README.md

metadata

license: mit

Bo Li*¹  Yuanhan Zhang*^,1  Liangyu Chen*^,1  Jinghao Wang*^,1  Fanyi Pu*^,1 
Jingkang Yang¹  Chunyuan Li²  Ziwei Liu¹

¹S-Lab, Nanyang Technological University  ²Microsoft Research, Redmond

---

An example of using this model to run on your video. Please first clone Otter to your local disk. Place following script inside the Otter folder to make sure it has the access to otter/modeling_otter.py.

import mimetypes
import os
from typing import Union
import cv2
import requests
import torch
import transformers
from PIL import Image

from otter.modeling_otter import OtterForConditionalGeneration

# Disable warnings
requests.packages.urllib3.disable_warnings()

# ------------------- Utility Functions -------------------


def get_content_type(file_path):
    content_type, _ = mimetypes.guess_type(file_path)
    return content_type


# ------------------- Image and Video Handling Functions -------------------


def extract_frames(video_path, num_frames=128):
    video = cv2.VideoCapture(video_path)
    total_frames = int(video.get(cv2.CAP_PROP_FRAME_COUNT))
    frame_step = total_frames // num_frames
    frames = []

    for i in range(num_frames):
        video.set(cv2.CAP_PROP_POS_FRAMES, i * frame_step)
        ret, frame = video.read()
        if ret:
            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            frame = Image.fromarray(frame).convert("RGB")
            frames.append(frame)

    video.release()
    return frames


def get_image(url: str) -> Union[Image.Image, list]:
    if "://" not in url:  # Local file
        content_type = get_content_type(url)
    else:  # Remote URL
        content_type = requests.head(url, stream=True, verify=False).headers.get(
            "Content-Type"
        )

    if "image" in content_type:
        if "://" not in url:  # Local file
            return Image.open(url)
        else:  # Remote URL
            return Image.open(requests.get(url, stream=True, verify=False).raw)
    elif "video" in content_type:
        video_path = "temp_video.mp4"
        if "://" not in url:  # Local file
            video_path = url
        else:  # Remote URL
            with open(video_path, "wb") as f:
                f.write(requests.get(url, stream=True, verify=False).content)
        frames = extract_frames(video_path)
        if "://" in url:  # Only remove the temporary video file if it was downloaded
            os.remove(video_path)
        return frames
    else:
        raise ValueError("Invalid content type. Expected image or video.")


# ------------------- OTTER Prompt and Response Functions -------------------


def get_formatted_prompt(prompt: str) -> str:
    return f"<image>User: {prompt} GPT:<answer>"


def get_response(input_data, prompt: str, model=None, image_processor=None) -> str:
    if isinstance(input_data, Image.Image):
        vision_x = (
            image_processor.preprocess([input_data], return_tensors="pt")[
                "pixel_values"
            ]
            .unsqueeze(1)
            .unsqueeze(0)
        )
    elif isinstance(input_data, list):  # list of video frames
        vision_x = (
            image_processor.preprocess(input_data, return_tensors="pt")["pixel_values"]
            .unsqueeze(1)
            .unsqueeze(0)
        )
    else:
        raise ValueError(
            "Invalid input data. Expected PIL Image or list of video frames."
        )

    lang_x = model.text_tokenizer(
        [
            get_formatted_prompt(prompt),
        ],
        return_tensors="pt",
    )

    generated_text = model.generate(
        vision_x=vision_x.to(model.device),
        lang_x=lang_x["input_ids"].to(model.device),
        attention_mask=lang_x["attention_mask"].to(model.device),
        max_new_tokens=512,
        num_beams=3,
        no_repeat_ngram_size=3,
    )
    parsed_output = (
        model.text_tokenizer.decode(generated_text[0])
        .split("<answer>")[-1]
        .lstrip()
        .rstrip()
        .split("<|endofchunk|>")[0]
        .lstrip()
        .rstrip()
        .lstrip('"')
        .rstrip('"')
    )
    return parsed_output


if __name__ == "__main__":
    # ------------------- Main Function -------------------
    load_bit = "fp16"
    if load_bit == "fp16":
        precision = {"torch_dtype": torch.float16}
    elif load_bit == "bf16":
        precision = {"torch_dtype": torch.bfloat16}
    elif load_bit == "fp32":
        precision = {"torch_dtype": torch.float32}

    # This model version is trained on MIMIC-IT DC dataset.
    model = OtterForConditionalGeneration.from_pretrained(
        "luodian/otter-9b-dc-hf", device_map="auto", **precision
    )
    model.text_tokenizer.padding_side = "left"
    tokenizer = model.text_tokenizer
    image_processor = transformers.CLIPImageProcessor()
    model.eval()

    while True:
        video_url = "demo.mp4"  # Replace with the path to your video file

        frames_list = get_image(video_url)

        prompts_input = input("Enter prompts (comma-separated): ")
        prompts = [prompt.strip() for prompt in prompts_input.split(",")]

        for prompt in prompts:
            print(f"\nPrompt: {prompt}")
            response = get_response(frames_list, prompt, model, image_processor)
            print(f"Response: {response}")

        if prompts_input.lower() == "quit":
            break