"""Run a verified trained GazeLIFT adapter on one image and fixation CSV.""" from __future__ import annotations import csv import hashlib import json import os import sys import logging from functools import lru_cache from pathlib import Path import numpy as np from PIL import Image import torch import torch.nn.functional as F # CPU preprocessing uses the Space's allocated cores. torch.set_num_threads(max(1, int(os.getenv("GAZELIFT_CPU_THREADS", "2")))) ROOT = Path(__file__).resolve().parents[1] sys.path.insert(0, str(ROOT / "model_code")) from heatmap import make_image_heatmap from model import TransposeCNNGazePyramidDecoder PRESETS = {"colonoscopy": "kvasir", "mri": "prostate"} def _manifest(): return json.loads((ROOT / "weights" / "manifest.json").read_text(encoding="utf-8")) def _checkpoint(dataset: str) -> tuple[Path, dict]: record = _manifest()[dataset] path = Path(os.getenv(record["environment_variable"], ROOT / "weights" / record["filename"])) if not path.is_file(): repository = os.getenv("GAZELIFT_WEIGHTS_REPO") if not repository: raise FileNotFoundError("The verified trained adapter is unavailable. Add it to weights/ or set GAZELIFT_WEIGHTS_REPO.") from huggingface_hub import hf_hub_download path = Path(hf_hub_download(repository, record["filename"], token=os.getenv("HF_TOKEN"))) expected = record["sha256"] if expected.startswith("REPLACE_"): raise ValueError("Set the verified safetensors SHA-256 in weights/manifest.json before inference.") if hashlib.sha256(path.read_bytes()).hexdigest() != expected: raise ValueError("Checkpoint hash does not match the selected GazeLIFT adapter.") return path, record @lru_cache(maxsize=2) def _adapter(dataset: str, device: str): path, record = _checkpoint(dataset) if path.suffix == ".safetensors": from safetensors.torch import load_file state = load_file(str(path), device="cpu") else: state = torch.load(path, map_location="cpu", weights_only=True)["model"] model = TransposeCNNGazePyramidDecoder() model.load_state_dict(state, strict=True) return model.eval().requires_grad_(False).to(device), record @lru_cache(maxsize=1) def _encoder(device: str): import timm checkpoint = os.getenv("GAZELIFT_DINO_CHECKPOINT") model = timm.create_model("vit_large_patch16_dinov3.lvd1689m", pretrained=not bool(checkpoint), num_classes=0, img_size=512, dynamic_img_size=True) if checkpoint: timm.models.load_checkpoint(model, checkpoint, strict=True) return model.eval().requires_grad_(False).to(device) def preload_for_zerogpu(): """Place available models on emulated CUDA during ZeroGPU startup. The real GPU is allocated only by the UI's decorated segmentation callback. Missing adapters leave the UI available and are rejected again at inference. """ if os.getenv("SPACES_ZERO_GPU", "").lower() not in ("true", "1", "t"): return loaded = False for dataset in ("kvasir", "prostate"): try: _adapter(dataset, "cuda") loaded = True except Exception as error: logging.getLogger(__name__).warning( "Could not preload the %s adapter (%s). Inference requires its verified trained weights.", dataset, type(error).__name__, ) if loaded: _encoder("cuda") def _load_points(path: str | Path, size: tuple[int, int]) -> np.ndarray: with open(path, encoding="utf-8-sig", newline="") as stream: rows = list(csv.DictReader(stream)) if not rows: raise ValueError("The fixation CSV is empty.") key = {item.lower(): item for item in rows[0]} x, y = key.get("x", key.get("current_fix_x")), key.get("y", key.get("current_fix_y")) duration = key.get("duration", key.get("current_fix_duration")) if not x or not y: raise ValueError("Fixation CSV needs x/y or CURRENT_FIX_X/CURRENT_FIX_Y columns.") values = np.asarray([[float(row[x]), float(row[y]), float(row[duration]) if duration else 1] for row in rows], np.float32) if not np.isfinite(values).all() or (values[:, 2] < 0).any(): raise ValueError("Fixations must be finite and have nonnegative durations.") w, h = size # The UI always writes native-pixel coordinates here. Never infer units # from their magnitude: a real click near (0, 1) is still in pixels. values[:, 0] = np.clip(values[:, 0], 0, w - 1) values[:, 1] = np.clip(values[:, 1], 0, h - 1) return values def _overlay_heatmap(image: Image.Image, heatmap: np.ndarray) -> Image.Image: heat = Image.fromarray(np.uint8(np.clip(heatmap, 0, 1) * 255)).resize(image.size, Image.Resampling.BILINEAR) hue = np.zeros((*np.asarray(heat).shape, 3), dtype=np.uint8) hue[..., 0], hue[..., 1] = np.asarray(heat), np.asarray(heat) // 2 return Image.blend(image.convert("RGB"), Image.fromarray(hue), 0.45) @torch.inference_mode() def predict(image_path, fixation_csv, preset="colonoscopy", threshold=0.5, device=None, return_all=False): if preset not in PRESETS: raise ValueError(f"Unknown configuration: {preset}") image = image_path.convert("RGB") if isinstance(image_path, Image.Image) else Image.open(image_path).convert("RGB") values = _load_points(fixation_csv, image.size) dataset = PRESETS[preset] zero_gpu = os.getenv("SPACES_ZERO_GPU", "").lower() in ("true", "1", "t") if zero_gpu and device is not None and str(device) != "cuda": raise ValueError("The ZeroGPU deployment requires device='cuda'.") runtime = "cuda" if zero_gpu else str(device or ("cuda" if torch.cuda.is_available() else "cpu")) adapter, record = _adapter(dataset, runtime) xy = values[:, :2] / np.array([max(image.width - 1, 1), max(image.height - 1, 1)], np.float32) duration = values[:, 2] duration = (duration - duration.min()) / (duration.max() - duration.min() + 1e-8) if duration.max() > duration.min() else np.ones_like(duration) heatmap = make_image_heatmap(xy, duration, image_size=(512, 512), sigma_px=record["sigma_px"])[None].to(runtime) reduced = image.resize((512, 512), Image.Resampling.BILINEAR) tensor = torch.from_numpy(np.asarray(reduced).copy()).permute(2, 0, 1).float() / 255 tensor = (tensor - torch.tensor([.485, .456, .406])[:, None, None]) / torch.tensor([.229, .224, .225])[:, None, None] encoder = _encoder(runtime) captured, hooks = {}, [] for index in [5, 11, 17, 23]: hooks.append(encoder.blocks[index].register_forward_hook(lambda module, args, output, index=index: captured.update({index: output[:, encoder.num_prefix_tokens:]}))) try: encoder(tensor[None].to(runtime)) finally: for hook in hooks: hook.remove() features = [captured[index].transpose(1, 2).reshape(1, 1024, 32, 32).half().float() for index in [5, 11, 17, 23]] logits = adapter(tensor[None].half().float().to(runtime), features, heatmap) probability = F.interpolate(logits.sigmoid(), size=(image.height, image.width), mode="bilinear", align_corners=False)[0, 0].cpu().numpy() mask = probability > threshold overlay = np.asarray(image).copy() overlay[mask] = (0.55 * overlay[mask] + 0.45 * np.array([0, 205, 235])).astype(np.uint8) result = {"mask": Image.fromarray((255 * mask).astype(np.uint8)), "mask_overlay": Image.fromarray(overlay), "gaze_overlay": _overlay_heatmap(image, heatmap[0, 0].cpu().numpy()), "preds": probability, "gaze_heatmap": heatmap[0, 0].cpu().numpy()} return result if return_all else result["mask"]