Quick Start

Installation

Use the package manager pip to install leap-ie.

pip install leap-ie

Generate an API key

To authenticate with the leap-ie library you will need an API key from the leap dashboard. Head over to the dashboard and create an account:

When first signing in, you will be prompted to generate an API key. Keep this key safe as it will only be displayed once.

If you lose your API key, a new one can be generated in the dashboard settings page.

A user can only have one active API key at a time, so generating a new key will invalidate the old one.

Get started

To quickly try out the leap-ie library on a prebuilt model we provide easy access to all image classification torchvision models via leap_ie.vision.models.get_model.

from leap_ie.vision.models import get_model

preprocessing_fn, model, class_list = get_model('resnet18', source='torchvision')

The get_model function takes the name of any image classification torchvision model. We can also automatically pull image classification models from HuggingFace:

from leap_ie.vision.models import get_model

preprocessing_fn, model, class_list = get_model('nateraw/vit-age-classifier', source='huggingface')

The get_model function returns a tuple with 3 values:

preprocessing_fn The preprocessing function used on inputs for inference.
model The model.
class_list List of class names corresponding to the model's output classes.

To run the Leap Interpretability Engine on this model, simply pass these values and some additional configuration to the engine.generate function:

from leap_ie.vision import engine

config = {"leap_api_key": "YOUR_API_KEY"}

results_df, results_dict = engine.generate(
    project_name="leap!",
    model=model,
    class_list=class_list,
    config=config,
    target_classes=[1],
    preprocessing=preprocessing_fn,
)

In the example above we use the following properties:

project_name Used to group multiple runs in the Leap Dashboard. This can be any string.
model The model value output from get_model.
class_list The class list value output from get_model.
config A config dictionary containing your generated Leap API key
target_classes The indexes of classes from the class list you want to analyse with Leap.
preprocessing The preprocessing function output from get_model.

For a full list of arguments and config values see API Reference.

On running this function, the results of your run will start to appear in the Leap Dashboard. Results will show immediately and automatically refresh as the run continues.

We also support logging directly to Weights and Biases. See Integrations.

If you want to analyse results locally, engine.generate also returns a DataFrame and Dictionary:

results_df contains the results of results of prototype, isolation and entanglement analyses including relative paths to image files in the local filesystem. For more information on the different types of results see Concepts.
results_dict contains probability values at each stage of the generation process. This is only required for advanced debugging use cases.

Usage with your own model

To run using your own model, simply replace the values previously returned by the get_model function with your own model, class list and preprocessing function (optional):

from leap_ie.vision import engine

df_results, dict_results = engine.generate(
    project_name="interpretability",
    model=your_model,
    class_list=["hotdog", "not_hotdog"],
    target_classes=[1],
    config={"leap_api_key": "YOUR_LEAP_API_KEY"},
)

Currently we support image classification and segmentation models only. We expect the model to take a batch of images as input, and return a batch of logits (NOT probabilities).

For most models this will work out of the box, but if your model returns something else (e.g. a dictionary, or probabilities) you might have to edit it, or add a wrapper before passing it to engine.generate:


class ModelWrapper(nn.Module):
    def __init__(self, model):
        super().__init__()
        self.model = model

    def forward(self, x):
        x = self.model(x)
        return x["logits"]

model = ModelWrapper(your_model)

Results

The generate function returns a Pandas DataFrame and a Dictionary of Numpy arrays. If you're in a Jupyter notebook, you can view these DataFrame inline using engine.display_df(df_results), but for the best experience we recommend you head to the leap app, or log directly to your weights and biases dashboard.

For more information about the data we return, see Concepts. If used with samples (see Sample Feature Isolation), the DataFrame contains feature isolations for each sample, for the target classes (if provided), or for the top 3 predicted classes.

Supported Frameworks

We support both PyTorch and Tensorflow. Specify your package with the mode parameter, using "tf" for Tensorflow and "pt" for PyTorch.

Please be aware that TensorFlow support in leap-ie is currently experimental and under active development. Generated prototypes may not be as clear as PyTorch, and may require additional configuration tuning.

If using PyTorch, we expect the model to take images to be in channels first format, e.g. of shape [1, channels, height, width].

If Tensorflow we expect channels last, e.g.[1, height, width, channels].

from leap_ie.vision import engine config = {"leap_api_key": "YOUR_API_KEY"} results_df, results_dict = engine.generate( project_name="leap!", model=model, class_list=class_list, config=config, target_classes=[1], preprocessing=preprocessing_fn, )

from leap_ie.vision import engine df_results, dict_results = engine.generate( project_name="interpretability", model=your_model, class_list=["hotdog", "not_hotdog"], target_classes=[1], config={"leap_api_key": "YOUR_LEAP_API_KEY"}, )

class ModelWrapper(nn.Module): def __init__(self, model): super().__init__() self.model = model def forward(self, x): x = self.model(x) return x["logits"] model = ModelWrapper(your_model)