JaronTHU/

Video-CCAM-4B-v1.1

import os
import torch
from PIL import Image
from transformers import AutoModel

from eval import load_decord

os.environ['TOKENIZERS_PARALLELISM'] = 'false'

videoccam = AutoModel.from_pretrained(
    '<your_local_path_1>',
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map='auto',
    _attn_implementation='flash_attention_2',
    # llm_name_or_path='<your_local_llm_path>',
    # vision_encoder_name_or_path='<your_local_vision_encoder_path>'
)


messages = [
    [
        {
            'role': 'user',
            'content': '<image>\nDescribe this image in detail.'
        }
    ], [
        {
            'role': 'user',
            'content': '<video>\nDescribe this video in detail.'
        }
    ]
]

images = [
    Image.open('assets/example_image.jpg').convert('RGB'),
    load_decord('assets/example_video.mp4', sample_type='uniform', num_frames=32)
]

response = videoccam.chat(messages, images, max_new_tokens=512, do_sample=False)

print(response)