TensorFlow Developer Skills Covered

Know tensors, rank, shape, dtype, eager execution, `tf.Variable`, and assignment behavior. Shape mistakes are especially common: a dense model may expect flat features, a CNN expects image height, width, and channels, and sequence models expect time steps or token positions. `tf.function` can improve execution but also changes debugging because Python code may be traced into a graph.

Use Sequential for simple stacks, Functional for non-linear architectures, and subclassing for custom control. Dense layers fit tabular or flattened inputs, Conv2D layers learn image features, embeddings turn token IDs into vectors, and recurrent layers process sequences. Choose the simplest API that expresses the model clearly before adding custom layers.

Match the output layer and loss to the task. Binary classification commonly pairs a sigmoid output with binary cross-entropy. Single-label multiclass classification commonly uses softmax with categorical or sparse categorical cross-entropy depending on label encoding. Regression usually uses a linear output with MAE or MSE. Accuracy can hide poor recall when classes are imbalanced.

`tf.data.Dataset` pipelines control how data reaches the model. Map transformations parse and preprocess examples, batching controls tensor shape, shuffling improves training for independent samples, caching can avoid repeated work, and prefetching overlaps input work with training. Caching placement and dataset size matter; not every pipeline should use every optimization.

Image models depend on consistent input size, color channels, normalization, augmentation, and validation behavior. Augmentation should affect training data without corrupting validation or test data. Transfer learning usually starts with a frozen pretrained base, then fine-tunes selected layers with a low learning rate once the new classifier head is stable.

Text classification depends on standardization, tokenization, vocabulary adaptation, padding, sequence length, and out-of-vocabulary handling. `TextVectorization` keeps preprocessing tied to the model workflow. Embeddings produce dense token representations, while RNNs, LSTMs, GRUs, or pooling layers summarize sequence information for classification.

Training loss, validation loss, and held-out test metrics answer different questions. Rising validation loss while training loss falls points to overfitting. Callbacks such as EarlyStopping, ModelCheckpoint, ReduceLROnPlateau, and TensorBoard help inspect and control training. Saving checkpoints is not the same as exporting a model for TensorFlow Serving, TensorFlow Lite, or TensorFlow.js.