strexp / captum /attr /_utils /visualization.py
markytools's picture
added strexp
d61b9c7
raw
history blame
22.9 kB
#!/usr/bin/env python3
import warnings
from enum import Enum
from typing import Any, Iterable, List, Tuple, Union
import numpy as np
from matplotlib import pyplot as plt
from matplotlib.colors import LinearSegmentedColormap
from matplotlib.figure import Figure
from matplotlib.pyplot import axis, figure
from mpl_toolkits.axes_grid1 import make_axes_locatable
from numpy import ndarray
try:
from IPython.core.display import display, HTML
HAS_IPYTHON = True
except ImportError:
HAS_IPYTHON = False
class ImageVisualizationMethod(Enum):
heat_map = 1
blended_heat_map = 2
original_image = 3
masked_image = 4
alpha_scaling = 5
class VisualizeSign(Enum):
positive = 1
absolute_value = 2
negative = 3
all = 4
def _prepare_image(attr_visual: ndarray):
return np.clip(attr_visual.astype(int), 0, 255)
def _normalize_scale(attr: ndarray, scale_factor: float):
assert scale_factor != 0, "Cannot normalize by scale factor = 0"
if abs(scale_factor) < 1e-5:
warnings.warn(
"Attempting to normalize by value approximately 0, visualized results"
"may be misleading. This likely means that attribution values are all"
"close to 0."
)
attr_norm = attr / scale_factor
return np.clip(attr_norm, -1, 1)
def _cumulative_sum_threshold(values: ndarray, percentile: Union[int, float]):
# given values should be non-negative
assert percentile >= 0 and percentile <= 100, (
"Percentile for thresholding must be " "between 0 and 100 inclusive."
)
sorted_vals = np.sort(values.flatten())
cum_sums = np.cumsum(sorted_vals)
threshold_id = np.where(cum_sums >= cum_sums[-1] * 0.01 * percentile)[0][0]
return sorted_vals[threshold_id]
def _normalize_image_attr(
attr: ndarray, sign: str, outlier_perc: Union[int, float] = 2
):
attr_combined = np.sum(attr, axis=2)
# Choose appropriate signed values and rescale, removing given outlier percentage.
if VisualizeSign[sign] == VisualizeSign.all:
threshold = _cumulative_sum_threshold(np.abs(attr_combined), 100 - outlier_perc)
elif VisualizeSign[sign] == VisualizeSign.positive:
attr_combined = (attr_combined > 0) * attr_combined
threshold = _cumulative_sum_threshold(attr_combined, 100 - outlier_perc)
elif VisualizeSign[sign] == VisualizeSign.negative:
attr_combined = (attr_combined < 0) * attr_combined
threshold = -1 * _cumulative_sum_threshold(
np.abs(attr_combined), 100 - outlier_perc
)
elif VisualizeSign[sign] == VisualizeSign.absolute_value:
attr_combined = np.abs(attr_combined)
threshold = _cumulative_sum_threshold(attr_combined, 100 - outlier_perc)
else:
raise AssertionError("Visualize Sign type is not valid.")
return _normalize_scale(attr_combined, threshold)
def visualize_image_attr(
attr: ndarray,
original_image: Union[None, ndarray] = None,
method: str = "heat_map",
sign: str = "absolute_value",
plt_fig_axis: Union[None, Tuple[figure, axis]] = None,
outlier_perc: Union[int, float] = 2,
cmap: Union[None, str] = None,
alpha_overlay: float = 0.5,
show_colorbar: bool = False,
title: Union[None, str] = None,
fig_size: Tuple[int, int] = (6, 6),
use_pyplot: bool = True,
):
r"""
Visualizes attribution for a given image by normalizing attribution values
of the desired sign (positive, negative, absolute value, or all) and displaying
them using the desired mode in a matplotlib figure.
Args:
attr (numpy.array): Numpy array corresponding to attributions to be
visualized. Shape must be in the form (H, W, C), with
channels as last dimension. Shape must also match that of
the original image if provided.
original_image (numpy.array, optional): Numpy array corresponding to
original image. Shape must be in the form (H, W, C), with
channels as the last dimension. Image can be provided either
with float values in range 0-1 or int values between 0-255.
This is a necessary argument for any visualization method
which utilizes the original image.
Default: None
method (string, optional): Chosen method for visualizing attribution.
Supported options are:
1. `heat_map` - Display heat map of chosen attributions
2. `blended_heat_map` - Overlay heat map over greyscale
version of original image. Parameter alpha_overlay
corresponds to alpha of heat map.
3. `original_image` - Only display original image.
4. `masked_image` - Mask image (pixel-wise multiply)
by normalized attribution values.
5. `alpha_scaling` - Sets alpha channel of each pixel
to be equal to normalized attribution value.
