AutoTrain Engine

Upload datasets, train tabular machine-learning models, import existing models, evaluate predictions, and connect production applications to stable prediction endpoints.

Standard workflow

1. Sign in and upload a dataset or choose Bring Your Own Model.
2. Configure preprocessing, task, target, and model parameters.
3. Train or upload the model and create a Model Testing tab.
4. Validate single or batch predictions.
5. On Basic or Advanced, use the endpoint shown in Testing Console from your application.

Task guides

Use cases and dataset references

Use the public experiment library to see how a decision question becomes a target, preprocessing plan, fixed benchmark, interpretation, and limitation review. Dataset pages then connect the source fields to the matching AutoTrain guide.

Supported models

Image model files

Complete .onnx, .keras, .h5, and .tflite models are inspected automatically. A .pb upload must be a frozen GraphDef with one image input and a terminal output. Upload a SavedModel directory as a .tf or .zip archive containing saved_model.pb and its variables directory.

An H5 file created with model.save_weights() must be paired with the matching architecture JSON exported by model.to_json(). For a single portable Keras artifact, export with model.save("model.keras").

Image preprocessing

Image models support a fixed input width and height, 1/3/4 channels, RGB/BGR/grayscale color mode, and pixel ranges of 0–1, -1–1, or original values. These settings must match model training.

Tabular prediction

curl -X POST "https://api.autotrain.app/api/autotrain/jobs/JOB_ID/predict" \
  -H "Content-Type: application/json" \
  -d '{"data":[{"age":42,"income":58000}]}'

{
  "success": true,
  "task": "classification",
  "predictions": [{
    "predicted_class": "approved",
    "confidence_score": 0.94,
    "probabilities": {"approved": 0.94, "rejected": 0.06}
  }]
}

Input schema

The schema returns feature columns, task, model type, target column, request format, and output fields.

Image prediction

curl -X POST "https://api.autotrain.app/api/autotrain/jobs/JOB_ID/predict-image" \
  -F "file=@sample.jpg"

Accepted browser formats are PNG, JPEG, and WebP. The response depends on the image task.

Image segmentation

Single-class masks use a transparent red overlay by default. Multiclass colors can be configured in Testing Console and are sent as a JSON array in the optional class_colors form field.

curl -X POST "https://api.autotrain.app/api/autotrain/jobs/JOB_ID/predict-image" \
  -F "file=@sample.jpg" \
  -F 'class_colors=["#ef4444","#22c55e","#3b82f6"]'

Image classification

The response returns the detected class, class index, confidence, and either a label-to-probability object or score array. Provide class labels when importing a model to receive named results.

{"predicted_class":"cat","class_index":1,"confidence":0.982,"probabilities":{"dog":0.018,"cat":0.982}}

Object detection

Detection models must expose one decoded output tensor with shape [batch, detections, 6]. Each row is:

[x1, y1, x2, y2, confidence, class_id]

Coordinates may be normalized to 0–1 or expressed in model-input pixels. A single-box model may return [batch, 4]. Detections below 25% confidence are omitted.

{
  "annotated_png": "data:image/png;base64,...",
  "detection_count": 1,
  "detections": [{"class_name":"person","confidence":0.97,"box":[31,18,210,248]}]
}

External deployment platforms

AutoTrain Engine provides managed training, testing, exports, and prediction endpoints. When a project needs a public demo, dedicated GPU, custom serving stack, or a third-party computer-vision workflow, the exported model can also be deployed through a specialized provider.

Platform comparison

Prices are indicative comparison points, not provider quotes. A continuously running GPU can range from a few hundred to several thousand US dollars per month depending on hardware, region, availability, and discounts.

Quick selection

• Use AutoTrain ML for no-code tabular experiments, managed model testing, and stable prediction endpoints.
• Use Hugging Face Spaces for an interactive demo with Gradio or Streamlit.
• Use HF Inference Endpoints or Baseten for a managed production API.
• Use Replicate for rapid generative-model API integration.
• Use RunPod when GPU cost and runtime control matter more than operational simplicity.
• Use Modal for serverless Python jobs and intermittent inference.
• Use Roboflow for a complete computer-vision dataset and deployment workflow.

General export workflow

1. Train and validate the model in AutoTrain ML.
2. Export the model artifact and record its exact preprocessing configuration, feature schema, class labels, and runtime version.
3. Package the model using the provider's supported repository, container, function, or SDK format.
4. Configure CPU/GPU, scaling, secrets, storage, and access controls.
5. Run representative validation inputs and compare outputs with AutoTrain Engine before routing production traffic.

Error responses

Limits and security

• Free includes one dashboard testing tab and no inference API access.
• Basic includes three testing tabs and three inference API endpoints; Advanced includes ten of each.
• Testing has no per-test counter while an eligible plan is active. Protective fair-use and hourly rate limits may still apply.
• When a paid subscription expires, projects above the Free limit are preserved but sealed and their inference API endpoints return an access error until renewal.
• Uploaded BYOM artifacts remain stored until deleted by the account owner.
• Images submitted for prediction are processed in memory and are not saved.
• Keep job IDs private and apply your own server-side access policy when exposing an endpoint to end users.
• Do not embed private credentials or unrestricted business logic in browser code.

Support: hello@autotrain.app