To classify the testing set and prepare a Kaggle submission for the Dog Breed Identification competition, the model is first retrained using the combined training and validation datasets to utilize all available labeled data. The trained network then iterates through the testing set to generate predictions. Because this specific competition evaluates submissions using multi-class log loss rather than simple classification accuracy, a softmax function is applied to the network's output to extract a probability distribution over all breed classes for each test image, instead of using an argmax operation to select a single label. The required submission.csv file must contain an id column followed by individual columns for every breed. A critical implementation detail is that standard dataset loaders read files in lexicographical order; consequently, the numerical test image IDs must be extracted from the filenames and sorted as strings to perfectly align with the sequence of the model's generated probabilities. These sorted IDs are then paired with their respective probability arrays and exported to the CSV format.