# examples/swin_export.py """Export genuine native TorchVision Swin-T V1 classification without network. Supply the checkpoint explicitly and licensed local RGB photos with provenance. The native TorchVision backbone is used directly; no custom-backbone parity or original-paper benchmark reproduction is claimed. All 1000 outputs and the exact normalized inputs are preserved. Install only requirements-swin.txt if needed. """ import argparse import hashlib import inspect import json import platform import shutil import sys from pathlib import Path import numpy as np import torch import torchvision import torchvision.models.swin_transformer as native_swin from PIL import Image from PIL import __version__ as pillow_version from torchvision import transforms from torchvision.models import Swin_T_Weights, swin_t CHECKPOINT_URL = "https://download.pytorch.org/models/swin_t-704ceda3.pth" CHECKPOINT_SHA256 = "704ceda373461b0a224fcdddd75cd2a5e9f8064512ed47adbddef7f343fd147b" ORIGINAL_REVISION = "f82860bfb5225915aca09c3227159ee9e1df874d" def digest(path): """Hash file bytes without buffering a large checkpoint.""" with Path(path).open("rb") as stream: return hashlib.file_digest(stream, "sha256").hexdigest() def export(checkpoint, images_dir, photo_sources, output_dir): """Run the native full classifier and preserve exact input/output evidence.""" if digest(checkpoint) != CHECKPOINT_SHA256: raise ValueError("Swin-T V1 checkpoint checksum mismatch") torch.set_num_threads(2) torch.manual_seed(13) weights = Swin_T_Weights.IMAGENET1K_V1 model = swin_t(weights=None).eval() model.load_state_dict( torch.load(checkpoint, map_location="cpu", weights_only=True), strict=True ) preprocessing = weights.transforms() records = json.loads(Path(photo_sources).read_text()) output_dir = Path(output_dir) output_dir.mkdir(parents=True, exist_ok=True) source_module = Path(inspect.getfile(native_swin)) presets_module = Path(inspect.getfile(type(preprocessing))) functional_module = Path(inspect.getfile(transforms.functional)) pillow_module = Path(inspect.getfile(transforms.functional.F_pil)) package = Path(torchvision.__file__).parent native_sources = { str(path.relative_to(package)): digest(path) for path in (source_module, presets_module, functional_module, pillow_module) } program = Path(__file__).with_name("swin_transformer_training.py") original_manifest = Path(__file__).with_name("swin-original-source-hashes.json") result = { "provenance": { "model": "torchvision.models.swin_t (Swin V1, not Swin V2)", "architecture": { "patchSize": 4, "windowSize": 7, "width": 96, "depths": [2, 2, 6, 2], "heads": [3, 6, 12, 24], }, "weights": "Swin_T_Weights.IMAGENET1K_V1", "checkpointUrl": CHECKPOINT_URL, "checkpointSha256": digest(checkpoint), "parameterCount": sum(p.numel() for p in model.parameters()), "strictStateDict": True, "device": "cpu", "dtype": "float32", "python": platform.python_version(), "torch": torch.__version__, "torchvision": torchvision.__version__, "numpy": np.__version__, "pillow": pillow_version, "torchvisionRevision": torchvision.version.git_version, "torchvisionSourceHashes": native_sources, "originalMicrosoftRevision": ORIGINAL_REVISION, "originalSourceManifestSha256": digest(original_manifest), "programSha256": digest(program), "exportSha256": digest(__file__), "photoSourcesSha256": digest(photo_sources), "command": [sys.executable, *sys.argv], "trainingRecipe": weights.meta["recipe"], "publishedTorchvisionTop1": weights.meta["_metrics"]["ImageNet-1K"][ "acc@1" ], "originalPaperSwinTTop1": 81.3, "scope": ( "Two real-photo native inferences; no ImageNet evaluation " "or original benchmark reproduction" ), }, "preprocessing": { "resizeSize": preprocessing.resize_size, "cropSize": preprocessing.crop_size, "interpolation": preprocessing.interpolation.value, "antialias": preprocessing.antialias, "mean": preprocessing.mean, "std": preprocessing.std, "rescale": "PIL RGB uint8 to float32 [0,1], then (x-mean)/std", }, "categories": weights.meta["categories"], "photos": [], } for photo in records: source = Path(images_dir) / photo["originalImage"] if digest(source) != photo["source"]["sha256"]: raise ValueError(f"Photo checksum mismatch: {source}") image = Image.open(source).convert("RGB") tensor = preprocessing(image).unsqueeze(0) processed = transforms.functional.center_crop( transforms.functional.resize( image, preprocessing.resize_size, interpolation=preprocessing.interpolation, antialias=preprocessing.antialias, ), preprocessing.crop_size, ) # Authenticate the visual crop against the exact normalized model input. explicit = transforms.functional.normalize( transforms.functional.pil_to_tensor(processed).float() / 255, preprocessing.mean, preprocessing.std, ).unsqueeze(0) if not torch.equal(tensor, explicit): raise AssertionError("Processed photo differs from native input") with torch.inference_mode(): logits = model(tensor)[0] probabilities = logits.softmax(0) scores, indices = probabilities.topk(5) stem = photo["id"] processed_name, arrays_name = f"input-{stem}.png", f"arrays-{stem}.npz" processed.save(output_dir / processed_name) shutil.copyfile(source, output_dir / photo["originalImage"]) np.savez_compressed( output_dir / arrays_name, normalized_input=tensor.numpy(), logits=logits.numpy(), probabilities=probabilities.numpy(), ) result["photos"].append( { **photo, "originalShape": [image.height, image.width], "inputImage": processed_name, "inputShape": list(tensor.shape), "inputImageSha256": digest(output_dir / processed_name), "normalizedInputSha256": hashlib.sha256( tensor.numpy().tobytes() ).hexdigest(), "processedInputParityMaxError": float((explicit - tensor).abs().max()), "arrays": arrays_name, "arraysSha256": digest(output_dir / arrays_name), "logits": logits.tolist(), "probabilities": probabilities.tolist(), "top5": [ { "index": int(index), "label": weights.meta["categories"][index], "score": float(score), } for score, index in zip(scores, indices) ], } ) (output_dir / "predictions.json").write_text(json.dumps(result, indent=2) + "\n") return result def main(): """Read explicit local inputs, never download model weights or photos.""" parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--checkpoint", required=True, type=Path) parser.add_argument("--images-dir", required=True, type=Path) parser.add_argument( "--photo-sources", type=Path, default=Path(__file__).with_name("swin-photo-sources.json"), ) parser.add_argument("--output-dir", required=True, type=Path) args = parser.parse_args() result = export( args.checkpoint, args.images_dir, args.photo_sources, args.output_dir ) print( json.dumps( { "photos": len(result["photos"]), "parameterCount": result["provenance"]["parameterCount"], } ) ) if __name__ == "__main__": main()