Skip to main content

Run an experiment

Running experiments is the core future of csquare. Experiments allow you to train deep-learning models while taking full advantage of the dedicated and specialized csquare hardware.

note

In order to run an experiment, you need to have created its parent model.

Run a new experiment#

Quick guide#

Step-by-step#

  1. Select the project you want your experiment to be a part of.
  2. Go to the Experiments page using the sidebar.
  3. Click on the + button: Create an experiment of the Experiments widget.
  4. Select to the model you want your experiment to be a part of by clicking on its ID.
  5. Adjust the experiments settings:
    1. Fill the Commit to clone field if you want to run your code from a specific commit. If left empty, the latest commit of the branch will be used.
    2. Set the Billing options as desired, by indicating the number of nodes to use and the experiment's priority. See more info about billing on the dedicated documentation.
  6. Click on the Submit experiment button.

A loader will appear: it means that your new experiment started running!

Experiments are automatically run at their creation. You can however re-run an experiment whenever you want.

Tips

You can check an experiment's status in its dedicated page or in its parent model's page.

Experiments batching#

You may want to run the same experiment with different set of hyper parameters. To serve this purpose, we have created a special syntax which allows you to generate multiple experiments from a single command.

List operator: %[foo,bar,baz]%#

The is list operator allows to specify multiple, arbitrary values.

List operator syntax
%[value1,value2,value3]%
Output:
value1
value2
value3
Example: train a model with 10, 20 and 50 epochs
python train.py --epochs=%[10,20,50]%
Output:
python train.py --epochs=10
python train.py --epochs=20
python train.py --epochs=50

Range operator: %(min,max,[increment])%#

Inspired from the range() Python function, the range operator allow to generate a range of numeric values.

Example: train a model with 10, 20 and 30 epochs
python train.py --epochs=%(10,30,10)%
Output:
python train.py --epochs=10
python train.py --epochs=20
python train.py --epochs=30
danger

Increment, if specified, cannot be equal to zero.

Limitations#

The experiment batching has the following limitations;

  • You cannot generate for than 10000 experiments using the range operator.
  • The total number of generated experiments cannot exceed 65535.

Re-run an experiment#

You can also use a previous experiment a template instead of creating a new one from scratch.

Quick guide#

Step-by-step#

  1. Go to the experiment page that you want to re-run.
  2. Fill the form:
    1. Fill the Commit to clone field if you want to run your code from a specific commit. If left empty, the latest commit of the branch will be used.
    2. Set the Billing options as desired, by indicating the number of nodes to use and the experiment's priority. See more info about billing on the dedicated documentation.
  3. Click on the Submit button.

A loader will appear: it means that your new experiment started running! You can check its status by returning to the corresponding model page.