import argparse from pathlib import Path import runpy from huggingface_hub import snapshot_download from PIL import Image import torch from torchvision import transforms BUNDLE = Path(__file__).absolute().parents[1] vision = runpy.run_path(str(BUNDLE / "inkling2_vision.py")) MODEL_ID = "immanuelpeter/Inkling-2-Small-Vision" RESOLUTION = 640 def main() -> None: parser = argparse.ArgumentParser(description="Extract Inkling 2 Small visual features.") parser.add_argument("image", type=Path) parser.add_argument("--model", default=MODEL_ID) args = parser.parse_args() bundle = Path(args.model) if Path(args.model).is_dir() else Path(snapshot_download(args.model)) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") dtype = torch.bfloat16 if device.type == "cuda" else torch.float32 tower, projector = vision["load_vision"](bundle, dtype) tower, projector = tower.to(device), projector.to(device) crop = transforms.Compose([ transforms.Resize(RESOLUTION, interpolation=transforms.InterpolationMode.BICUBIC), transforms.CenterCrop(RESOLUTION), transforms.ToTensor(), ]) pixels = vision["normalize"](crop(Image.open(args.image).convert("RGB")))[None].to(device, dtype) with torch.inference_mode(): tokens = tower.forward_features(pixels)[:, 1:] merged = vision["merge"](tokens, RESOLUTION // vision["PATCH_SIZE"]) projected = projector(merged) print("tower", tuple(tokens.shape)) print("merged", tuple(merged.shape)) print("projected", tuple(projected.shape)) if __name__ == "__main__": main()