Take the exam and earn your certificate.

Predict vertical, horizontal or cross for a generated array. Train on training groups, fit statistics there, select minimum clean validation loss over 16 fixed epochs and freeze the selected state before the test command. A training-label majority tie chooses class 0, giving a 12/36 validation baseline. Neither a high score nor completing the exercise establishes real-image performance or NVIDIA certification.

This worked answer supports the ungraded activity. It does not establish independent practical competence or add exam marks.

Keep both views in one source group and split groups first. Use 54/18/18 groups, giving 108/36/36 images. Fit scalar mean and population standard deviation on training only and reuse them unchanged. Inputs are float32 [108,1,8,8] and targets long [108]. Moving a training group into validation must fail the disjointness check.

This worked answer supports the ungraded activity. It does not establish independent practical competence or add exam marks.

Shapes are [5,1,8,8], [5,4,8,8], [5,4,4,4], [5,64] and [5,3]. The convolution has 40 parameters and the linear layer 195, total 235. Zero logits give log(3) loss for valid class-index targets. Floating class indices of shape [5] do not meet this target contract.

This worked answer supports the ungraded activity. It does not establish independent practical competence or add exam marks.

There are six batches per epoch and 96 updates. Select minimum clean validation loss with earliest exact ties; the author selected 15, but read your own history. Training updates weights; validation evaluates and chooses among candidates. eval changes dropout behaviour, inference_mode disables gradient recording, and deep-copying preserves selected parameters without later aliasing.

This worked answer supports the ungraded activity. It does not establish independent practical competence or add exam marks.

Clean validation has 36 correct and blank validation 12 in the saved author run. Rows are true classes and columns predictions. Two blank-validation error IDs in the saved run are c1-g18-v0 and c1-g18-v1: both have true class horizontal (1) and predicted class vertical (0). They are correlated views from one source group, not two independent real-world cases. Retain blank-input errors rather than relabelling them. A next experiment may change a validation stress or training-only transformation with an explicit label rule, but remains development evidence. Keep final testing closed until the revised decision rule is frozen.

This worked answer supports the ungraded activity. It does not establish independent practical competence or add exam marks.

For the saved author experiment, epoch 15 was selected from 16 fixed epochs. Clean validation and final test were each 36/36; the majority baseline and blank validation stress were each 12/36 on their stated populations. Record actual checkpoint and payload hashes, not shortened examples. Preserve CPU and synthetic-data limits. Different-author review, real-image evaluation and publication checks are separate responsibilities.

This worked answer supports the ungraded activity. It does not establish independent practical competence or add exam marks.