Desktop app (legacy)¶
The desktop app, alchemist, is the original CustomTkinter interface to ALchemist. This page says how to install it, what it still does that the web app does not, and why it refuses to save some sessions.
Deprecated
The desktop app is deprecated and will be removed in a future release. It gets no new features. The web app, started with alchemist-web, is the supported interface; start with Your first loop.
Install and start¶
The desktop app is no longer part of the default install. Its packages (customtkinter, tksheet, ctkmessagebox, mplcursors and tabulate) come with the desktop extra, and it needs a Python build that includes Tk:
alchemist first prints that the desktop app is deprecated and points to the web app. Without the extra it stops with the name of the missing package and the command above.
What it still does¶
Most of the workflow is in both apps: variables, the initial and optimal designs, training with BoTorch or scikit-learn, the diagnostic plots, suggestions and Find Model Optimum. A few things are only in one of them.
| Only in the desktop app, of the two apps | Only in the web app |
|---|---|
| Multi-objective suggestions (qEHVI and qNEHVI) with a direction per objective and a reference point | The work queue, provenance and the monitoring dashboard (Lab automation, Monitoring) |
| Column roles when loading data (Variable, Target or Drop) and changing the targets after loading | Observed variables, hard-to-change variables and split-plot designs (Split-plot designs) |
| An editable experiment table, and Save Experiments to CSV | Input constraints in designs and suggestions (Constraints) |
| Variable files in CSV as well as JSON | AI effect suggestion for optimal designs (Optimal design) |
| Matplotlib figures with a customize dialog, saved in any Matplotlib format | The Hold list for suggestions, and the 30-second recovery copy |
In Python, two of the desktop's extras are available: multi-objective suggestions with suggest_next(strategy="qNEHVI", goal=[...], ref_point=[...]) (Multi-objective), and target columns with load_data(path, target_columns=[...]). Variables can be saved and loaded as JSON with SearchSpace.save_to_json and load_from_json, but there are no CSV variable files, no editable experiment table and no figure customize dialog; the plot methods return Matplotlib figures that you save yourself.
Known issue
D3: the desktop app ignores input constraints in its designs and suggestions, although Find Model Optimum respects them. D4: its calibration checkbox has no effect (F1 in the web app). D5: Help, About and Settings open placeholder dialogs.
Opening a session from the web app¶
The desktop app keeps its own copy of the variables and the experiment table, and writes that copy back into the session before it saves, trains or logs. Its copy has no place for several things a web or Python session can hold, so writing it back would delete them.
Instead, the desktop app checks first and refuses. A dialog titled "Can't sync this session" lists what would be lost, ends with "Nothing was changed.", and leaves the session as it was. It refuses when the session has:
- an observed variable, or a variable marked hard or very hard to change;
- derived variables;
- input constraints;
- the columns that provenance and split-plot designs add to the experiment table (
ProvenanceId,WholePlot,Subplot,RunOrder), as any session recorded through the work queue or the web app's Add Point... has; - rows with a result whose inputs are empty or not numbers, which it would drop.
The check runs on Save Session, Save Session As..., saving when you quit, Train Model, Log to Audit Trail, and Save & Close in the variable editor. For the first kinds, the dialog says to use the web app for that session. For rows with empty inputs, fill in or delete the empty cells in the experiment table and try again. A session the desktop app made itself is never refused.
Rows that have no result yet, such as suggestions you have not run, are still dropped from the table when it syncs, as in earlier versions.
Moving to the web app¶
The two apps use the same session file. In the web app, choose Load Session on the welcome screen and pick the file the desktop app saved; see Sessions and audit log. The variables, results and model settings come across, and the Model tab shows the settings the session was last trained with. The desktop app's text logs in ./logs are not read; the session's audit log replaces them.