Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Katib documentation for DARTS algorithm #1991

Merged
merged 4 commits into from
Jun 11, 2020
Merged

Katib documentation for DARTS algorithm #1991

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c25ac3a
Init commit
andreyvelich Jun 9, 2020
7e902a9
Add darts docs
andreyvelich Jun 10, 2020
1dd128b
Fix
andreyvelich Jun 10, 2020
fdf93a1
Merge remote-tracking branch 'upstream/master' into add-darts-algorithm
andreyvelich Jun 11, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
125 changes: 125 additions & 0 deletions content/en/docs/components/hyperparameter-tuning/experiment.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -155,6 +155,7 @@ descriptions on this page:
* [Tree of Parzen Estimators (TPE)](#tpe-search)
* [Covariance Matrix Adaptation Evolution Strategy (CMA-ES)](#cmaes)
* [Neural Architecture Search based on ENAS](#enas)
* [Differentiable Architecture Search (DARTS)](#darts)

More algorithms are under development. You can add an algorithm to Katib
yourself. See the guide to [adding a new
Expand Down Expand Up @@ -464,6 +465,130 @@ For more information, see:
The example aims to show all the possible operations. Due to the large search
space, the example is not likely to generate a good result.

<a id="darts"></a>
#### Differentiable Architecture Search (DARTS)

{{% alert title="Alpha version" color="warning" %}}
Neural architecture search is currently in <b>alpha</b> with limited support.
The Kubeflow team is interested in any feedback you may have, in particular with
regards to usability of the feature. You can log issues and comments in
the [Katib issue tracker](https://github.com/kubeflow/katib/issues).
{{% /alert %}}

The algorithm name in Katib is `darts`.

Currently, you can't view results of this algorithm in the Katib UI and
you can run experiment only on single GPU.

Katib supports the following algorithm settings:

<div class="table-responsive">
<table class="table table-bordered">
<thead class="thead-light">
<tr>
<th>Setting Name</th>
<th>Type</th>
<th>Default value</th>
<th>Description</th>
</tr>
</thead>
<tbody>
<tr>
<td>num_epochs</td>
<td>int</td>
<td>50</td>
<td>Number of epochs to train model</td>
</tr>
<tr>
<td>w_lr</td>
<td>float</td>
<td>0.025</td>
<td>Initial learning rate for training model weights.
This learning rate annealed down to <code>w_lr_min</code>
following a cosine schedule without restart.</td>
</tr>
<tr>
<td>w_lr_min</td>
<td>float</td>
<td>0.001</td>
<td>Minimum learning rate for training model weights.</td>
</tr>
<tr>
<td>w_momentum</td>
<td>float</td>
<td>0.9</td>
<td>Momentum for training training model weights.</td>
</tr>
<tr>
<td>w_weight_decay</td>
<td>float</td>
<td>3e-4</td>
<td>Training model weight decay.</td>
</tr>
<tr>
<td>w_grad_clip</td>
<td>float</td>
<td>5.0</td>
<td>Max norm value for clipping gradient norm of training model weights.</td>
</tr>
<tr>
<td>alpha_lr</td>
<td>float</td>
<td>3e-4</td>
<td>Initial learning rate for alphas weights.</td>
</tr>
<tr>
<td>alpha_weight_decay</td>
<td>float</td>
<td>1e-3</td>
<td>Alphas weight decay.</td>
</tr>
<tr>
<td>batch_size</td>
<td>int</td>
<td>128</td>
<td>Batch size for dataset.</td>
</tr>
<tr>
<td>num_workers</td>
<td>int</td>
<td>4</td>
<td>Number of subprocesses to download dataset.</td>
</tr>
<tr>
<td>init_channels</td>
<td>int</td>
<td>16</td>
<td>Initial number of channels.</td>
</tr>
<tr>
<td>print_step</td>
<td>int</td>
<td>50</td>
<td>Number of training or validation steps before logging it.</td>
</tr>
<tr>
<td>num_nodes</td>
<td>int</td>
<td>4</td>
<td>Number of DARTS nodes.</td>
</tr>
<tr>
<td>stem_multiplier</td>
<td>int</td>
<td>3</td>
<td>Multiplier for initial channels. It is used in first stem cell.</td>
</tr>
</tbody>
</table>
</div>

For more information, see:

* Information in the Katib repository on [Differentiable Architecture Search](https://github.com/kubeflow/katib/tree/master/pkg/suggestion/v1alpha3/nas/darts).
* As a DARTS example, see the YAML file for the
[darts-example-gpu](https://github.com/kubeflow/katib/blob/master/examples/v1alpha3/nas/darts-example-gpu.yaml).

<a id="metrics-collector"></a>
### Metrics collector

Expand Down