#!/usr/bin/env python3
# pyre-strict
from collections import namedtuple
from itertools import cycle
from typing import (
Any,
Callable,
Dict,
Iterable,
List,
NamedTuple,
Optional,
Tuple,
Union,
)
import torch
from captum.attr import IntegratedGradients
from captum.attr._utils.batching import _batched_generator
from captum.insights.attr_vis.attribution_calculation import (
AttributionCalculation,
OutputScore,
)
from captum.insights.attr_vis.config import (
ATTRIBUTION_METHOD_CONFIG,
ATTRIBUTION_NAMES_TO_METHODS,
)
from captum.insights.attr_vis.features import BaseFeature
from captum.insights.attr_vis.server import namedtuple_to_dict
from captum.log import log_usage
from torch import Tensor
from torch.nn import Module
_CONTEXT_COLAB = "_CONTEXT_COLAB"
_CONTEXT_IPYTHON = "_CONTEXT_IPYTHON"
_CONTEXT_NONE = "_CONTEXT_NONE"
def _get_context() -> str:
"""Determine the most specific context that we're in.
Implementation from TensorBoard: https://git.io/JvObD.
Returns:
_CONTEXT_COLAB: If in Colab with an IPython notebook context.
_CONTEXT_IPYTHON: If not in Colab, but we are in an IPython notebook
context (e.g., from running `jupyter notebook` at the command
line).
_CONTEXT_NONE: Otherwise (e.g., by running a Python script at the
command-line or using the `ipython` interactive shell).
"""
# In Colab, the `google.colab` module is available, but the shell
# returned by `IPython.get_ipython` does not have a `get_trait`
# method.
try:
# To avoid fbsource//third-party/pypi/google-cloud-pubsub:google-cloud-pubsub
# which will cause "Duplicate extension: grpc/_cython/cygrpc.so!"
# @manual
# pyre-fixme[21]: Could not find module `google.colab`.
import google.colab # noqa: F401
import IPython
except ImportError:
pass
else:
if IPython.get_ipython() is not None:
# We'll assume that we're in a Colab notebook context.
return _CONTEXT_COLAB
# In an IPython command line shell or Jupyter notebook, we can
# directly query whether we're in a notebook context.
try:
import IPython
except ImportError:
pass
else:
ipython = IPython.get_ipython()
if ipython is not None and ipython.has_trait("kernel"):
return _CONTEXT_IPYTHON
# Otherwise, we're not in a known notebook context.
return _CONTEXT_NONE
# pyre-fixme[4]: Attribute annotation cannot be `Any`.
# pyre-fixme[2]: Parameter annotation cannot be `Any`.
VisualizationOutput = namedtuple(
"VisualizationOutput", "feature_outputs actual predicted active_index model_index"
)
# pyre-fixme[4]: Attribute annotation cannot be `Any`.
# pyre-fixme[2]: Parameter annotation cannot be `Any`.
Contribution = namedtuple("Contribution", "name percent")
# pyre-fixme[4]: Attribute annotation cannot be `Any`.
# pyre-fixme[2]: Parameter annotation cannot be `Any`.
SampleCache = namedtuple("SampleCache", "inputs additional_forward_args label")
class FilterConfig(NamedTuple):
attribution_method: str = IntegratedGradients.get_name()
# issue with mypy github.com/python/mypy/issues/8376
attribution_arguments: Dict[str, Any] = {
arg: config.value # type: ignore
for arg, config in ATTRIBUTION_METHOD_CONFIG[
IntegratedGradients.get_name()
].params.items()
}
prediction: str = "all"
classes: List[str] = []
num_examples: int = 4
[docs]
class Batch:
def __init__(
self,
inputs: Union[Tensor, Tuple[Tensor, ...]],
labels: Optional[Tensor],
# pyre-fixme[2]: Parameter must be annotated.
additional_args=None,
) -> None:
r"""
Constructs batch of inputs to be attributed and visualized.
Args:
inputs (Tensor or tuple[Tensor, ...]): Batch of inputs for a model.
