POMA_BENCH/eval_scene_retrieval.py

import os
import glob
import json
import argparse
from typing import Dict, List, Tuple

import torch
import torch.nn as nn
import torch.nn.functional as F
from safetensors.torch import load_file
from transformers import AutoImageProcessor, AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import hf_hub_download
from peft import LoraConfig, get_peft_model

# -----------------------------
# Utils
# -----------------------------
def load_json(path: str) -> dict:
    with open(path, "r") as f:
        return json.load(f)

def find_scan_safetensor(scan_root: str, scan_id: str) -> str:
    direct = os.path.join(scan_root, f"{scan_id}.safetensors")
    if os.path.exists(direct):
        return direct

    pattern = os.path.join(scan_root, "**", f"{scan_id}.safetensors")
    matches = glob.glob(pattern, recursive=True)
    if not matches:
        raise FileNotFoundError(f"Cannot find safetensor for scan_id={scan_id} under {scan_root}")
    matches = sorted(matches, key=len)
    return matches[0]

def to_vchw(point_map: torch.Tensor) -> torch.Tensor:
    """
    Convert point_map to (V, 3, H, W) float tensor.
    Accepts:
      (V, 3, H, W)
      (V, H, W, 3)
    """
    if point_map.dim() != 4:
        raise ValueError(f"Expected 4D point_map, got shape={tuple(point_map.shape)}")

    V, a, b, c = point_map.shape
    if a == 3:
        out = point_map
    elif c == 3:
        out = point_map.permute(0, 3, 1, 2).contiguous()
    else:
        raise ValueError(f"Unrecognized point_map layout: shape={tuple(point_map.shape)}")

    return out.float()

def load_safetensor_from_hf(repo_id, filename, repo_type="dataset"):
    cached_path = hf_hub_download(
        repo_id=repo_id,
        filename=filename,
        repo_type=repo_type,
        local_files_only=False
    )
    return load_file(cached_path)

def load_pretrain(model, pretrain_ckpt_path: str):
    print(f"📂 Loading pretrained weights from: {str(pretrain_ckpt_path)}")

    model_weight_path_pattern = os.path.join(pretrain_ckpt_path, "model*.safetensors")
    model_weight_paths = glob.glob(model_weight_path_pattern)

    if len(model_weight_paths) == 0:
        raise FileNotFoundError(f"❌ Cannot find any model*.safetensors in {str(pretrain_ckpt_path)}")

    weights = {}
    for model_weight_path in model_weight_paths:
        print(f"📥 Loading weights from: {model_weight_path}")
        weights.update(load_file(model_weight_path, device="cpu"))

    result = model.load_state_dict(weights, strict=False)

    model_keys = set(model.state_dict().keys())
    loaded_keys = model_keys.intersection(weights.keys())
    print(f"✅ Loaded keys:      {len(loaded_keys)} / {len(model_keys)}")
    print(f"❌ Missing keys:     {len(result.missing_keys)}")
    print(f"⚠️ Unexpected keys:  {len(result.unexpected_keys)}")


# -----------------------------
# Model wrapper
# -----------------------------
class RepModel(nn.Module):
    def __init__(self, model_root: str = "fg-clip-base"):
        super().__init__()

        self.pm_encoder = AutoModelForCausalLM.from_pretrained(f'../{model_root}', trust_remote_code=True)
        self.tokenizer = AutoTokenizer.from_pretrained(f'../{model_root}', trust_remote_code=True, use_fast=True)
        self.image_processor = AutoImageProcessor.from_pretrained(f'../{model_root}')

        # Optional: print trainable params
        try:
            self.pm_encoder.print_trainable_parameters()
        except Exception:
            pass

    def encode_views_batched(self, pm_vchw: torch.Tensor, batch_views: int = 32) -> torch.Tensor:
        """
        pm_vchw: (V,3,H,W) on device
        returns: (V,D) normalized
        """
        feats_all = []
        V = pm_vchw.shape[0]
        for s in range(0, V, batch_views):
            chunk = pm_vchw[s : s + batch_views]  # (b,3,H,W)
            _, feats = self.pm_encoder.get_image_features(chunk)
            feats = F.normalize(feats.float(), dim=-1)
            feats_all.append(feats)
        return torch.cat(feats_all, dim=0)

    @torch.no_grad()
    def encode_text(self, texts: List[str]) -> torch.Tensor:
        """
        texts: list[str]
        returns: (B,D) normalized
        """
        tok = self.tokenizer(
            texts,
            padding="max_length",
            truncation=True,
            max_length=248,
            return_tensors="pt",
        ).to(next(self.parameters()).device)

        feats = self.pm_encoder.get_text_features(tok["input_ids"], walk_short_pos=False)
        feats = F.normalize(feats.float(), dim=-1)
        return feats

