Keras

Setup

On first use, check setup.md for integration guidelines. The skill stores preferences in ~/keras/ when the user confirms.

When to Use

User builds neural networks with Keras or TensorFlow. Agent handles model architecture, layer configuration, training loops, callbacks, debugging loss issues, and deployment preparation.

Architecture

Memory lives in ~/keras/. See memory-template.md for setup.

~/keras/
├── memory.md          # Preferred architectures, hyperparams
└── models/            # Saved model configs (optional)

Quick Reference

Core Rules

1. Sequential vs Functional API

• Sequential: simple stacks, no branching
• Functional: multi-input/output, skip connections, shared layers
• Subclassing: custom forward pass, dynamic architectures

# Sequential - simple stack
model = keras.Sequential([
    layers.Dense(64, activation='relu'),
    layers.Dense(10, activation='softmax')
])

# Functional - flexible graphs
inputs = keras.Input(shape=(784,))
x = layers.Dense(64, activation='relu')(inputs)
outputs = layers.Dense(10, activation='softmax')(x)
model = keras.Model(inputs, outputs)

2. Input Shape Patterns

• First layer needs input_shape (exclude batch)
• Images: (height, width, channels) for channels_last
• Sequences: (timesteps, features)
• Tabular: (features,)

# Image input
layers.Conv2D(32, 3, input_shape=(224, 224, 3))

# Sequence input
layers.LSTM(64, input_shape=(100, 50))  # 100 timesteps, 50 features

# Tabular input
layers.Dense(64, input_shape=(20,))  # 20 features

3. Activation Functions

4. Regularization Stack

Apply in this order for overfitting:

1. Dropout - after dense/conv layers (0.2-0.5)
2. BatchNorm - before or after activation
3. L2 regularization - in layer (0.01-0.001)
4. Early stopping - callback with patience

layers.Dense(64, activation='relu', kernel_regularizer=keras.regularizers.l2(0.01))
layers.Dropout(0.3)
layers.BatchNormalization()

5. Callbacks Essentials

callbacks = [
    keras.callbacks.EarlyStopping(
        monitor='val_loss', patience=5, restore_best_weights=True
    ),
    keras.callbacks.ModelCheckpoint(
        'best_model.keras', save_best_only=True
    ),
    keras.callbacks.ReduceLROnPlateau(
        monitor='val_loss', factor=0.5, patience=3
    ),
    keras.callbacks.TensorBoard(log_dir='./logs')
]