TensorFlow Developer Certificate — the certifying body practice for the TensorFlow Developer Certificate credential

Normalizing time-series inputs by the TRAIN-set mean/std is preferable because:

1. Computing stats over the test set leaks information
2. The mean/std are zero anyway
3. Normalization is unnecessary for LSTMs
4. Only categorical features need normalization

Answer: A — Computing stats over the test set leaks information

A) Correct — fit on train, apply to val/test. B) Untrue. C) Helps optimization. D) Continuous features need it.

Subword tokenization helps when:

1. Vocabulary is small and fixed
2. You only need character-level model
3. You want to handle OOV gracefully by splitting unknown words into known subwords
4. You want one-hot inputs

Answer: C — You want to handle OOV gracefully by splitting unknown words into known subwords

A) Word-level is fine then. B) Subword sits between word and char. C) Correct — BPE / WordPiece. D) Unrelated.

Sign of overfitting?

1. Training loss + validation loss both high
2. Training loss low, validation loss increases or plateaus high
3. Both losses very low
4. Loss is NaN

Answer: B — Training loss low, validation loss increases or plateaus high

A) Underfit. B) Correct — diverging train + val = overfit. Mitigate with dropout, augmentation, smaller model, early stop. C) Well-fit. D) Numerical instability.

TextVectorization.adapt(dataset) must be called:

1. After training
2. Only on validation set
3. On a text-only dataset before training to build the vocabulary
4. Never — vocabulary is fixed

Answer: C — On a text-only dataset before training to build the vocabulary

A) Too late. B) Should see training distribution. C) Correct. D) Vocabulary is learned.

tf.train.Checkpoint differs from model.save() because:

1. Checkpoints save only weights/optimizer state without architecture; great for resuming training
2. Checkpoints save architecture only
3. They are equivalent
4. Checkpoints can't track variables

Answer: A — Checkpoints save only weights/optimizer state without architecture; great for resuming training

A) Correct. B) Reversed. C) Different scopes. D) Designed exactly for tracking.

tf.keras.optimizers.schedules.ExponentialDecay is passed to:

1. The loss function
2. A callback only
3. A metric
4. An optimizer's learning_rate argument (so LR drops over steps)

Answer: D — An optimizer's learning_rate argument (so LR drops over steps)

A) Unrelated. B) Schedules are not callbacks. C) Not a metric. D) Correct — the schedule object behaves like a callable LR.

Dropout with rate 0.5 means:

1. 50% of weights deleted
2. Used during inference
3. Halve learning rate
4. 50% of activations randomly zeroed during training

Answer: D — 50% of activations randomly zeroed during training

A) Weights remain. D) Correct — dropout randomly zeros activations during training; rescaled to preserve expected value. C) Different. B) Disabled at inference.

Which is the correct Functional API pattern for a two-input model?

1. Sequential([Input(a), Input(b), Dense(1)])
2. inputs=[Input(a), Input(b)]; x=concat([..]); out=Dense(1)(x); Model(inputs, out)
3. Model().add(Input(a)).add(Input(b))
4. tf.keras.layers.MultiInput(...)

Answer: B — inputs=[Input(a), Input(b)]; x=concat([..]); out=Dense(1)(x); Model(inputs, out)

A) Sequential cannot accept multiple inputs. B) Correct Functional pattern. C) Not a valid API. D) No such layer exists.

Strided convolution (stride=2) achieves:

1. Identical spatial size
2. Channel reduction
3. Upsampling
4. Downsampling spatial dims by ~factor of 2 without a pooling layer

Answer: D — Downsampling spatial dims by ~factor of 2 without a pooling layer

A) No downsampling at stride 1. B) Channels are set by filters. C) Use Conv2DTranspose. D) Correct.

Fine-tuning a frozen pretrained backbone is best done by:

1. Unfreezing the entire base immediately at full LR
2. Adding more Dropout
3. Replacing the backbone
4. Unfreezing top layers and continuing training at a much lower LR (e.g., 1e-5)

Answer: D — Unfreezing top layers and continuing training at a much lower LR (e.g., 1e-5)

A) Destroys pretrained features. B) Doesn't accomplish fine-tuning. C) Defeats the purpose. D) Correct.

When using an LSTM for forecasting, the input shape is typically:

1. (samples, features)
2. (time_steps,)
3. (samples, time_steps, features)
4. (samples,)

Answer: C — (samples, time_steps, features)

A) Missing time. B) Missing batch. C) Correct 3D RNN input. D) Missing time and features.

tf.data.AUTOTUNE is used with:

1. Only batch()
2. prefetch and num_parallel_calls (in map) — TF auto-tunes buffer size
3. shuffle() only
4. Optional, not actually a feature

Answer: B — prefetch and num_parallel_calls (in map) — TF auto-tunes buffer size

A) Mainly with batching for tuning batch size. B) Correct — common AUTOTUNE places are .prefetch(AUTOTUNE) and .map(..., num_parallel_calls=AUTOTUNE). C) AUTOTUNE applies. D) Real TF feature.

Image classification model output for 10 classes?

Dense(10, activation='softmax'). Loss: categorical_crossentropy (one-hot) or sparse_categorical_crossentropy (integer labels).

When use pre-trained embeddings (GloVe, Word2Vec)?

Small dataset, want to leverage general semantics. Initialize Embedding with vectors, optionally freeze.

Allowed during exam?

Internet, IDE (PyCharm with plugin), TF documentation. Tasks must run + return acceptable accuracy.

Learning rate too low?

Loss decreases very slowly. Increase the LR or use scheduler.

TFLite?

Convert TF models to mobile/edge format. Use tf.lite.TFLiteConverter. Apply quantization for size + speed.

Exploding gradient — fix?

Gradient clipping: clipnorm or clipvalue in optimizer.