tensorflow-ml

When to Use

• Building neural network models with Keras Sequential or Functional API.
• Setting up efficient data pipelines with tf.data.
• Training, evaluating, and tuning deep learning models.
• Applying transfer learning from pre-trained models (ResNet, MobileNet, BERT).
• Saving, loading, and serving models (SavedModel, H5, TFLite).
• Debugging training issues: overfitting, vanishing gradients, slow convergence.
• Configuring GPU/TPU acceleration and mixed precision training.
• Visualizing training with TensorBoard.

Critical Patterns

• Keras Is THE API: Use tf.keras for all model building. Raw tf.Module and tf.GradientTape are only for advanced custom training loops. Start with the high-level API.
• Functional API Over Sequential: Use Functional API for anything beyond a simple linear stack. It supports multi-input, multi-output, shared layers, and residual connections.
• tf.data for Everything: Never use Python generators or numpy loading for training data in production. tf.data.Dataset handles prefetching, parallel mapping, and memory-efficient streaming.
• Callbacks Are Your Safety Net: Always use EarlyStopping (patience-based), ModelCheckpoint (save best weights), and ReduceLROnPlateau. Never train blind without them.
• Validate on Held-Out Data: Always split data into train/validation/test. Use validation loss ( not training loss) for all tuning decisions. Test set is touched exactly once.
• Mixed Precision for Speed: Enable tf.keras.mixed_precision.set_global_policy("mixed_float16") on modern GPUs (Volta+) for ~2x speedup with minimal accuracy impact.
• SavedModel for Deployment: Always export as SavedModel format (not H5) for production serving. SavedModel preserves the computation graph and is framework-agnostic.

Code Examples

Model Building: Functional API

import tensorflow as tf
from tensorflow import keras
from keras import layers

def build_classifier(input_shape: tuple[int, ...], num_classes: int) -> keras.Model:
    """Build a CNN classifier with residual connections."""
    inputs = keras.Input(shape=input_shape, name="image_input")

    # Convolutional block 1
    x = layers.Conv2D(32, 3, padding="same", activation="relu")(inputs)
    x = layers.BatchNormalization()(x)
    x = layers.Conv2D(32, 3, padding="same", activation="relu")(x)
    x = layers.BatchNormalization()(x)
    x = layers.MaxPooling2D()(x)
    x = layers.Dropout(0.25)(x)

    # Convolutional block 2 with residual
    shortcut = layers.Conv2D(64, 1, strides=2, padding="same")(x)
    x = layers.Conv2D(64, 3, padding="same", activation="relu")(x)
    x = layers.BatchNormalization()(x)
    x = layers.Conv2D(64, 3, padding="same")(x)
    x = layers.BatchNormalization()(x)
    x = layers.MaxPooling2D()(x)
    x = layers.Add()([x, shortcut])  # Residual connection
    x = layers.Activation("relu")(x)
    x = layers.Dropout(0.25)(x)

    # Classification head
    x = layers.GlobalAveragePooling2D()(x)
    x = layers.Dense(128, activation="relu")(x)
    x = layers.Dropout(0.5)(x)
    outputs = layers.Dense(num_classes, activation="softmax", name="predictions")(x)

    return keras.Model(inputs=inputs, outputs=outputs, name="cnn_classifier")

model = build_classifier(input_shape=(224, 224, 3), num_classes=10)
model.summary()