Skip to main content

TensorFlow Sweeps

Try in a Colab Notebook here →

Use Weights & Biases for machine learning experiment tracking, dataset versioning, and project collaboration.

Weights & Biases

Use Weights & Biases Sweeps to automate hyperparameter optimization and explore the space of possible models, complete with interactive dashboards like this:

🤔 Why Should I Use Sweeps?

  • Quick setup: With just a few lines of code you can run W&B sweeps.
  • Transparent: We cite all the algorithms we're using, and our code is open source.
  • Powerful: Our sweeps are completely customizable and configurable. You can launch a sweep across dozens of machines, and it's just as easy as starting a sweep on your laptop.

Check out the official documentation $\rightarrow$

What this notebook covers

  • Simple steps to get started with W&B Sweep with custom training loop in TensorFlow.
  • We will find best hyperparameters for our image classification task.

Note: Sections starting with Step are all you need to perform hyperparameter sweep in existing code. The rest of the code is there to set up a simple example.

🚀 Install, Import, and Log in

Step 0️⃣: Install W&B

!pip install wandb

Step 1️⃣: Import W&B and Login

import tqdm
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.datasets import cifar10

import os
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import wandb
from wandb.keras import WandbCallback


Side note: If this is your first time using W&B or you are not logged in, the link that appears after running wandb.login() will take you to sign-up/login page. Signing up is as easy as a few clicks.

👩‍🍳 Prepare Dataset

# Prepare the training dataset
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()

x_train = x_train/255.
x_test = x_test/255.
x_train = np.reshape(x_train, (-1, 784))
x_test = np.reshape(x_test, (-1, 784))

🧠 Define the Model and Training Loop

🏗️ Build a Simple Classifier MLP

def Model():
inputs = keras.Input(shape=(784,), name="digits")
x1 = keras.layers.Dense(64, activation="relu")(inputs)
x2 = keras.layers.Dense(64, activation="relu")(x1)
outputs = keras.layers.Dense(10, name="predictions")(x2)

return keras.Model(inputs=inputs, outputs=outputs)

def train_step(x, y, model, optimizer, loss_fn, train_acc_metric):
with tf.GradientTape() as tape:
logits = model(x, training=True)
loss_value = loss_fn(y, logits)

grads = tape.gradient(loss_value, model.trainable_weights)
optimizer.apply_gradients(zip(grads, model.trainable_weights))

train_acc_metric.update_state(y, logits)

return loss_value

def test_step(x, y, model, loss_fn, val_acc_metric):
val_logits = model(x, training=False)
loss_value = loss_fn(y, val_logits)
val_acc_metric.update_state(y, val_logits)

return loss_value

🔁 Write a Training Loop

Step 3️⃣: Log metrics with wandb.log

def train(train_dataset,

for epoch in range(epochs):
print("\nStart of epoch %d" % (epoch,))

train_loss = []
val_loss = []

# Iterate over the batches of the dataset
for step, (x_batch_train, y_batch_train) in tqdm.tqdm(enumerate(train_dataset), total=len(train_dataset)):
loss_value = train_step(x_batch_train, y_batch_train,
model, optimizer,
loss_fn, train_acc_metric)

# Run a validation loop at the end of each epoch
for step, (x_batch_val, y_batch_val) in enumerate(val_dataset):
val_loss_value = test_step(x_batch_val, y_batch_val,
model, loss_fn,

# Display metrics at the end of each epoch
train_acc = train_acc_metric.result()
print("Training acc over epoch: %.4f" % (float(train_acc),))

val_acc = val_acc_metric.result()
print("Validation acc: %.4f" % (float(val_acc),))

# Reset metrics at the end of each epoch

# 3️⃣ log metrics using wandb.log
wandb.log({'epochs': epoch,
'loss': np.mean(train_loss),
'acc': float(train_acc),
'val_loss': np.mean(val_loss),

Step 4️⃣: Configure the Sweep

This is where you will:

  • Define the hyperparameters you're sweeping over
  • Provide your hyperparameter optimization method. We have random, grid and bayes methods.
  • Provide an objective and a metric if using bayes, for example to minimize the val_loss.
  • Use hyperband for early termination of poorly-performing runs

Check out more on Sweep Configs $\rightarrow$

sweep_config = {
'method': 'random',
'metric': {
'name': 'val_loss',
'goal': 'minimize'
'type': 'hyperband',
'min_iter': 5
'parameters': {
'batch_size': {
'values': [32, 64, 128, 256]
'values': [0.01, 0.005, 0.001, 0.0005, 0.0001]

Step 5️⃣: Wrap the Training Loop

You'll need a function, like sweep_train below, that uses wandb.config to set the hyperparameters before train gets called.

def sweep_train(config_defaults=None):
# Set default values
config_defaults = {
"batch_size": 64,
"learning_rate": 0.01
# Initialize wandb with a sample project name
wandb.init(config=config_defaults) # this gets over-written in the Sweep

# Specify the other hyperparameters to the configuration, if any
wandb.config.epochs = 2
wandb.config.log_step = 20
wandb.config.val_log_step = 50
wandb.config.architecture_name = "MLP"
wandb.config.dataset_name = "MNIST"

# build input pipeline using
train_dataset =, y_train))
train_dataset = (train_dataset.shuffle(buffer_size=1024)

val_dataset =, y_test))
val_dataset = (val_dataset.batch(wandb.config.batch_size)

# initialize model
model = Model()

# Instantiate an optimizer to train the model.
optimizer = keras.optimizers.SGD(learning_rate=wandb.config.learning_rate)
# Instantiate a loss function.
loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True)

# Prepare the metrics.
train_acc_metric = keras.metrics.SparseCategoricalAccuracy()
val_acc_metric = keras.metrics.SparseCategoricalAccuracy()


Step 6️⃣: Initialize Sweep and Run Agent

sweep_id = wandb.sweep(sweep_config, project="sweeps-tensorflow")

You can limit the number of total runs with the count parameter, we will limit a 10 to make the script run fast, feel free to increase the number of runs and see what happens.

wandb.agent(sweep_id, function=sweep_train, count=10)

👀 Visualize Results

Click on the Sweep URL link above to see your live results.

🎨 Example Gallery

See examples of projects tracked and visualized with W&B in our Gallery →

📏 Best Practices

  1. Projects: Log multiple runs to a project to compare them. wandb.init(project="project-name")
  2. Groups: For multiple processes or cross validation folds, log each process as a runs and group them together. wandb.init(group='experiment-1')
  3. Tags: Add tags to track your current baseline or production model.
  4. Notes: Type notes in the table to track the changes between runs.
  5. Reports: Take quick notes on progress to share with colleagues and make dashboards and snapshots of your ML projects.

🤓 Advanced Setup

  1. Environment variables: Set API keys in environment variables so you can run training on a managed cluster.
  2. Offline mode: Use dryrun mode to train offline and sync results later.
  3. On-prem: Install W&B in a private cloud or air-gapped servers in your own infrastructure. We have local installations for everyone from academics to enterprise teams.
Was this page helpful?👍👎