keras（keras 运维）

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')
]

6. Data Pipeline

# tf.data for performance
dataset = tf.data.Dataset.from_tensor_slices((x, y))
dataset = dataset.shuffle(10000).batch(32).prefetch(tf.data.AUTOTUNE)

# ImageDataGenerator for augmentation
datagen = keras.preprocessing.image.ImageDataGenerator(
    rotation_range=20,
    horizontal_flip=True,
    validation_split=0.2
)

7. Compile Checklist

model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=0.001),
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

• Learning rate: start 0.001, reduce on plateau
• Batch size: 32-128 typical, larger = smoother gradients

Common Traps

• Input shape mismatch → check data shape vs model input_shape, exclude batch dim
• Loss is NaN → reduce learning rate, check for inf/nan in data, add gradient clipping
• Validation loss diverges → add regularization, reduce model capacity, more data
• Model not learning → check labels are correct, verify loss function matches task
• GPU OOM → reduce batch size, use mixed precision, gradient checkpointing
• Slow training → use tf.data pipeline with prefetch, enable XLA compilation

External Endpoints

Note: Transfer learning examples download pretrained weights on first use. Use weights=None for fully offline operation.

Security & Privacy

Data that stays local:

• Model architectures and configs in ~/keras/
• Training preferences and hyperparameters

This skill does NOT:

• Upload models or data anywhere
• Access files outside ~/keras/ and working directory
• Store training data

Related Skills

Install with clawhub install <slug> if user confirms:

• tensorflow — TensorFlow operations and deployment
• pytorch — Alternative deep learning framework
• ai — General AI and ML patterns
• models — Model architecture design

Feedback

• If useful: clawhub star keras
• Stay updated: clawhub sync