app.py

import os, json, traceback
from typing import Any, Dict, Tuple
import gradio as gr
from PIL import Image
import torch
from transformers import AutoProcessor, AutoTokenizer, AutoModelForCausalLM, AutoConfig

# -------- Env / params --------
MODEL_ID = os.environ.get("MODEL_ID", "inference-net/ClipTagger-12b")
HF_TOKEN = os.environ.get("HF_TOKEN")
TEMP = 0.1
MAX_NEW_TOKENS = 768  # faster demo; raise later if needed
DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32

# -------- Prompts (yours) --------
SYSTEM_PROMPT = (
    "You are an image annotation API trained to analyze YouTube video keyframes. "
    "You will be given instructions on the output format, what to caption, and how to perform your job. "
    "Follow those instructions. For descriptions and summaries, provide them directly and do not lead them with "
    "'This image shows' or 'This keyframe displays...', just get right into the details."
)

USER_PROMPT = """You are an image annotation API trained to analyze YouTube video keyframes. You must respond with a valid JSON object matching the exact structure below.

Your job is to extract detailed **factual elements directly visible** in the image. Do not speculate or interpret artistic intent, camera focus, or composition. Do not include phrases like "this appears to be", "this looks like", or anything about the image itself. Describe what **is physically present in the frame**, and nothing more.

Return JSON in this structure:

{
    "description": "A detailed, factual account of what is visibly happening (4 sentences max). Only mention concrete elements or actions that are clearly shown. Do not include anything about how the image is styled, shot, or composed. Do not lead the description with something like 'This image shows' or 'this keyframe is...', just get right into the details.",
    "objects": ["object1 with relevant visual details", "object2 with relevant visual details", ...],
    "actions": ["action1 with participants and context", "action2 with participants and context", ...],
    "environment": "Detailed factual description of the setting and atmosphere based on visible cues (e.g., interior of a classroom with fluorescent lighting, or outdoor forest path with snow-covered trees).",
    "content_type": "The type of content it is, e.g. 'real-world footage', 'video game', 'animation', 'cartoon', 'CGI', 'VTuber', etc.",
    "specific_style": "Specific genre, aesthetic, or platform style (e.g., anime, 3D animation, mobile gameplay, vlog, tutorial, news broadcast, etc.)",
    "production_quality": "Visible production level: e.g., 'professional studio', 'amateur handheld', 'webcam recording', 'TV broadcast', etc.",
    "summary": "One clear, comprehensive sentence summarizing the visual content of the frame. Like the description, get right to the point.",
    "logos": ["logo1 with visual description", "logo2 with visual description", ...]
}

Rules:
- Be specific and literal. Focus on what is explicitly visible.
- Do NOT include interpretations of emotion, mood, or narrative unless it's visually explicit.
- No artistic or cinematic analysis.
- Always include the language of any text in the image if present as an object, e.g. "English text", "Japanese text", "Russian text", etc.
- Maximum 10 objects and 5 actions.
- Return an empty array for 'logos' if none are present.
- Always output strictly valid JSON with proper escaping.
- Output **only the JSON**, no extra text or explanation.
"""

# -------- Utils --------
def extract_top_level_json(s: str):
    """Parse JSON; if extra text around it, extract the first balanced {...} block."""
    # Fast path
    try:
        return json.loads(s)
    except Exception:
        pass
    # Brace-stack extraction
    start = None
    depth = 0
    for i, ch in enumerate(s):
        if ch == '{':
            if depth == 0:
                start = i
            depth += 1
        elif ch == '}':
            if depth > 0:
                depth -= 1
                if depth == 0 and start is not None:
                    chunk = s[start:i+1]
                    try:
                        return json.loads(chunk)
                    except Exception:
                        # continue scanning for the next candidate
                        start = None
    return None

def build_messages(image: Image.Image):
    return [
        {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
        {"role": "user",   "content": [{"type": "image", "image": image},
                                       {"type": "text",  "text": USER_PROMPT}]}
    ]

def downscale_if_huge(pil: Image.Image, max_side: int = 1792) -> Image.Image:
    if pil is None:
        return pil
    w, h = pil.size
    m = max(w, h)
    if m <= max_side:
        return pil.convert("RGB")
    s = max_side / m
    return pil.convert("RGB").resize((int(w*s), int(h*s)), Image.BICUBIC)

# -------- Load model (A100) --------
processor = tokenizer = model = None
LOAD_ERROR = None

try:
    cfg = AutoConfig.from_pretrained(MODEL_ID, token=HF_TOKEN, trust_remote_code=True)
    if "clip" in cfg.__class__.__name__.lower():
        raise RuntimeError(f"MODEL_ID '{MODEL_ID}' is a CLIP/encoder repo; need a causal VLM.")

