Take the exam and earn your certificate.

The source encoder trained on generated vertical, horizontal and cross drawings is reused for line-versus-cross target drawings. Compare scratch, frozen and fine-tuned recipes with target validation. The experiment can support a bounded statement about that generator and recipe; it cannot establish ImageNet-scale or real-camera transfer, NVIDIA affiliation or professional competence.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

Assign whole groups to training, validation or final test before generating views. Use 16/24/40 target groups, giving 32/48/80 views. Fit mean and population standard deviation on 480 source-training images, reuse them unchanged, and reserve target validation for selection. A manifest can expose final group IDs and counts without creating final pixel or label tensors.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

Pooling gives[5,6,4,4], flatten gives[5,96], encoder output[5,12], and target logits[5,2]. The head has 12*2+2=26 learned values; encoder 1,236 plus head 26 gives 1,262. BatchNorm running statistics and counter are buffers, not learned parameters. Class order and the source-training normaliser must accompany compatible shapes.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

The observed changed keys are 1.running_mean,1.running_var and 1.num_batches_tracked. No parameter changed; BatchNorm updated buffers during a training-mode forward. Use encoder.requires_grad_(False), an optimiser containing only the head, and encoder.eval(). Snapshot all encoder state, adapt the head and require no changed keys. no_grad alone does not stop BatchNorm buffer updates.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

Select fine_tune because 0.518323 is the smallest selected target-validation loss. The other losses are 0.591373 and 0.688436. Frozen computes features once, so equal epochs do not imply equal compute. Encoder learning rates differ and only one seed/generator was used. This supports the chosen recipe for this development experiment, not a universal advantage or real-image claim.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

Matrix[[19,5],[8,16]] gives 35/48 accuracy, line recall 19/24 and cross recall 16/24. The first clean error is target_val-0-000-v1. Blank errors begin target_val-2-000-v0 and target_val-2-000-v1; all inputs become identical, so a constant line prediction gets 24/48. This destructive control does not establish robustness to blur or real lighting changes.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

The validation-selected fine-tuned model achieved 68/80 correct views from 40 groups, versus a 40/80 constant-line baseline. Its final matrix[[30,10],[2,38]] gives line recall 75 percent and cross recall 95 percent. The 12 errors come from six groups. Only this recipe received final evaluation, so this does not establish a final-test ranking of three strategies, real-camera performance or professional suitability.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.

Preserve 14 passing tests, fit.json, selected.pt, final.json and command logs with interpreter versions. The recorded run selects fine_tune, validation 35/48 and final 68/80. The bad freeze changes three BatchNorm buffers and no parameters; good frozen adaptation changes no encoder state. Different machine or package results require investigation and accurate reporting. One replay is not a new independent test or learner validation.

This worked response connects the stated activity to the actual experiment and its limits. It is ungraded practice.