏi trắc nghiệm sau: Một doanh nghiệp có vốn đầu tư nước ngoài có trụ sở chính ở Việt Nam, thì: Lựa Chọn: A. Được ĐKDN và HĐKD theo pháp luật Việt Nam B. Được ĐKDN và HĐKD theo pháp luật nước ngoài C. Được ĐKDN và HĐKD theo pháp luật Việt Nam và pháp luật nước ngoài tùy theo từng vấn đề cụ thể D. Cả A, B và C đều sai"
model.chat(tokenizer, None, question, generation_config)

PhoGPT Benchmark

The Vintern-4B-v1 improved its performance compared to Sailor-4B-Chat, increasing from 21 correct answers to 49 correct answers out of 147 Vietnam-specific questions on the PhoGPT benchmark.

Model Vietnam-specific
PhoGPT-4B-Chat 43.5 (64 / 147)
Vistral-7B-Chat 42.9 (63 / 147)
GPT-4-0125-preview 39.5 (58 / 147)
Gemini Pro 1.0 34.7 (51 / 147)
Vintern-4B-v1 33.3 (49 / 147)
Sailor-7B-Chat 27.9 (41 / 147)
GPT-3.5-turbo 22.4 (33 / 147)
*Sailor-4B-Chat 14.3 (21 / 147)
SeaLLM-7B-v2 13.6 (20 / 147)
VBD-Llama2-7B-50B-Chat 10.9 (16 / 147)
Vinallama-7B-Chat 8.2 (12 / 147)
Gemma-7B-it 6.1 (9 / 147) 
generation_config = dict(max_new_tokens= 128, do_sample=False, num_beams = 2, repetition_penalty=1.5)
question = "Hãy trả lời câu hỏi liên quan đến kiến thức về Việt Nam sau: Anh hùng dân tộc nào dời đô từ Hoa Lư về Thăng Long" 
output = model.chat(tokenizer, None, question, generation_config, return_history=True)
# Anh hùng dân tộc dời đô từ Hoa Lư về Thăng Long là Lý Công Uẩn.

VLSP2023: ViVRC Challenge Benchmark

Name F1
ICNLP 3.6384
Vintern-4B-v1 3.5514
Vintern-1B-v2 3.4616
linh 3.4293
DS@ViVRC 3.4121
DS@UIT Dynasty 3.3172
NTQ Solution 3.2926
I, Me & Myself 3.2396
AVQA_AIO 2.9018
Vintern-1B-v1 2.7256
NguyenLe 2.7053
nowj2 1.6808

Other benchmarks are being run .......

................................................................................................

Quickstart

Here provides a code snippet to show you how to load the tokenizer and model and how to generate contents. To run inference using the model, follow the steps outlined in our Colab inference notebook

import numpy as np
import torch
import torchvision.transforms as T
# from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

model = AutoModel.from_pretrained(
    "5CD-AI/Vintern-4B-v1",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained("5CD-AI/Vintern-4B-v1", trust_remote_code=True, use_fast=False)

test_image = 'test-image.jpg'

pixel_values = load_image(test_image, max_num=6).to(torch.bfloat16).cuda()
generation_config = dict(max_new_tokens= 512, do_sample=False, num_beams = 3, repetition_penalty=3.5)

question = '<image>\nMô tả hình ảnh một cách chi tiết.'

response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

#question = "Câu hỏi khác ......"
#response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
#print(f'User: {question}\nAssistant: {response}')
Downloads last month
0
Safetensors
Model size
4.27B params
Tensor type
BF16
·
Inference Examples
Visual Question Answering
Inference API (serverless) is not available, repository is disabled.

Model tree for 5CD-AI/Vintern-4B-v1

Base model

OpenGVLab/InternViT-300M-448px
Finetuned
this model