| """Export PET and SUVR embeddings from Stage 1 checkpoints for Figure 2.""" |
| from __future__ import annotations |
|
|
| import argparse |
| from pathlib import Path |
|
|
| import numpy as np |
| import torch |
| from torch.utils.data import DataLoader |
|
|
| from pet_vlm_dataset import PETSUVRDataset, collate_pet_suvr |
| from train_pet_foundation import PETSUVRFoundationModel, build_encoder |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser() |
| parser.add_argument("--checkpoint", type=Path, required=True) |
| parser.add_argument("--split", type=Path, default=Path("data/metadata/splits/test.csv")) |
| parser.add_argument("--backbone", default=None) |
| parser.add_argument("--medicalnet-weights", type=Path, default=Path("pretrained/medicalnet/resnet_50_23dataset.pth")) |
| parser.add_argument("--brainfm-weights", type=Path, default=Path("pretrained/brainfm/assets/brainfm_pretrained.pth")) |
| parser.add_argument("--brainfm-code-root", type=Path, default=Path("pretrained/brainfm")) |
| parser.add_argument("--batch-size", type=int, default=4) |
| parser.add_argument("--num-workers", type=int, default=2) |
| parser.add_argument("--output-size", type=int, nargs=3, default=None) |
| parser.add_argument("--embed-dim", type=int, default=None) |
| parser.add_argument("--freeze-encoder", type=bool, default=None) |
| parser.add_argument("--out", type=Path, required=True) |
| args = parser.parse_args() |
|
|
| ckpt = torch.load(args.checkpoint, map_location="cpu", weights_only=False) |
| saved_args = ckpt.get("args", {}) |
| for name in ("backbone", "embed_dim", "freeze_encoder"): |
| if getattr(args, name, None) is None and name in saved_args: |
| setattr(args, name, saved_args[name]) |
| if args.output_size is None: |
| args.output_size = tuple(saved_args.get("output_size", (96, 96, 96))) |
|
|
| device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
| dataset = PETSUVRDataset(args.split, output_size=tuple(args.output_size)) |
| loader = DataLoader(dataset, batch_size=args.batch_size, shuffle=False, |
| num_workers=args.num_workers, collate_fn=collate_pet_suvr) |
| n_regions = int(dataset[0]["suvr"].numel()) |
|
|
| encoder = build_encoder(args) |
| model = PETSUVRFoundationModel(encoder, n_regions, args.embed_dim or 256, bool(args.freeze_encoder)).to(device) |
| model.load_state_dict(ckpt["model"], strict=True) |
| model.eval() |
|
|
| pet_embeddings = [] |
| suvr_embeddings = [] |
| pred_suvr = [] |
| true_suvr = [] |
|
|
| with torch.no_grad(): |
| for batch in loader: |
| image = batch["image"].to(device) |
| suvr = batch["suvr"].to(device) |
| outputs = model(image, suvr) |
|
|
| pet_feat = model.pet_encoder(image) |
| pet_z = torch.nn.functional.normalize(model.pet_projector(pet_feat), dim=-1) |
| suvr_z = torch.nn.functional.normalize(model.suvr_encoder(suvr), dim=-1) |
|
|
| pet_embeddings.append(pet_z.cpu().numpy()) |
| suvr_embeddings.append(suvr_z.cpu().numpy()) |
| pred_suvr.append(outputs["pred_suvr"].cpu().numpy()) |
| true_suvr.append(suvr.cpu().numpy()) |
|
|
| pet_z = np.concatenate(pet_embeddings, axis=0) |
| suvr_z = np.concatenate(suvr_embeddings, axis=0) |
| pred = np.concatenate(pred_suvr, axis=0) |
| true = np.concatenate(true_suvr, axis=0) |
|
|
| args.out.parent.mkdir(parents=True, exist_ok=True) |
| np.savez(args.out, pet_z=pet_z, suvr_z=suvr_z, pred_suvr=pred, true_suvr=true) |
| print(f"wrote {pet_z.shape[0]} samples, pet_z={pet_z.shape}, suvr_z={suvr_z.shape} to {args.out}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
