Section 1 · Lesson 7

Inference, ensembling, and submission

At prediction time the project averages probabilities over flipped copies and over models, applies the tuned rules, and writes each mask as an RLE string.

Test-time augmentation

Test-time augmentation (TTA) predicts on a few altered copies of the same image and averages the results. predict_probs in src/train.py uses the original image and two flips. It flips each prediction back before averaging:

p = torch.sigmoid(model(img))
if tta:
    p = p + torch.sigmoid(model(img.flip(-1))).flip(-1)   # horizontal flip
    p = p + torch.sigmoid(model(img.flip(-2))).flip(-2)   # vertical flip
    p = p / 3
out.append((p * 255).round().byte().cpu().numpy())

The model saw flipped images in training, so the flipped views are valid inputs. Averaging smooths out view-specific mistakes. On the baseline, TTA lifts the holdout score at tau 0.3 from 0.6284 to 0.6420.

Ensembling

An ensemble averages the probabilities of several models, each of which already includes its TTA. src/evaluate.py does this for holdout predictions, and src/predict.py does the same on the test images:

# predict.py
acc = None
for name in args.runs:
    cfg = json.load(open(KAGGLE / "runs" / name / "args.json"))
    model = build_model(cfg["arch"], cfg["encoder"], pretrained=False).to(device)
    model.load_state_dict(torch.load(KAGGLE / "runs" / name / "model.pt", map_location=device))
    p = predict_probs(model, dl, device, tta=True).astype(np.float32)
    acc = p if acc is None else acc + p
probs = np.round(acc / len(args.runs)).astype(np.uint8)

Writing the submission

Once the probabilities are averaged, predict.py applies the tuned parameters (--params runs/ens/postproc.json) and encodes each non-empty mask. The docstring shows models retrained on all the data (--full), whose names end with _full:

uv run python src/predict.py r34_full effb3_full --params runs/ens/postproc.json

for i, image_id in enumerate(ids.image_id):
    mask = postproc.apply(probs[i], params)
    for c, cls in enumerate(CLASSES):
        rles[(image_id, cls)] = rle_encode(mask[c]) if mask[c].any() else " "
sub["mask_rle"] = [rles[(i, c)] for i, c in zip(sub.image_id, sub.defect_type)]
sub.to_csv(args.out, index=False)

The submission mirrors sample_submission.csv: the columns ID, image_id, defect_type, and mask_rle, with four rows for each of the 3,000 test images, so 12,000 rows. A single space marks "no defect". The encoder is the inverse of the decoder from lesson 1:

def rle_encode(mask):
    pixels = np.concatenate([[0], mask.flatten(order="F").astype(np.uint8), [0]])
    runs = np.where(pixels[1:] != pixels[:-1])[0] + 1
    runs[1::2] -= runs[::2]
    return " ".join(map(str, runs))

The code writes submission.csv locally. The project notebook states that the file is written locally and isn't uploaded, and the folder doesn't contain one, so the course reports no leaderboard score.