Supported models
PeekML works with models that take a few separate values, such as numbers and categories, or a picture, and return numbers, labels or probabilities.
Works
| Kind of model | Shown as |
|---|---|
| Classification (scikit-learn, XGBoost, LightGBM, Keras / TensorFlow, PyTorch) | Predicted class and probability bars |
| Regression | The predicted number |
| Regression with several outputs | The list of predicted numbers |
| Clustering (e.g. KMeans) | Cluster number and distances |
| Anomaly detection (e.g. IsolationForest) | Normal / anomaly label and score |
All features in one input (e.g. X with 4 columns) | One number box per feature (naming them) |
| One input per feature, including text columns | Number boxes, text boxes or dropdowns |
| Image models that take any picture (how to export) | A picture upload |
Not supported yet
| Kind of model | Why |
|---|---|
| Text mixed with numbers in one input | Text can't share an input with numbers. Export one input per column: Text columns. |
| CatBoost classifiers | CatBoost's ONNX export gives probabilities as {class: probability} pairs, which PeekML can't show; only the predicted class appears. CatBoost regressors work. |
Image models that need one fixed size (e.g. [1, 1, 28, 28]) | PeekML passes pictures as they are. Build the resizing into the model: Image models. |
| Text and language models | These need a tokenizer to turn text into word ids, which isn't supported yet. |
| Models over 100 MB | Too slow to download in a browser. |
scikit-learn classifiers exported without zipmap: False still work, but their probabilities can't be shown, only the predicted class.