run_custom_model.py

#!/usr/bin/env python3
"""
Run BEAR inference with ANY local model — no VLMEvalKit required.

Instead of VLMEvalKit's `supported_VLM`, this runner calls a tiny pluggable
adapter that you point at with `--model_impl module:ClassName`. An adapter is
just a class with:

    class MyModel:
        def __init__(self, model_name, **kwargs): ...
        def generate(self, text: str, images: list[PIL.Image.Image]) -> str: ...

The runner prepares, per question:
  * single-image tasks (pointing / bbox / trajectory): images = [the image]
  * video tasks       : images = 16 sampled frames
  * interleaved tasks : images = 16 sampled frames + the observation image
and a text prompt, then calls `adapter.generate(text, images)`.

Ready-made adapters live in bear_models.py (Cosmos, generic Qwen2.5-VL, Echo).

Examples:
    # NVIDIA Cosmos-Reason1-7B
    cd task_planning
    python ../run_custom_model.py --model_impl bear_models:CosmosReason1 \
        --input_json_path next_action_prediction_official.json

    # smoke test with the dependency-free Echo adapter
    python ../run_custom_model.py --model_impl bear_models:EchoModel \
        --input_json_path next_action_prediction_official.json
"""
import os
import sys
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

import json
import argparse
import importlib

import numpy as np
from PIL import Image

from util.prompt_generation import generate_question_prompt

SAMPLE_FRAMES = 16


def sample_frames(mp4_path, num_frames=SAMPLE_FRAMES):
    """Return `num_frames` evenly-spaced frames from a video as PIL images."""
    import cv2
    cap = cv2.VideoCapture(mp4_path)
    if not cap.isOpened():
        raise ValueError(f"Cannot open video: {mp4_path}")
    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    idx = np.linspace(0, max(total - 1, 0), num=num_frames, dtype=int)
    frames = []
    for i in idx:
        cap.set(cv2.CAP_PROP_POS_FRAMES, int(i))
        ok, frame = cap.read()
        if ok:
            frames.append(Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)))
    cap.release()
    return frames


def make_input(category, item, fmt, base):
    """Build (text, images) for one item / one prompt format ('direct' or 'cot')."""
    question = item.get("question", "")
    options = item.get("options", [])
    video = (item.get("video") or "").strip()
    image = (item.get("image") or "").strip()

    prompt = generate_question_prompt(fmt, category, question, options, model_category="general")

    images = []
    if isinstance(prompt, tuple):
        text = "".join(str(p) for p in prompt)
        if video:
            images += sample_frames(os.path.join(base, video))
        # 3-part prompts (relative direction / path planning) append the observation image
        if len(prompt) == 3 and image:
            images.append(Image.open(os.path.join(base, image)).convert("RGB"))
    else:
        text = prompt
        if image:
            images = [Image.open(os.path.join(base, image)).convert("RGB")]
        elif video:
            images = sample_frames(os.path.join(base, video))
    return text, images


def load_adapter(spec, model_name):
    """spec = 'module:ClassName' -> instantiated adapter."""
    if ":" not in spec:
        raise ValueError("--model_impl must be 'module:ClassName', e.g. bear_models:CosmosReason1")
    mod_name, cls_name = spec.split(":", 1)
    cls = getattr(importlib.import_module(mod_name), cls_name)
    return cls(model_name) if model_name else cls()


ROUTE = {
    "image": ["pointing", "trajectory", "bbox"],
    "interleaved": ["path planning", "relative direction"],
    "video": ["object localization", "next action prediction", "task progress reasoning"],
}


if __name__ == "__main__":
    ap = argparse.ArgumentParser(description="BEAR inference with a custom local model (no VLMEvalKit).")
    ap.add_argument("--model_impl", default="bear_models:CosmosReason1",
                    help="Adapter as 'module:ClassName' (default: bear_models:CosmosReason1).")
    ap.add_argument("--model_name", default="",
                    help="Passed to the adapter (e.g. HF id). Blank = adapter default.")
    ap.add_argument("--input_json_path", required=True, help="Task JSON file.")
    ap.add_argument("--formats", default="direct,cot",
                    help="Comma-separated prompt formats to run (default: direct,cot).")
    ap.add_argument("--evaluate_output_category", default=None, help="Output tag. Default: input filename stem.")
    args = ap.parse_args()

    base = os.path.dirname(os.path.abspath(args.input_json_path))
    tag = args.evaluate_output_category or os.path.splitext(os.path.basename(args.input_json_path))[0]
    formats = [f.strip() for f in args.formats.split(",") if f.strip()]

    model = load_adapter(args.model_impl, args.model_name)
    model_label = args.model_name or getattr(model, "model_name", type(model).__name__)

    with open(args.input_json_path) as f:
        data = json.load(f)

    out = []
    for i, item in enumerate(data):
        category = item.get("category", "")
        item_copy = item.copy()
        for fmt in formats:
            try:
                text, images = make_input(category, item, fmt, base)
                item_copy[f"{fmt}_reply"] = model.generate(text, images)
            except Exception as e:
                item_copy[f"{fmt}_reply"] = f"error: {e}"
        out.append(item_copy)
        print(f"[{i + 1}/{len(data)}] {item.get('idx', i)} done")

    out_path = f"final_{model_label.replace('/', '_')}_evaluate_{tag}.json"
    with open(out_path, "w") as f:
        json.dump(out, f, indent=4, ensure_ascii=False)
    print(f"Saved -> {out_path}")