TensorFlow Developer Practice Test

Add metrics that better reflect minority-class performance, such as precision and recall is correct because Precision and recall are more informative than plain accuracy when the positive class is rare and important. The cited source, Classification on imbalanced data, supports this answer for the Training and Evaluation scenario rather than the adjacent distractors.

The correct choice is the one describing TensorFlow as an open-source machine learning platform, because that captures its role in model building, training, and deployment. The other options describe tools for unrelated infrastructure or office tasks.

tf.data.experimental.make_csv_dataset is the high-level function designed to load CSV files into a tf.data.Dataset where each element is a tuple of (features_dict, labels). It automatically handles header reading, column type inference, batching, and shuffling. tf.data.TextLineDataset only reads raw lines as strings and requires manual parsing using tf.io.decode_csv.

Dropout is correct because Dropout is commonly used to reduce overfitting by randomly dropping units during training, which can help the model generalize better. The cited source, Overfit and underfit, supports this answer for the Neural Network Building Blocks scenario rather than the adjacent distractors.

The current TensorFlow certificate page says that credentials remain valid for three years from the date the exam was passed. Closing the exam for new scheduling is not the same thing as voiding credentials already earned.

A saved full model rather than only a screenshot or metric table is correct because A full saved model is the right artifact because it preserves the information needed to reload and reuse the model directly. The cited source, Save and load models, supports this answer for the Saving and Deployment scenario rather than the adjacent distractors.

Longer reviews can expose issues with sequence length handling, truncation, padding, or model capacity, so that is the most relevant area to inspect. Font size and model file extension do not explain why performance changes with longer textual inputs.

Using information from the future when building features or labels is a form of data leakage, because it gives the model an unrealistic advantage during training or evaluation. That can make the forecast look better than it will be in real use.

Use transfer learning from a pretrained vision model is correct because Transfer learning is often the stronger option on small image datasets because the model starts from useful pretrained features instead of learning everything from scratch. The cited source, Transfer learning and fine-tuning, supports this answer for the Computer Vision scenario rather than the adjacent distractors.

To build a custom stateful metric in Keras, developers subclass tf.keras.metrics.Metric. They must override four methods: 1) __init__() to declare state variables using self.add_weight(), 2) update_state() to accumulate statistics across batches, 3) result() to return the current metric value, and 4) reset_state() to clear statistics at the end of each epoch.