These may be either a Tensor or tuple of tensors. Each tensor
must correspond to a feature for AttributionVisualizer, and
the corresponding input transform function of the feature
is applied to each input tensor prior to passing it to the
model. It is assumed that the first dimension of each
input tensor corresponds to the number of examples
(batch size) and is aligned for all input tensors.
labels (Tensor): Tensor containing correct labels for input examples.
This must be a 1D tensor with length matching the first
dimension of each input tensor.
additional_args (tuple, optional): If the forward function
requires additional arguments other than the inputs for
which attributions should not be computed, this argument
can be provided. It must be either a single additional
argument of a Tensor or arbitrary (non-tuple) type or a
tuple containing multiple additional arguments including
tensors or any arbitrary python types. These arguments
are provided to ``forward_func`` in order following the
arguments in inputs.
For a tensor, the first dimension of the tensor must
correspond to the number of examples.
"""
self.inputs = inputs
self.labels = labels
# pyre-fixme[4]: Attribute must be annotated.
self.additional_args = additional_args
[docs]
class AttributionVisualizer:
def __init__(
self,
models: Union[List[Module], Module],
classes: List[str],
features: Union[List[BaseFeature], BaseFeature],
dataset: Iterable[Batch],
# pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
score_func: Optional[Callable] = None,
use_label_for_attr: bool = True,
) -> None:
r"""
Args:
models (torch.nn.Module): One or more PyTorch modules (models) for
attribution visualization.
classes (list[str]): List of strings corresponding to the names of
classes for classification.
features (list[BaseFeature]): List of BaseFeatures, which correspond
to input arguments to the model. Each feature object defines
relevant transformations for converting to model input,
constructing baselines, and visualizing. The length of the
features list should exactly match the number of (tensor)
arguments expected by the given model.
For instance, an image classifier should only provide
a single BaseFeature, while a multimodal classifier may
provide a list of features, each corresponding to a different
tensor input and potentially different modalities.
dataset (Iterable of Batch): Defines the dataset to visualize attributions
for. This must be an iterable of batch objects, each of which
may contain multiple input examples.
score_func (Callable, optional): This function is applied to the model
output to obtain the score for each class. For instance,
this function could be the softmax or final non-linearity
of the network, applied to the model output. The indices
of the second dimension of the output should correspond
to the class names provided. If None, the model outputs
are taken directly and assumed to correspond to the
class scores.
Default: None
use_label_for_attr (bool, optional): If true, the class index is passed
to the relevant attribution method. This is necessary in most
cases where there is an output neuron corresponding to each
class. When the model output is a scalar and class index
(e.g. positive, negative) is inferred from the output value,
this argument should be False.
Default: True
"""
if not isinstance(models, List):
models = [models]
if not isinstance(features, List):
features = [features]
self.classes = classes
self.features = features
self.dataset = dataset
self.models = models
self.attribution_calculation = AttributionCalculation(
models, classes, features, score_func, use_label_for_attr
)
self._outputs: List[VisualizationOutput] = []
self._config = FilterConfig(prediction="all", classes=[], num_examples=4)
# pyre-fixme[4]: Attribute must be annotated.
self._dataset_iter = iter(dataset)
self._dataset_cache: List[Batch] = []
def _calculate_attribution_from_cache(
self, input_index: int, model_index: int, target: Optional[Tensor]
) -> Optional[VisualizationOutput]:
c = self._outputs[input_index][1]
result = self._calculate_vis_output(
c.inputs,
c.additional_forward_args,
c.label,
torch.tensor(target),
model_index,
)
if not result:
return None
return result[0]
# pyre-fixme[2]: Parameter must be annotated.
def _update_config(self, settings) -> None:
self._config = FilterConfig(
attribution_method=settings["attribution_method"],
attribution_arguments=settings["arguments"],
prediction=settings["prediction"],
classes=settings["classes"],
num_examples=4,
)
@log_usage()
def render(self, debug: bool = True) -> None:
from captum.insights.attr_vis.widget.widget import CaptumInsights
from IPython.display import display
widget = CaptumInsights(visualizer=self)
display(widget)
if debug:
display(widget.out)
@log_usage()
