AIM Web UI¶
AIM web interface is quite intuitive and the official documentation already provides a general purpose tutorial.
In this mini guide we limit to showcase a basic set of operations to navigate the ML artifacts using some artifacts from our IMC23 paper.
To replicate the following, make sure you installed the needed artifacts.
Output
Running Aim UI on repo `<Repo#-3653246895908991301 path=./notebooks/imc23/campaigns/ucdavis-icdm19/augmentation-at-loading-with-dropout/.aim read_only=None>`
Open http://127.0.0.1:43800
Press Ctrl+C to exit
Run aim up --help for more options (e.g., specifying a different port or hostname).
When visiting the URL reported in the output you land on the home page of the AIM repository.
This collects a variety of aggregate metrics and track activity over time. Hence, in our scenario the home page of the ML artifacts are mostly empty because all campaigns were generated in a specific moment in time.
The left side bar allows switch the view. In particular, "Runs" show a tabular view of the runs collected in the repository.
From the view you can see the hash of each run and scrolling horizontally you can glance over the metadata stored for each run.
The search bar on the top of the page
allows to filter runs.
It accept python expression bounded
to a run entry point.
For instance, in the following example we filter one specific run based on hyper parameters.
Using the search box
The search box accept python expressions and run.hparams
is a dictionary of key-value pairs related to the different runs.
As from the example, you can use the traditional python
syntax of dict[<key>] == <value> to filter, but the search
box supports also a dot-notated syntax hparams.<key> == <value>
which has an autocomplete.
In the example, the search is based on equality but any other python operation is allowed.
When clicking the hash of a run (e.g., the one we filtered) we switch to a per-run view which further details the collected metadata of the selected run.
For instance, when scrolling at the bottom of the per-run page we can see that AIM details
-
The specific git commit used when executing the run.
-
The specific python packages and related versions available in the environment when executing the run.
Both are automatically tracked by AIM with no extra code required (beside activating the their collection when creating the run).
The per-run view offers a variety of information organized in multiple tabs.
For instance, the tab "Logs" details the console output.
