Keras Integration 🧠

Deep learning for humans, orchestrated by flowyml.

What you'll learn

How to track Keras training runs automatically with zero boilerplate. Add one callback, get full experiment tracking and auto-extracted model metadata for free.

Why Keras + flowyml?

• Zero-Boilerplate Tracking: No need to write log_metric loops.
• Auto-Extracted Metadata: Framework, parameters, layers, optimizer—all captured automatically.
• Training History Visualization: Interactive charts in the UI for loss, accuracy, and all metrics.
• Model Versioning: Every model.fit() produces a versioned artifact.
• Reproducibility: Capture exact hyperparameters and architecture.

🧠 FlowymlKerasCallback

The core of the integration is the FlowymlKerasCallback. It automatically logs metrics, parameters, and model checkpoints during training.

Real-World Pattern: Auto-Tracking

import keras
from flowyml.integrations.keras import FlowymlKerasCallback
from flowyml import step, Model

@step
def train_model(x_train, y_train):
    model = keras.Sequential([
        keras.layers.Dense(128, activation='relu'),
        keras.layers.Dense(10, activation='softmax')
    ])

    model.compile(
        optimizer='adam',
        loss='sparse_categorical_crossentropy',
        metrics=['accuracy']  # All metrics automatically tracked!
    )

    # Just add the callback!
    callback = FlowymlKerasCallback(
        experiment_name="mnist_classifier",
        log_model=True  # Auto-save to artifact store
    )

    model.fit(
        x_train, y_train,
        epochs=10,
        callbacks=[callback]
    )

    # 🎯 Use convenience method for full extraction
    return Model.from_keras(
        model,
        name="mnist_classifier",
        callback=callback,  # Auto-extracts training_history!
    )

🎯 Model.from_keras() Convenience Method

The Model.from_keras() method provides the easiest way to create Model assets with full metadata extraction:

from flowyml import Model
from flowyml.integrations.keras import FlowymlKerasCallback

# Train with callback
callback = FlowymlKerasCallback(experiment_name="demo")
model.fit(X, y, callbacks=[callback])

# Create asset with full auto-extraction
model_asset = Model.from_keras(
    model,
    name="my_model",
    callback=callback,  # Training history auto-extracted!
    # Additional custom properties (optional)
    custom_param="value",
)

# Access auto-extracted properties
print(f"Framework: {model_asset.framework}")  # 'keras'
print(f"Parameters: {model_asset.parameters:,}")  # Total param count
print(f"Layers: {model_asset.num_layers}")
print(f"Layer Types: {model_asset.layer_types}")  # ['Dense', 'Dropout', ...]
print(f"Optimizer: {model_asset.optimizer}")  # 'Adam'
print(f"Learning Rate: {model_asset.learning_rate}")
print(f"Loss Function: {model_asset.loss_function}")  # 'mse'
print(f"Metrics: {model_asset.metrics}")  # ['loss', 'accuracy']

# Training info from callback
training_info = model_asset.get_training_info()
print(f"Epochs Trained: {training_info['epochs_trained']}")
print(f"Final Loss: {training_info['final_train_loss']:.4f}")
print(f"Final Val Loss: {training_info['final_val_loss']:.4f}")

📊 Training History Visualization

When you use FlowymlKerasCallback, your training history is automatically visualized in the FlowyML UI:

• Interactive charts for loss, accuracy, and all compiled metrics
• Per-epoch values with hover tooltips
• Log/linear scale toggle for better visualization
• Show/hide individual metrics by clicking the legend
• Zoom and pan for detailed analysis

All metrics passed to model.compile(metrics=[...]) are automatically captured—no configuration needed!

📦 Model Management

Models trained with the callback are automatically registered in flowyml's Model Registry.

from flowyml import ModelRegistry

# Load the latest champion
registry = ModelRegistry()
model = registry.get_model("mnist_classifier", version="latest").load()

🔧 Auto-Extracted Properties

The following properties are automatically extracted from Keras models:

Property	Description
framework	Always 'keras'
parameters	Total parameter count
trainable_parameters	Trainable parameters
architecture	Model name
model_class	Class name (Sequential, Functional, etc.)
num_layers	Number of layers
layer_types	List of layer types
input_shape	Model input shape
output_shape	Model output shape
optimizer	Optimizer name (Adam, SGD, etc.)
learning_rate	Learning rate value
loss_function	Loss function name
metrics	List of compiled metrics
is_compiled	Whether model is compiled
is_built	Whether model is built