Configure experiments
Use the wandb.config object to save your training configuration such as:
- hyperparameters
- input settings such as the dataset name or model type
- any other independent variables for your experiments.
The wandb.config attribute makes it easy to analyze your experiments and reproduce your work in the future. You can group by configuration values in the W&B App, compare the settings of different W&B Runs and view how different training configurations affect the output. A Run's config attribute is a dictionary-like object, and it can be built from lots of dictionary-like objects.
Dependent variables (like loss and accuracy) or output metrics should be saved with wandb.loginstead.
Set up an experiment configuration
Configurations are typically defined in the beginning of a training script. Machine learning workflows may vary, however, so you are not required to define a configuration at the beginning of your training script.
We recommend that you avoid using dots in your config variable names. Instead, use a dash or underscore instead. Use the dictionary access syntax ["key"]["foo"] instead of the attribute access syntax config.key.foo if your script accesses wandb.config keys below the root.
The following sections outline different common scenarios of how to define your experiments configuration.
Set the configuration at initialization
Pass a dictionary at the beginning of your script when you call the wandb.init() API to generate a background process to sync and log data as a W&B Run.
The proceeding code snippet demonstrates how to define a Python dictionary with configuration values and how to pass that dictionary as an argument when you initialize a W&B Run.
import wandb
# Define a config dictionary object
config = {
"hidden_layer_sizes": [32, 64],
"kernel_sizes": [3],
"activation": "ReLU",
"pool_sizes": [2],
"dropout": 0.5,
"num_classes": 10,
}
# Pass the config dictionary when you initialize W&B
run = wandb.init(project="config_example", config=config)
You can pass a nested dictionary to wandb.config(). W&B will flatten the names using dots in the W&B backend.
Access the values from the dictionary similarly to how you access other dictionaries in Python:
# Access values with the key as the index value
hidden_layer_sizes = wandb.config["hidden_layer_sizes"]
kernel_sizes = wandb.config["kernel_sizes"]
activation = wandb.config["activation"]
# Python dictionary get() method
hidden_layer_sizes = wandb.config.get("hidden_layer_sizes")
kernel_sizes = wandb.config.get("kernel_sizes")
activation = wandb.config.get("activation")
Throughout the Developer Guide and examples we copy the configuration values into separate variables. This step is optional. It is done for readability.
Set the configuration with argparse
You can set your configuration with an argparse object. argparse, short for argument parser, is a standard library module in Python 3.2 and above that makes it easy to write scripts that take advantage of all the flexibility and power of command line arguments.
This is useful for tracking results from scripts that are launched from the command line.
The proceeding Python script demonstrates how to define a parser object to define and set your experiment config. The functions train_one_epoch and evaluate_one_epoch are provided to simulate a training loop for the purpose of this demonstration:
# config_experiment.py
import wandb
import argparse
import numpy as np
import random
# Training and evaluation demo code
def train_one_epoch(epoch, lr, bs):
acc = 0.25 + ((epoch / 30) + (random.random() / 10))
loss = 0.2 + (1 - ((epoch - 1) / 10 + random.random() / 5))
return acc, loss
def evaluate_one_epoch(epoch):
acc = 0.1 + ((epoch / 20) + (random.random() / 10))
loss = 0.25 + (1 - ((epoch - 1) / 10 + random.random() / 6))
return acc, loss
def main(args):
# Start a W&B Run
run = wandb.init(project="config_example", config=args)
# Access values from config dictionary and store them
# into variables for readability
lr = wandb.config["learning_rate"]
bs = wandb.config["batch_size"]
epochs = wandb.config["epochs"]
# Simulate training and logging values to W&B
for epoch in np.arange(1, epochs):
train_acc, train_loss = train_one_epoch(epoch, lr, bs)
val_acc, val_loss = evaluate_one_epoch(epoch)
wandb.log(
{
"epoch": epoch,
"train_acc": train_acc,
"train_loss": train_loss,
"val_acc": val_acc,
"val_loss": val_loss,
}
)
if __name__ == "__main__":
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument("-b", "--batch_size", type=int, default=32, help="Batch size")
parser.add_argument(
"-e", "--epochs", type=int, default=50, help="Number of training epochs"
)
parser.add_argument(
"-lr", "--learning_rate", type=int, default=0.001, help="Learning rate"
)
args = parser.parse_args()
main(args)
Set the configuration throughout your script
You can add more parameters to your config object throughout your script. The proceeding code snippet demonstrates how to add new key-value pairs to your config object:
import wandb
# Define a config dictionary object
config = {
"hidden_layer_sizes": [32, 64],
"kernel_sizes": [3],
"activation": "ReLU",
"pool_sizes": [2],
"dropout": 0.5,
"num_classes": 10,
}
# Pass the config dictionary when you initialize W&B
run = wandb.init(project="config_example", config=config)
# Update config after you initialize W&B
wandb.config["dropout"] = 0.2
wandb.config.epochs = 4
wandb.config["batch_size"] = 32
You can update multiple values at a time:
wandb.init(config={"epochs": 4, "batch_size": 32})
# later
wandb.config.update({"lr": 0.1, "channels": 16})
Set the configuration after your Run has finished
Use the W&B Public API to update your config (or anything else about from a complete Run) after your Run. This is particularly useful if you forgot to log a value during a Run.
Provide your entity, project name, and the Run ID to update your configuration after a Run has finished. Find these values directly from the Run object itself wandb.run or from the W&B App UI:
api = wandb.Api()
# Access attributes directly from the run object
# or from the W&B App
username = wandb.run.entity
project = wandb.run.project
run_id = wandb.run.id
run = api.run(f"{username}/{project}/{run_id}")
run.config["bar"] = 32
run.update()
absl.FLAGS
You can also pass in absl flags.
flags.DEFINE_string("model", None, "model to run") # name, default, help
wandb.config.update(flags.FLAGS) # adds absl flags to config
File-Based Configs
Key-value pairs are automatically passed to wandb.config if you create a file called config-defaults.yaml.
The proceeding code snippet demonstrates a sample config-defaults.yaml YAML file:
# config-defaults.yaml
# sample config defaults file
epochs:
desc: Number of epochs to train over
value: 100
batch_size:
desc: Size of each mini-batch
value: 32
You can overwrite automatically passed by a config-defaults.yaml. To do so , pass values to the config argument of wandb.init.
You can also load different config files with the command line argument --configs.
Example use case for file-based configs
Suppose you have a YAML file with some metadata for the run, and then a dictionary of hyperparameters in your Python script. You can save both in the nested config object:
hyperparameter_defaults = dict(
dropout=0.5,
batch_size=100,
learning_rate=0.001,
)
config_dictionary = dict(
yaml=my_yaml_file,
params=hyperparameter_defaults,
)
wandb.init(config=config_dictionary)
TensorFlow v1 flags
You can pass TensorFlow flags into the wandb.config object directly.
wandb.init()
wandb.config.epochs = 4
flags = tf.app.flags
flags.DEFINE_string("data_dir", "/tmp/data")
flags.DEFINE_integer("batch_size", 128, "Batch size.")
wandb.config.update(flags.FLAGS) # add tensorflow flags as config