Neu.ro allows you to run model training in parallel with different hyperparameter combinations via integration with Weights & Biases. W&B is an experiment tracking tool for deep learning. The ML engineer only needs to initiate the process: prepare the code for training the model, set up the hyperparameter space, and start the search with just one command. Neu.ro is in charge of the rest.
To see this guide in action, check out our recipe in which we apply hyperparemeter tuning with W&B to an image classification task.
The Neu.ro project template contains an integration with Weights and Biases. To create a new project from a template, you need to follow a couple of steps.
Then, create a project:
> neuro project init
This command will then promt you to enter some information about the project:
project_name \[Neuro Project]: Hyperparameter tuning testproject_slug \[hyperparameter-tuning-test]:code_directory \[modules]:
Enter if you don't want to change the suggested value.
Then, change the working directory:
> cd hyperparameter-tuning-test
Now, connect your project with Weights & Biases:
Find your API key (also called a token) on W&B’s settings page (“API keys” section). It should be a sequence like
Add this API key as a secret to the platform:
> neuro secret add wandb-token cf23df2207d99a74fbe169e3eba035e633b63d13
hypertrain.yml from here to
.neuro/hypertrain.yml in your project's directory.
If you have completed the previous steps, W&B is ready to use. To run hyperparameter tuning for the model, you need to:
Define the list of hyperparameters (in a
Send the metrics to W&B after each run (by using the
neuro-flow bake hypertrain command).
config/wandb-sweep.yaml have links to
train.py (you can look at an example here). If you want to run
hypertrain for another script, you can change the
program property in
config/wandb-sweep.yaml (see below). The script must contain the description of the model and the training loop.
The Python script must also receive parameters with the same names as specified in
config/wandb-sweep.yaml as arguments of the command line and use them for model training/evaluation. For example, you can use command line parameters such as the argparse Python module.
Here are some additional details:
train.py is a file that contains the model training code. It should log the metrics with W&B - here's an example for our case:
config/wandb-sweep.yaml has the following structure:
program: /project/src/train.pymethod: gridmetric:name: accuracy01/validgoal: maximizeparameters:lr:values: [0.1, 0.01, 0.001]optimizer:values: ['sgd', 'adam']scheduler:values: ['const', 'cosine']
/../train.py is a default path to a file with the model training code.
Line 2: a method that is used for hyperparameter tuning. For more information, refer to the W&B docs.
Lines 4, 5: the name of the metric that is supposed to be optimized. The ML engineer can change this metric according to the task.
Lines 6 -12: hyperparameter settings. The ML engineer should use them in the
train.py file. Names, values, and ranges are changeable as well.
The name of the file
wandb-sweep.yaml and the path to it can also be modified in
.neuro/hypertrain.yml (look for
start_sweep task definition).
Now that you have set up both Neu.ro and W&B and prepared your training script, it’s time to try hyperparameter tuning. To do this, run the following command:
> neuro-flow bake hypertrain --param token_secret_name wandb-token
This starts jobs that run the
train.py script (or whatever name you have chosen for it) on the platform with different sets of hyperparameters in parallel. By default, just 2 jobs run at the same time. You can change this number by modifying the
id list within the
worker_... task definition in
tasks:...- id: worker_$[[ matrix.id ]]image: $[[ images.myimage.ref ]]preset: gpu-k80-small-pneeds: [start_sweep]strategy:matrix:id: [1, 2] # <- e.g., replace with [1, 2, 3, 4, 5]
To monitor the hyperparameter tuning process, follow the link provided by W&B at the beginning of the process.
If you want to stop the hyperparameter tuning, terminate all related jobs:
> neuro-flow kill hypertrain
After that, verify that the jobs stopped by running