Project: classify receipt image quality with a checked training contract

Define the task

Choose labels such as readable, blurred and cut-off. Record how annotators handle a photo with more than one defect: use a priority rule or a multi-label target with a matching loss. For single-label training, pin class order and image input format. Split original receipt identities before augmentation. Tensor checks] and split checks] must run.

Train against a baseline

Start with class prevalence and a simple image model. Show a tiny-batch overfit check, then train under a fixed optimizer and epoch budget. Log examples seen, loss, invalid gradients and validation counts. Use raw logits for cross-entropy and the correct evaluation mode. The loss contract] and weighted validation] are review criteria.

Exercise recovery and serving

Interrupt training after a checkpoint and resume with the same class mapping and input manifest. Compare one fixed-batch update. Publish an inference manifest with preprocessing and maximum request size. Reject a corrupt image and oversized request; measure loaded high-percentile latency at a stated concurrency. Serving] must match the selected checkpoint.

Report failure slices

Show a confusion table and error examples for dim light, angled photos and duplicate captures. State the count per slice; do not give percentages without denominators. Compare the selected model with the baseline on an untouched test partition once. Submit code, run manifest, sample IDs and failure-case outputs so another engineer can reproduce the decision.

Implementation

def check_image_project(train_ids, validation_ids, test_ids, class_names):
    partitions = [set(train_ids), set(validation_ids), set(test_ids)]
    if any(not group for group in partitions):
        raise ValueError("empty image partition")
    if any(partitions[left] & partitions[right]
           for left, right in ((0, 1), (0, 2), (1, 2))):
        raise ValueError("identity leaked across partitions")
    if len(class_names) != len(set(class_names)):
        raise ValueError("duplicate class label")

Performance and operating cost

Each training epoch reads O(N) images; decoding and augmentation may dominate loader CPU. Include peak device memory, total training time and loaded inference tail latency in the project report.

Common Mistakes

Do not split variants independently of original receipts.
Do not promote a checkpoint without preprocessing metadata.
Do not hide tiny failure slices behind overall accuracy.