    print("[boot] loading processor…", flush=True)
    try:
        processor = AutoProcessor.from_pretrained(
            MODEL_ID, token=HF_TOKEN, trust_remote_code=True, use_fast=True
        )
    except TypeError:
        processor = AutoProcessor.from_pretrained(
            MODEL_ID, token=HF_TOKEN, trust_remote_code=True
        )

    print("[boot] loading model…", flush=True)
    # Force full-precision path on A100 first; add quantized path later if desired
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        token=HF_TOKEN,
        device_map="auto",
        torch_dtype=DTYPE,
        trust_remote_code=True,
        # quantization_config=None,  # keep commented if you want to honor repo quant; uncomment to force dequant
    )

    tokenizer = getattr(processor, "tokenizer", None) or AutoTokenizer.from_pretrained(
        MODEL_ID, token=HF_TOKEN, trust_remote_code=True, use_fast=True
    )
    print("[boot] ready.", flush=True)

except Exception as e:
    LOAD_ERROR = f"{e}\n\n{traceback.format_exc()}"

# -------- Inference --------
def generate(image: Image.Image) -> Tuple[str, Dict[str, Any] | None, bool]:
    if image is None:
        return "Please upload an image.", None, False
    if model is None or processor is None:
        return f"❌ Load error:\n{LOAD_ERROR}", None, False

    image = downscale_if_huge(image)

    # Build prompt
    if hasattr(processor, "apply_chat_template"):
        prompt = processor.apply_chat_template(build_messages(image), add_generation_prompt=True, tokenize=False)
    else:
        # fallback join (rare)
        prompt = USER_PROMPT

    # Tokenize with vision
    inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)

    # Common gen kwargs
    eos = getattr(model.config, "eos_token_id", None)

    tried = []

    # (1) Greedy, no sampling (most stable, no temperature arg)
    try:
        g = dict(do_sample=False, max_new_tokens=MAX_NEW_TOKENS)
        if eos is not None:
            g["eos_token_id"] = eos
        with torch.inference_mode():
            out = model.generate(**inputs, **g)
        text = (processor.decode(out[0], skip_special_tokens=True)
                if hasattr(processor, "decode")
                else tokenizer.decode(out[0], skip_special_tokens=True))
        parsed = extract_top_level_json(text)
        if isinstance(parsed, dict):
            return json.dumps(parsed, indent=2), parsed, True
        tried.append(("greedy", "parsed-failed"))
    except Exception as e:
        tried.append(("greedy", f"err={e}"))

    # (2) Sampling with temperature=0.1
    try:
        g = dict(do_sample=True, temperature=TEMP, max_new_tokens=MAX_NEW_TOKENS)
        if eos is not None:
            g["eos_token_id"] = eos
        with torch.inference_mode():
            out = model.generate(**inputs, **g)
        text = (processor.decode(out[0], skip_special_tokens=True)
                if hasattr(processor, "decode")
                else tokenizer.decode(out[0], skip_special_tokens=True))
        parsed = extract_top_level_json(text)
        if isinstance(parsed, dict):
            return json.dumps(parsed, indent=2), parsed, True
        tried.append(("sample_t0.1", "parsed-failed"))
    except Exception as e:
        tried.append(("sample_t0.1", f"err={e}"))

    # (3) Shorter greedy
    try:
        g = dict(do_sample=False, max_new_tokens=min(512, MAX_NEW_TOKENS))
        if eos is not None:
            g["eos_token_id"] = eos
        with torch.inference_mode():
            out = model.generate(**inputs, **g)
        text = (processor.decode(out[0], skip_special_tokens=True)
                if hasattr(processor, "decode")
                else tokenizer.decode(out[0], skip_special_tokens=True))
        parsed = extract_top_level_json(text)
        if isinstance(parsed, dict):
            return json.dumps(parsed, indent=2), parsed, True
        tried.append(("greedy_short", "parsed-failed"))
    except Exception as e:
        tried.append(("greedy_short", f"err={e}"))

    # Debug info if all fail
    return "Generation failed.\nTried: " + "\n".join([f"{t[0]} -> {t[1]}" for t in tried]), None, False

# -------- UI --------
with gr.Blocks(theme=gr.themes.Soft(), analytics_enabled=False, title="Keyframe Annotator (Gemma-3 VLM · A100)") as demo:
    gr.Markdown("# Keyframe Annotator (Gemma-3-12B FT · A100)\nUpload an image to get **strict JSON** annotations.")
    if LOAD_ERROR:
        with gr.Accordion("Startup Error Details", open=False):
            gr.Markdown(f"```\n{LOAD_ERROR}\n```")

    with gr.Row():
        with gr.Column(scale=1):
            image = gr.Image(type="pil", label="Upload Image", image_mode="RGB")
            btn = gr.Button("Annotate", variant="primary")
        with gr.Column(scale=1):
            out_text = gr.Code(label="Output (JSON or error)")
            out_json = gr.JSON(label="Parsed JSON")
            ok_flag = gr.Checkbox(label="Valid JSON", value=False, interactive=False)

    btn.click(generate, inputs=[image], outputs=[out_text, out_json, ok_flag])

demo.queue(max_size=32).launch()