brycewang-stanford/keras-deep-learning

Keras Deep Learning Guide

Overview

Keras is the high-level deep learning API that ships as part of TensorFlow 2.x and is the recommended interface for building, training, and deploying neural networks. Its Sequential and Functional APIs provide a progressive disclosure of complexity: beginners can stack layers in minutes, while researchers can build arbitrary DAG architectures, custom training loops, and multi-output models with the same framework.

This guide covers practical patterns for academic research with Keras, from image classification and sequence modeling to custom loss functions and experiment reproducibility. The focus is on patterns that appear repeatedly in published work -- data loading pipelines, callback orchestration, hyperparameter search, and model introspection -- rather than toy examples.

Keras is particularly strong in rapid prototyping for research papers. Its integration with TensorBoard, Weights & Biases, and tf.data pipelines makes it straightforward to go from idea to reproducible experiment to publication-quality results.

Model Architecture Patterns

Sequential API for Standard Architectures

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

# Image classification baseline
model = keras.Sequential([
    layers.Input(shape=(224, 224, 3)),
    layers.Rescaling(1.0 / 255),
    layers.Conv2D(32, 3, activation="relu", padding="same"),
    layers.BatchNormalization(),
    layers.MaxPooling2D(2),
    layers.Conv2D(64, 3, activation="relu", padding="same"),
    layers.BatchNormalization(),
    layers.MaxPooling2D(2),
    layers.Conv2D(128, 3, activation="relu", padding="same"),
    layers.GlobalAveragePooling2D(),
    layers.Dropout(0.3),
    layers.Dense(256, activation="relu"),
    layers.Dense(10, activation="softmax"),
])

model.compile(
    optimizer=keras.optimizers.AdamW(learning_rate=1e-3, weight_decay=1e-4),
    loss="sparse_categorical_crossentropy",
    metrics=["accuracy"],
)

Functional API for Multi-Input/Multi-Output Models

# Multi-input model for multimodal research
image_input = keras.Input(shape=(224, 224, 3), name="image")
text_input = keras.Input(shape=(128,), dtype="int32", name="text")

# Image branch
x_img = keras.applications.EfficientNetV2B0(
    include_top=False, weights="imagenet", input_tensor=image_input
).output
x_img = layers.GlobalAveragePooling2D()(x_img)

# Text branch
x_txt = layers.Embedding(10000, 128)(text_input)
x_txt = layers.Bidirectional(layers.LSTM(64))(x_txt)

# Merge
merged = layers.Concatenate()([x_img, x_txt])
merged = layers.Dense(256, activation="relu")(merged)
merged = layers.Dropout(0.4)(merged)
output = layers.Dense(5, activation="softmax", name="classification")(merged)

model = keras.Model(inputs=[image_input, text_input], outputs=output)

Data Pipeline with tf.data

Efficient data loading is critical for GPU utilization in research experiments:

def build_dataset(file_pattern, batch_size=32, training=True):
    """Build a tf.data pipeline with augmentation for research experiments."""
    dataset = tf.data.Dataset.list_files(file_pattern, shuffle=training)

    def parse_image(path):
        img = tf.io.read_file(path)
        img = tf.image.decode_jpeg(img, channels=3)
        img = tf.image.resize(img, [256, 256])
        label = tf.strings.split(path, os.sep)[-2]
        return img, label

    dataset = dataset.map(parse_image, num_parallel_calls=tf.data.AUTOTUNE)

    if training:
        dataset = dataset.shuffle(1000)
        dataset = dataset.map(
            lambda x, y: (tf.image.random_flip_left_right(x), y),
            num_parallel_calls=tf.data.AUTOTUNE,
        )

    dataset = dataset.batch(batch_size)
    dataset = dataset.prefetch(tf.data.AUTOTUNE)
    return dataset

Training and Callback Orchestration

Reproducible Training Setup

import os
import random
import numpy as np

def set_seed(seed=42):
    """Ensure reproducibility across runs for paper results."""
    os.environ["PYTHONHASHSEED"] = str(seed)
    random.seed(seed)
    np.random.seed(seed)
    tf.random.set_seed(seed)

set_seed(42)

callbacks = [
    keras.callbacks.ModelCheckpoint(
        "best_model.keras", monitor="val_loss", save_best_only=True
    ),
    keras.callbacks.EarlyStopping(
        monitor="val_loss", patience=10, restore_best_weights=True
    ),
    keras.callbacks.ReduceLROnPlateau(
        monitor="val_loss", factor=0.5, patience=5, min_lr=1e-6
    ),
    keras.callbacks.TensorBoard(log_dir="./logs", histogram_freq=1),
    keras.callbacks.CSVLogger("training_log.csv"),
]

history = model.fit(
    train_dataset,
    validation_data=val_dataset,
    epochs=100,
    callbacks=callbacks,
)

Custom Training Loop for Research

@tf.function
def train_step(model, optimizer, x, y, loss_fn):
    with tf.GradientTape() as tape:
        predictions = model(x, training=True)
        loss = loss_fn(y, predictions)
    gradients = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(gradients, model.trainable_variables))
    return loss

# Custom metric tracking
train_loss = keras.metrics.Mean(name="train_loss")
for epoch in range(num_epochs):
    train_loss.reset_state()
    for x_batch, y_batch in train_dataset:
        loss = train_step(model, optimizer, x_batch, y_batch, loss_fn)
        train_loss.update_state(loss)
    print(f"Epoch {epoch+1}, Loss: {train_loss.result():.4f}")

Debugging and Common Pitfalls

| Issue | Symptom | Solution | |-------|---------|----------| | Exploding gradients | Loss becomes NaN | Add gradient clipping, reduce learning rate | | Overfitting | Val loss diverges from train loss | Add Dropout, data augmentation, weight decay | | Underfitting | Both losses plateau high | Increase model capacity, reduce regularization | | Slow training | Low GPU utilization | Use tf.data with prefetch, increase batch size | | Memory errors | OOM on GPU | Reduce batch size, use mixed precision | | Non-deterministic results | Different results per run | Call set_seed(), set TF_DETERMINISTIC_OPS=1 |

Mixed Precision Training

# Enable mixed precision for 2x speedup on modern GPUs
keras.mixed_precision.set_global_policy("mixed_float16")

# Ensure the output layer uses float32 for numerical stability
output = layers.Dense(10, activation="softmax", dtype="float32")(x)

Best Practices for Research

• Version pin everything. Record tensorflow, keras, numpy, and cuda versions in your paper appendix.
• Use keras.utils.set_random_seed(42) for full determinism (TF 2.12+).
• Save models in .keras format (not HDF5) for forward compatibility.
• Profile with TensorBoard to identify data pipeline bottlenecks before scaling up.
• Use tf.debugging.enable_check_numerics() during development to catch NaN/Inf early.
• Export with tf.saved_model for deployment; export ONNX for cross-framework comparison.

References

• Deep Learning with Python, 2nd Edition -- Francois Chollet (Keras creator)
• Keras documentation -- Official API reference and guides
• TensorFlow tutorials -- End-to-end examples
• fchollet/deep-learning-with-python-notebooks -- Code companion to the book
• Keras examples gallery -- 100+ community-contributed examples