| """Demo: run the packaged VesselCNN weights over the packaged sample |
| volumes (assets/samples/) and the first misclassified test volumes, |
| printing the predictions exactly as measured — errors included. |
| |
| The sample volumes were selected from the test evaluation by a fixed |
| rule (see assets/samples/index.json): a false negative and a false |
| positive are in the set on purpose. A demo that only shows correct |
| predictions would misrepresent a model whose aneurysm recall is the |
| number that matters. |
| |
| Usage: |
| python examples/predict_demo.py [--weights model.safetensors] |
| [--test-data assets/vesselmnist3d_test.npz] |
| """ |
| import argparse |
| import glob |
| import json |
| import os |
| import sys |
|
|
| import numpy as np |
|
|
| sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "src")) |
|
|
| from vesselcnn.infer import LABELS, load_model, predict |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| PKG = os.path.dirname(HERE) |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--weights", default=os.path.join(PKG, "model.safetensors")) |
| ap.add_argument("--test-data", default=os.path.join(PKG, "assets", "vesselmnist3d_test.npz")) |
| args = ap.parse_args() |
|
|
| model = load_model(args.weights) |
|
|
| print("== packaged samples (assets/samples/) ==") |
| index = json.load(open(os.path.join(PKG, "assets", "samples", "index.json"))) |
| truth = {s["file"]: s for s in index["samples"]} |
| for path in sorted(glob.glob(os.path.join(PKG, "assets", "samples", "*.npy"))): |
| name = os.path.basename(path) |
| r = predict(model, path) |
| t = truth[name] |
| mark = "OK " if r["label"] == t["label"] else "ERR" |
| print(f"[{mark}] {name}: true={t['label']} predicted={r['label']} " |
| f"(confidence {r['confidence']:.3f}, role: {t['role']})") |
|
|
| if os.path.exists(args.test_data): |
| print("\n== first misclassified test volumes ==") |
| import torch |
| from vesselcnn.infer import TRAIN_MEAN, TRAIN_STD |
| d = np.load(args.test_data) |
| x, y = d["test_images"], d["test_labels"].ravel().astype(np.int64) |
| shown = 0 |
| with torch.no_grad(): |
| for i in range(len(x)): |
| r = predict(model, x[i]) |
| if r["label"] != LABELS[int(y[i])] and shown < 5: |
| print(f"test[{i}]: true={LABELS[int(y[i])]} predicted={r['label']} " |
| f"(confidence {r['confidence']:.3f}, " |
| f"occupancy {(x[i] > 0).sum()} voxels)") |
| shown += 1 |
| print(f"(first {shown} errors shown; the full error inventory is in " |
| "assets/evaluation_results.json)") |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|
