minerva.analysis.model_analysis

Attributes

_plot_tsne_written_dirs

Classes

TSNEAnalysis	Perform t-SNE analysis on the embeddings generated by a model.
_ModelAnalysis	Main interface for model analysis. A model analysis is a post-training

Module Contents

class minerva.analysis.model_analysis.TSNEAnalysis(label_names=None, height=1000, width=1000, text_size=30, title=None, x_axis_title='x', y_axis_title='y', legend_title='Label', output_filename='tsne.png', seed=42, n_components=2, marker_symbols=None, colors=None, write_html=True)

Bases: _ModelAnalysis

Perform t-SNE analysis on the embeddings generated by a model. A t-SNE plot is generated using the embeddings and saved in the path directory. The plot is saved as both PNG and HTML files.

Plot a t-SNE plot of the embeddings generated by a model.

Parameters

label_namesOptional[Dict[Union[int, str], str]], optional

Labels to use for the plot, instead of the original labels in the data (y). The keys are the original labels and the values are the new labels to use in the plot. If None, the original labels are used as they are. By default None

heightint, optional

Height of the figure, by default 1000

widthint, optional

Width of the figure, by default 1000

text_sizeint, optional

Size of font used in plot, by default 30

titlestr, optional

Title of graph, by default None

x_axis_titlestr, optional

Name of x-axis, by default “x”

y_axis_titlestr, optional

Name of y-axis, by default “y”

legend_titlestr, optional

Name for legend title, by default “Label”

output_filenamePathLike, optional

Name of the output file to save the plot as a PNG image file. The file will be saved in the path directory with this name. By default “tsne.png”

seedint, optional

Random seed for t-SNE, by default 42

n_componentsint, optional

Number of components to use in t-SNE, by default 2

marker_symbolsOptional[Dict[str, str]], optional

Dictionary mapping labels to marker symbols. If None, will use default symbols.

colorsOptional[Dict[str, str]], optional

Dictionary mapping labels to color values (hex codes or names). If None, will use plotly’s default color sequence.

write_html: bool, optional

If True, saves the plot as an HTML file with interactive controls. Default is True.

Examples

Create and run a t-SNE analysis for a fine-tuned model on the MotionSense dataset:

>>> analysis = TSNEAnalysis(
        height=1000,
        width=1000,
        legend_title="Activity",
        title=" ",
        output_filename="tsne_analysis.png",
        label_names={
            0: "sit",
            1: "stand",
            2: "walk",
            3: "stair up",
            4: "stair down",
            5: "run",
            6: "stair up and down",
        },
        marker_symbols={
            "sit": "x-open",
            "stand": "cross-open",
            "stair up": "triangle-up-open",
            "stair down": "triangle-down-open",
            "walk": "circle-open",
            "run": "star-open"
        },
        text_size=30,
        x_axis_title='1st Component',
        y_axis_title='2nd Component',
    )

colors = None

compute(model, data)

Run the t-SNE analysis on the provided model and dataset.

This method extracts embeddings from the given model using the provided LightningDataModule, applies t-SNE to reduce the embeddings to the specified number of components, and generates a scatter plot with points colored and marked according to their labels.

The resulting plot is saved as: - A static PNG file (output_filename) in the directory self.path. - An interactive HTML file (if write_html=True).

Parameters

modellightning.LightningModule

A trained PyTorch Lightning model with a .backbone attribute that produces embeddings from the input data.

datalightning.LightningDataModule

A Lightning DataModule providing the dataset to project. Data is extracted using the “predict” split.

Parameters:

• model (lightning.LightningModule)

• data (lightning.LightningDataModule)

height = 1000

label_names = None

legend_title = 'Label'

marker_symbols = None

n_components = 2

output_filename

seed = 42

set_path(path)

Set the output path for saving the plots.

Parameters:

path (minerva.utils.typing.PathLike)

text_size = 30

title = None

width = 1000

write_html = True

x_axis_title = 'x'

y_axis_title = 'y'

Parameters:

• label_names (Optional[Dict[Union[int, str], str]])

• height (int)

• width (int)

• text_size (int)

• title (Optional[str])

• x_axis_title (str)

• y_axis_title (str)

• legend_title (str)

• output_filename (minerva.utils.typing.PathLike)

• seed (int)

• n_components (int)

• marker_symbols (Optional[Dict[str, str]])

• colors (Optional[Dict[str, str]])

• write_html (bool)

class minerva.analysis.model_analysis._ModelAnalysis(path=None)

Main interface for model analysis. A model analysis is a post-training analysis that can be run on a trained model to generate insights about the model’s performance. It has a path attribute that specifies the directory where the analysis results will be saved. The compute method should be implemented by subclasses to perform the actual analysis. All insights generated by the analysis should be saved in the path directory. Note that, differently from Metric, _ModelAnalysis does not return any value. Instead, the results of the analysis should be saved in the path directory. All subclasses of _ModelAnalysis should implement the compute method. Inside a pipeline the path will be automatically set to the pipeline.log_dir attribute.

Parameters:

path (Optional[minerva.utils.typing.PathLike])

_path = None

abstract compute(model, data)

Parameters:

• model (lightning.LightningModule)

• data (lightning.LightningDataModule)

property path

minerva.analysis.model_analysis._plot_tsne_written_dirs