# pyre-fixme[3]: Return type must be annotated.
def serve(
self,
blocking: bool = False,
debug: bool = False,
# pyre-fixme[2]: Parameter must be annotated.
port=None,
bind_all: bool = False,
):
context = _get_context()
if context == _CONTEXT_COLAB:
return self._serve_colab(blocking=blocking, debug=debug, port=port)
else:
return self._serve(
blocking=blocking, debug=debug, port=port, bind_all=bind_all
)
# pyre-fixme[3]: Return type must be annotated.
def _serve(
self,
blocking: bool = False,
debug: bool = False,
# pyre-fixme[2]: Parameter must be annotated.
port=None,
bind_all: bool = False,
):
from captum.insights.attr_vis.server import start_server
return start_server(
self, blocking=blocking, debug=debug, _port=port, bind_all=bind_all
)
def _serve_colab(
self,
blocking: bool = False,
debug: bool = False,
# pyre-fixme[2]: Parameter must be annotated.
port=None,
) -> None:
import ipywidgets as widgets
from captum.insights.attr_vis.server import start_server
from IPython.display import display, HTML
# TODO: Output widget only captures beginning of server logs. It seems
# the context manager isn't respected when the web server is run on a
# separate thread. We should fix to display entirety of the logs
out = widgets.Output()
with out:
port = start_server(self, blocking=blocking, debug=debug, _port=port)
shell = """
<div id="root"></div>
<script>
(function() {
document.querySelector("base").href = "http://localhost:%PORT%";
function reloadScriptsAndCSS(root) {
// Referencing TensorBoard's method for reloading scripts,
// we remove and reinsert each script
for (const script of root.querySelectorAll("script")) {
const newScript = document.createElement("script");
newScript.type = script.type;
if (script.src) {
newScript.src = script.src;
}
if (script.textContent) {
newScript.textContent = script.textContent;
}
root.appendChild(newScript);
script.remove();
}
// A similar method is used to reload styles
for (const link of root.querySelectorAll("link")) {
const newLink = document.createElement("link");
newLink.rel = link.rel;
newLink.href = link.href;
document.querySelector("head").appendChild(newLink);
link.remove();
}
}
const root = document.getElementById("root");
fetch(".")
.then(x => x.text())
.then(html => void (root.innerHTML = html))
.then(() => reloadScriptsAndCSS(root));
})();
</script>
""".replace(
"%PORT%", str(port)
)
html = HTML(shell)
display(html)
display(out)
def _predictions_matches_labels(
self, predicted_scores: List[OutputScore], labels: Union[str, List[str]]
) -> bool:
if len(predicted_scores) == 0:
return False
predicted_label = predicted_scores[0].label
if isinstance(labels, List):
return predicted_label in labels
return labels == predicted_label
def _should_keep_prediction(
self, predicted_scores: List[OutputScore], actual_label: Optional[OutputScore]
) -> bool:
# filter by class
if len(self._config.classes) != 0:
if not self._predictions_matches_labels(
predicted_scores, self._config.classes
):
return False
if not actual_label:
return True
# filter by accuracy
label_name = actual_label.label
if self._config.prediction == "all":
pass
elif self._config.prediction == "correct":
if not self._predictions_matches_labels(predicted_scores, label_name):
return False
elif self._config.prediction == "incorrect":
if self._predictions_matches_labels(predicted_scores, label_name):
return False
else:
raise Exception(f"Invalid prediction config: {self._config.prediction}")
return True
def _calculate_vis_output(
self,
# pyre-fixme[2]: Parameter must be annotated.
inputs,
# pyre-fixme[2]: Parameter must be annotated.
additional_forward_args,
# pyre-fixme[2]: Parameter must be annotated.
label,
# pyre-fixme[2]: Parameter must be annotated.
target=None,
# pyre-fixme[2]: Parameter must be annotated.
single_model_index=None,
) -> Optional[List[VisualizationOutput]]:
# Use all models, unless the user wants to render data for a particular one
models_used = (
[self.models[single_model_index]]
if single_model_index is not None
else self.models
)
results = []