# -----------------------------
# Scene retrieval
# -----------------------------
def build_queries_from_caption_json(caption_json: dict) -> List[dict]:
    """
    Convert:
      { scene_id: { "captions": [c1,c2,...] }, ... }
    into:
      [ { "scene_id": scene_id, "caption": c }, ... ]
    """
    queries = []
    for scene_id, payload in caption_json.items():
        caps = payload.get("captions", [])
        for c in caps:
            c = (c or "").strip()
            if c:
                queries.append({"scene_id": scene_id, "caption": c})
    return queries


@torch.no_grad()
def eval_scene_retrieval(
    model: RepModel,
    caption_json: dict,
    scan_root: str,
    device: str = "cuda",
    batch_views: int = 32,
    recall_ks: Tuple[int, ...] = (1, 5, 10),
) -> Dict[str, float]:
    """
    For each caption, retrieve the correct scene among all scenes in caption_json.
    Scene embedding = mean pooling over view embeddings.
    """
    model.eval().to(device)

    scene_ids = sorted(list(caption_json.keys()))
    if len(scene_ids) == 0:
        return {"n": 0}

    # Cache: scene_id -> pooled scene feature (D,) on CPU
    scene_feat_cache: Dict[str, torch.Tensor] = {}

    # Precompute all scene pooled features once (so retrieval is fast)
    for sid in scene_ids:
        filename = f'light_scannet/{sid}.safetensors'
        data = load_safetensor_from_hf('MatchLab/ScenePoint', filename, repo_type="dataset")

        pm = to_vchw(data["point_map"])  # (V,3,H,W) on CPU
        pm = pm.to(device, non_blocking=True)

        view_feats = model.encode_views_batched(pm, batch_views=batch_views)  # (V,D) on GPU
        scene_feat = view_feats.mean(dim=0)  # (D,)
        scene_feat = F.normalize(scene_feat, dim=-1)

        scene_feat_cache[sid] = scene_feat.detach().cpu()

    # Stack gallery: (N,D)
    gallery = torch.stack([scene_feat_cache[sid] for sid in scene_ids], dim=0)  # CPU
    gallery = gallery.to(device)

    # Build queries
    queries = build_queries_from_caption_json(caption_json)

    total = 0
    top1_correct = 0
    recall_correct = {k: 0 for k in recall_ks}

    for q in queries:
        gt_sid = q["scene_id"]
        caption = q["caption"]

        if gt_sid not in scene_feat_cache:
            continue

        text_feat = model.encode_text([caption])[0]  # (D,) on GPU

        # similarity over all scenes: (N,)
        sims = gallery @ text_feat.unsqueeze(-1)  # (N,1)
        sims = sims.squeeze(-1)

        ranked = torch.argsort(sims, descending=True)  # indices into scene_ids
        pred_sid = scene_ids[int(ranked[0].item())]

        total += 1
        if pred_sid == gt_sid:
            top1_correct += 1

        for k in recall_ks:
            k_eff = min(k, len(scene_ids))
            topk_idx = ranked[:k_eff].tolist()
            topk_sids = [scene_ids[i] for i in topk_idx]
            if gt_sid in topk_sids:
                recall_correct[k] += 1

        # optional debug print
        print(f"[Q] GT={gt_sid} | Pred={pred_sid} | caption={caption[:80]}...")

    if total == 0:
        return {"n": 0}

    out = {"n": total, "top1_acc": top1_correct / total}
    for k in recall_ks:
        out[f"recall@{k}"] = recall_correct[k] / total
    return out


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--caption_json", type=str, required=True, help="JSON mapping scene_id -> {captions:[...]}")
    ap.add_argument("--scan_root", type=str, required=True, help="Root dir containing scene safetensors")
    ap.add_argument("--ckpt", type=str, default="", help="Optional: dir with model*.safetensors")
    ap.add_argument("--model_root", type=str, default="fg-clip-base")
    ap.add_argument("--device", type=str, default="cuda")
    ap.add_argument("--batch_views", type=int, default=32)
    args = ap.parse_args()

    caption_json = load_json(args.caption_json)

    model = RepModel(model_root=args.model_root)
    if args.ckpt:
        load_pretrain(model, args.ckpt)

    metrics = eval_scene_retrieval(
        model=model,
        caption_json=caption_json,
        scan_root=args.scan_root,
        device=args.device,
        batch_views=args.batch_views,
        recall_ks=(1, 5, 10),
    )

    print("\n=== Scene Retrieval Results ===")
    for k, v in metrics.items():
        if isinstance(v, float):
            print(f"{k:>10}: {v:.4f}")
        else:
            print(f"{k:>10}: {v}")


if __name__ == "__main__":
    main()