kinairi / mosaic Goto Github PK

Mosaic provides fast and flexible data structures for working with complex machine learning datasets.

Getting Started | What is Mosaic? | Supported Columns | Docs | Contributing | About

pip install mosaicml

Note: some parts of Mosaic rely on optional dependencies. If you know which optional dependencies you'd like to install, you can do so using something like pip install mosaicml[dev,text] instead. See setup.py for a full list of optional dependencies.

Load your dataset into a DataPanel and get going!

from mosaic import DataPanel
dp = DataPanel.from_csv("...")

Mosaic makes it easier for ML practitioners to interact with high-dimensional, multi-modal data. It provides simple abstractions for data inspection, model evaluation and model training supported by efficient and robust IO under the hood.

Mosaic's core contribution is the DataPanel, a simple columnar data abstraction. The Mosaic DataPanel can house columns of arbitrary type – from integers and strings to complex, high-dimensional objects like videos, images, medical volumes and graphs.

DataPanel loads high-dimensional data lazily. A full high-dimensional dataset won't typically fit in memory. Behind the scenes, DataPanel handles this by only materializing these objects when they are needed.

from mosaic import DataPanel, ImageColumn

# Images are NOT read from disk at DataPanel creation...
dp = DataPanel({
    'text': ['The quick brown fox.', 'Jumped over.', 'The lazy dog.'],
    'image': ImageColumn(['fox.png', 'jump.png', 'dog.png']),
    'label': [0, 1, 0]
}) 

# ...only at this point is "fox.png" read from disk
dp["image"][0]

DataPanel supports advanced indexing. Using indexing patterns similar to those of Pandas and NumPy, we can access a subset of a DataPanel's rows and columns.

from mosaic import DataPanel, AbstractColumn
dp = ... # create DataPanel

# Pull a column out of the DataPanel
new_col: ImageColumn = dp["image"]

# Create a new DataPanel from a subset of the columns in an existing one
new_dp: DataPanel = dp[["image", "label"]] 

# Create a new DataPanel from a subset of the rows in an existing one
new_dp: DataPanel = dp[10:20] 
new_dp: DataPanel = dp[np.array([0,2,4,8])]

# Pull a column out of the DataPanel and get a subset of its rows 
new_col: ImageColumn = dp["image"][10:20]

DataPanel supports map, update and filter operations. When training and evaluating our models, we often perform operations on each example in our dataset (e.g. compute a model's prediction on each example, tokenize each sentence, compute a model's embedding for each example) and store them . The DataPanel makes it easy to perform these operations and produce new columns (via DataPanel.map), store the columns alongside the original data (via DataPanel.update), and extract an important subset of the datset (via DataPanel.filter). Under the hood, dataloading is multiprocessed so that costly I/O doesn't bottleneck our computation. Consider the example below where we use update a DataPanel with two new columns holding model predictions and probabilities.

    # A simple evaluation loop using Mosaic 
    dp: DataPane = ... # get DataPane
    model: nn.Module = ... # get the model
    model.to(0).eval() # prepare the model for evaluation

    @torch.no_grad()
    def predict(batch: dict):
        probs = torch.softmax(model(batch["input"].to(0)), dim=-1)
        return {"probs": probs.cpu(), "pred": probs.cpu().argmax(dim=-1)}

    # updated_dp has two new `TensorColumn`s: 1 for probabilities and one
    # for predictions
    updated_dp: DataPane = dp.update(function=predict, batch_size=128, batched=True)

DataPanel is extendable. Mosaic makes it easy for you to make custom column types for our data. The easiest way to do this is by subclassing AbstractCell. Subclasses of AbstractCell are meant to represent one element in one column of a DataPanel. For example, say we want our DataPanel to include a column of videos we have stored on disk. We want these videos to be lazily loaded using scikit-video, so we implement a VideoCell class as follows:

    from mosaic import AbstractCell, CellColumn
    import skvideo.io

    class VideoCell(AbstractCell):
        
        # What information will we eventually  need to materialize the cell? 
        def __init__(filepath: str):
            super().__init__()
            self.filepath = filepath
        
        # How do we actually materialize the cell?
        def get(self):
            return skvideo.io.vread(self.filepath)
        
        # What attributes should be written to disk on `VideoCell.write`?
        @classmethod
        def _state_keys(cls) -> Collection:
            return {"filepath"}

    # We don't need to define a `VideoColumn` class and can instead just
    # create a CellColumn fro a list of `VideoCell`
    vid_column = CellColumn(map(VideoCell, ["vid1.mp4", "vid2.mp4", "vid3.mp4"]))

Mosaic ships with a number of core column types and the list is growing.

Mosaic is being developed at Stanford's Hazy Research Lab. Please reach out to kgoel [at] cs [dot] stanford [dot] edu if you would like to use or contribute to Mosaic.