Types

Mixtrain types add semantics to your data values. That meaning follows a value through the whole platform: mark a dataset column as an Image and you get Image previews in dataset, in SQL query results, as an input or output of models, workflows, and routines, and in evaluation views. Each type also provides convenient helper methods for common tasks, such as downloading a file or converting an image to numpy.

Media Types

File, Image, Video, and Audio reference a file by URL (https://, gs://, or s3://).

File

Base class for file references. Use for generic files. See File for details.

from mixtrain import MixModel, File

class Archiver(MixModel):
    def run(self, data: File) -> dict[str, File]:
        path = data.to_file()              # download to a local path
        archive(path, out="out.zip")
        return {"result": File.from_file("out.zip").save()}

Helpers:

• to_file(path=None) - Download to local file. Returns path. Cached for subsequent calls.
• save() - Upload file to Mixtrain storage and return a saved reference. Saved files can be accessed by other models and workflows.
• save(path) - Upload file to specific path in Mixtrain storage and return a saved reference.

Image

Renders an image inline. Extends File with image-specific helpers. See Image for details.

from mixtrain import MixModel, Image

class Upscaler(MixModel):
    def run(self, image: Image) -> dict[str, Image]:
        arr = image.to_numpy()             # numpy array (H, W, C)
        upscale(arr, out="out.png")
        return {"result": Image.from_file("out.png").save()}

Helpers:

• to_file(path=None) - Download to local file
• to_pil() - Return as PIL.Image object
• to_numpy() - Return as numpy array (H, W, C)

Video

Renders a video with a player. Extends File with video-specific helpers. See Video for details.

from mixtrain import MixFlow, Video

class FrameInterpolator(MixFlow):
    def run(self, video: Video) -> dict[str, Video]:
        frames = [interpolate(f) for f in video.to_frames()]  # numpy arrays (H, W, C)
        write_video(frames, "out.mp4")
        return {"result": Video.from_file("out.mp4").save()}

Helpers:

• to_file(path=None) - Download to local file
• to_cv2() - Return as cv2.VideoCapture
• to_frames() - Yield frames as numpy arrays (H, W, C) in BGR

Audio

Renders an audio player. Extends File with audio-specific helpers. See Audio for details.

from mixtrain import MixFlow, Audio

class Denoiser(MixFlow):
    def run(self, audio: Audio) -> dict[str, Audio]:
        samples, sample_rate = audio.to_numpy()
        denoise(samples, sample_rate, out="out.wav")
        return {"result": Audio.from_file("out.wav").save()}

Helpers:

• to_file(path=None) - Download to local file
• to_numpy() - Return as (samples array, sample_rate) tuple

Model3D

Renders a 3D model viewer. url points to the model file; format is one of glb, gltf, or obj.

from mixtrain import MixModel, Model3D

class MeshGenerator(MixModel):
    def run(self, prompt: str) -> dict[str, Model3D]:
        generate_mesh(prompt, out="out.glb")
        return {"result": Model3D.from_file("out.glb").save()}

Array Types

Numeric vectors and N-dimensional arrays for ML features, robotics, and multimodal data.

Embedding

ML embedding vectors for semantic representations. See Embedding for details.

from mixtrain import MixModel, Embedding

class TextEmbedder(MixModel):
    def run(self, text: str) -> dict[str, Embedding]:
        return {"embedding": Embedding(values=embed(text))}

Tensor

N-dimensional numeric arrays for ML features, robotics, and multimodal data — such as action sequences, joint state, depth maps, or point clouds. See Tensor for details.

from mixtrain import MixModel, Tensor

class Policy(MixModel):
    def run(self, observation: Tensor) -> dict[str, Tensor]:
        actions = predict(observation.to_numpy())   # action sequence (timesteps, joints)
        return {"action": Tensor.from_numpy(actions)}

Helpers:

• Tensor.from_numpy(array, dim_names=None) - Create from a numpy array, capturing its shape and dtype
• to_numpy() - Return as a numpy array

As a dataset column type, tensors in a column should share the same shape and dtype. Reading the column back with to_tensors() restores each row's full N-dimensional shape.

Text Types

Inline text content rendered in the UI:

from mixtrain import Text, Markdown, JSON

Text(content="Processing complete!")              # plain text
Markdown(content="# Results\n- Item 1")           # rich Markdown
JSON(data={"key": "value", "items": [1, 2, 3]})   # syntax-highlighted JSON

See Text, Markdown, and JSON for details.