Documentation
Search…
English
Running Sweeps in Jupyter
Run Sweeps using Jupyter notebooks
Sweeps allow you to easily try out a large number of hyperparameters while tracking model performance and logging all the information you need to reproduce experiments. There are two major APIs for Sweeps: one using the command line (guide here) and the other using pure Python. This guide describes how to use the second API to run a sweep inside a Jupyter notebook.
0. How do sweeps work?
Two components work together in a sweep: a controller on the central sweep server, which picks out new hyperparameter combinations to try, and agents, running in any number of processes on any number of machines, which query the server for hyperparameters, use them to run model training, and then report the results back to the controller.
The benefit of this setup is in parallelization: we handle tricky aspects like tracking state over time and communicating between machines, so you can focus on the machine learning. That means it's as easy as, say, opening two copies of the same Google Colab notebook in two tabs to get a 2x speed-up in your hyperparameter search!
1. Initialize the sweep
To get started, we need to configure the sweep: decide which hyperparameters we will search over, which
method we will use to pick new hyperparameters, and so on. We store all of this information in a dictionary (or dictionary-like) Python data structure, the sweep_config. For details, see this guide.We then initialize our sweep by calling
wandb.sweep. Note that no experiments are launched yet! We still need to define a training function and pass it to an agent.1
import wandb
2
3
sweep_config = {
4
"name" : "my-sweep",
5
"method" : "random",
6
"parameters" : {
7
"epochs" : {
8
"values" : [10, 20, 50]
9
},
10
"learning_rate" :{
11
"min": 0.0001,
12
"max": 0.1
13
}
14
}
15
}
16
17
sweep_id = wandb.sweep(sweep_config)
Copied!
wandb.sweep returns an identifier for the sweep. Agents need that identifier to join in.For more details, check out the reference docs for
wandb.sweep here:2. Run an agent
Now, in any process on any machine (including the one we just used to initialize the sweep), we just
- 1.define a function to run training based on those hyperparameters, and
- 2.pass that function, plus the
sweep_id, towandb.agent.
That's all we need to do! If we want to search more quickly, we can repeat this procedure on more machines.
1
def train():
2
with wandb.init() as run:
3
config = wandb.config
4
model = make_model(config)
5
for epoch in range(config["epochs"]):
6
loss = model.fit() # your model training code here
7
wandb.log({"loss": loss, "epoch": epoch})
8
9
count = 5 # number of runs to execute
10
wandb.agent(sweep_id, function=train, count=count)
Copied!
For more details, check out the reference docs for
wandb.agent here:Last modified 5mo ago
Copy link