Export a model

A trained model can leave dashAI as a single .dashai-model file and be used from your own Python code, without opening the app or running its server.

Model file or session file

	.dashai-model	.dashai
Holds	One trained run	A whole session, its dataset and chosen runs
Opened with	Python, load_model	The dashAI app
Predict	Yes	Yes
Retrain or add models	No	Yes

To share a whole session, see Share a session.

Export

Open a session, click a finished model to open its detail view, and click the Export model button (upload icon) next to Retrain. The file contains the trained model, the preprocessing fitted for its session, and the types of the training columns. The dataset itself is not included, but some models keep training rows inside them to predict (for example KNN stores all of them). Do not share the file if the training data is private.

Use it from Python

Install dashAI in the environment where you want to predict:

pip install dashai

Then:

from DashAI import load_model
import pandas as pd

model = load_model("my_model.dashai-model")
print(model.info)  # model, task, columns, versions and test metrics

new_rows = pd.read_csv("new_data.csv")
predictions = model.predict(new_rows)

predict takes a pandas DataFrame or a list of dicts with the original columns of the dataset. Preprocessing is applied for you, and the target column is ignored if it is present.

Serving it

dashAI does not include a server for exported models yet. A minimal one with FastAPI:

from fastapi import FastAPI
from DashAI import load_model

app = FastAPI()
model = load_model("my_model.dashai-model")


@app.post("/predict")
def predict(rows: list[dict]):
    return {"predictions": model.predict(rows)}

Compatibility

• If the libraries installed differ from the ones the model was trained with, dashAI shows a warning and still tries to load it.
• If the model needs a plugin that is not installed, loading fails and the error says which one to install.

warning

A model file contains Python objects, and loading it runs code. Only load files from people you trust.

Tabular classification and regression models are tested. Other model types (images, text, time series) can be exported but are experimental.