# pyre-fixme[6]: For 1st argument expected `Iterable[_T]` but got
# `Union[List[Any], Module]`.
for model_index, model in enumerate(models_used):
# Get list of model visualizations for each input
actual_label_output = None
if label is not None and len(label) > 0:
label_index = int(label[0])
actual_label_output = OutputScore(
score=100, index=label_index, label=self.classes[label_index]
)
(
predicted_scores,
baselines,
transformed_inputs,
) = self.attribution_calculation.calculate_predicted_scores(
inputs, additional_forward_args, model
)
# Filter based on UI configuration
if actual_label_output is None or not self._should_keep_prediction(
predicted_scores, actual_label_output
):
continue
if target is None:
target = (
predicted_scores[0].index if len(predicted_scores) > 0 else None
)
# attributions are given per input*
# inputs given to the model are described via `self.features`
#
# *an input contains multiple features that represent it
# e.g. all the pixels that describe an image is an input
attrs_per_feature = self.attribution_calculation.calculate_attribution(
baselines,
transformed_inputs,
additional_forward_args,
target,
self._config.attribution_method,
self._config.attribution_arguments,
model,
)
net_contrib = self.attribution_calculation.calculate_net_contrib(
attrs_per_feature
)
# the features per input given
features_per_input = [
feature.visualize(attr, data, contrib)
for feature, attr, data, contrib in zip(
# pyre-fixme[6]: For 1st argument expected
# `Iterable[Variable[_T1]]` but got `Union[List[BaseFeature],
# BaseFeature]`.
self.features,
attrs_per_feature,
inputs,
net_contrib,
)
]
results.append(
VisualizationOutput(
feature_outputs=features_per_input,
actual=actual_label_output,
predicted=predicted_scores,
active_index=(
target if target is not None else actual_label_output.index
),
# Even if we only iterated over one model, the index should be fixed
# to show the index the model would have had in the list
model_index=(
single_model_index
if single_model_index is not None
else model_index
),
)
)
return results if results else None
def _get_outputs(self) -> List[Tuple[List[VisualizationOutput], SampleCache]]:
# If we run out of new batches, then we need to
# display data which was already shown before.
# However, since the dataset given to us is a generator,
# we can't reset it to return to the beginning.
# Because of this, we store a small cache of stale
# data, and iterate on it after the main generator
# stops returning new batches.
try:
batch_data = next(self._dataset_iter)
self._dataset_cache.append(batch_data)
if len(self._dataset_cache) > self._config.num_examples:
self._dataset_cache.pop(0)
except StopIteration:
self._dataset_iter = cycle(self._dataset_cache)
batch_data = next(self._dataset_iter)
vis_outputs = []
# Type ignore for issue with passing union to function taking generic
# https://github.com/python/mypy/issues/1533
for (
inputs,
additional_forward_args,
label,
) in _batched_generator( # type: ignore
inputs=batch_data.inputs,
additional_forward_args=batch_data.additional_args,
target_ind=batch_data.labels,
internal_batch_size=1, # should be 1 until we have batch label support
):
output = self._calculate_vis_output(inputs, additional_forward_args, label)
if output is not None:
cache = SampleCache(inputs, additional_forward_args, label)
vis_outputs.append((output, cache))
return vis_outputs
@log_usage()
# pyre-fixme[3]: Return type must be annotated.
def visualize(self):
self._outputs = []
while len(self._outputs) < self._config.num_examples:
# pyre-fixme[6]: For 1st argument expected
# `Iterable[VisualizationOutput]` but got
# `List[Tuple[List[VisualizationOutput], SampleCache]]`.
self._outputs.extend(self._get_outputs())
return [o[0] for o in self._outputs]
def get_insights_config(self) -> Dict[str, Any]:
return {
"classes": self.classes,
"methods": list(ATTRIBUTION_NAMES_TO_METHODS.keys()),
"method_arguments": namedtuple_to_dict(
{k: v.params for (k, v) in ATTRIBUTION_METHOD_CONFIG.items()}
),
"selected_method": self._config.attribution_method,
}