Default: `heat_map`
sign (string, optional): Chosen sign of attributions to visualize. Supported
options are:
1. `positive` - Displays only positive pixel attributions.
2. `absolute_value` - Displays absolute value of
attributions.
3. `negative` - Displays only negative pixel attributions.
4. `all` - Displays both positive and negative attribution
values. This is not supported for `masked_image` or
`alpha_scaling` modes, since signed information cannot
be represented in these modes.
Default: `absolute_value`
plt_fig_axis (tuple, optional): Tuple of matplotlib.pyplot.figure and axis
on which to visualize. If None is provided, then a new figure
and axis are created.
Default: None
outlier_perc (float or int, optional): Top attribution values which
correspond to a total of outlier_perc percentage of the
total attribution are set to 1 and scaling is performed
using the minimum of these values. For sign=`all`, outliers
and scale value are computed using absolute value of
attributions.
Default: 2
cmap (string, optional): String corresponding to desired colormap for
heatmap visualization. This defaults to "Reds" for negative
sign, "Blues" for absolute value, "Greens" for positive sign,
and a spectrum from red to green for all. Note that this
argument is only used for visualizations displaying heatmaps.
Default: None
alpha_overlay (float, optional): Alpha to set for heatmap when using
`blended_heat_map` visualization mode, which overlays the
heat map over the greyscaled original image.
Default: 0.5
show_colorbar (boolean, optional): Displays colorbar for heatmap below
the visualization. If given method does not use a heatmap,
then a colormap axis is created and hidden. This is
necessary for appropriate alignment when visualizing
multiple plots, some with colorbars and some without.
Default: False
title (string, optional): Title string for plot. If None, no title is
set.
Default: None
fig_size (tuple, optional): Size of figure created.
Default: (6,6)
use_pyplot (boolean, optional): If true, uses pyplot to create and show
figure and displays the figure after creating. If False,
uses Matplotlib object oriented API and simply returns a
figure object without showing.
Default: True.
Returns:
2-element tuple of **figure**, **axis**:
- **figure** (*matplotlib.pyplot.figure*):
Figure object on which visualization
is created. If plt_fig_axis argument is given, this is the
same figure provided.
- **axis** (*matplotlib.pyplot.axis*):
Axis object on which visualization
is created. If plt_fig_axis argument is given, this is the
same axis provided.
Examples::
>>> # ImageClassifier takes a single input tensor of images Nx3x32x32,
>>> # and returns an Nx10 tensor of class probabilities.
>>> net = ImageClassifier()
>>> ig = IntegratedGradients(net)
>>> # Computes integrated gradients for class 3 for a given image .
>>> attribution, delta = ig.attribute(orig_image, target=3)
>>> # Displays blended heat map visualization of computed attributions.
>>> _ = visualize_image_attr(attribution, orig_image, "blended_heat_map")
"""
# Create plot if figure, axis not provided
if plt_fig_axis is not None:
plt_fig, plt_axis = plt_fig_axis
else:
if use_pyplot:
plt_fig, plt_axis = plt.subplots(figsize=fig_size)
else:
plt_fig = Figure(figsize=fig_size)
plt_axis = plt_fig.subplots()
if original_image is not None:
if np.max(original_image) <= 1.0:
original_image = _prepare_image(original_image * 255)
else:
assert (
ImageVisualizationMethod[method] == ImageVisualizationMethod.heat_map
), "Original Image must be provided for any visualization other than heatmap."
# Remove ticks and tick labels from plot.
plt_axis.xaxis.set_ticks_position("none")
plt_axis.yaxis.set_ticks_position("none")
plt_axis.set_yticklabels([])
plt_axis.set_xticklabels([])
plt_axis.grid(b=False)
heat_map = None
# Show original image
if ImageVisualizationMethod[method] == ImageVisualizationMethod.original_image:
if len(original_image.shape) > 2 and original_image.shape[2] == 1:
original_image = np.squeeze(original_image, axis=2)
plt_axis.imshow(original_image)
else:
# Choose appropriate signed attributions and normalize.
norm_attr = _normalize_image_attr(attr, sign, outlier_perc)
# Set default colormap and bounds based on sign.
if VisualizeSign[sign] == VisualizeSign.all:
default_cmap = LinearSegmentedColormap.from_list(
"RdWhGn", ["red", "white", "green"]
)
vmin, vmax = -1, 1
elif VisualizeSign[sign] == VisualizeSign.positive:
default_cmap = "Greens"
vmin, vmax = 0, 1
elif VisualizeSign[sign] == VisualizeSign.negative:
default_cmap = "Reds"
vmin, vmax = 0, 1
elif VisualizeSign[sign] == VisualizeSign.absolute_value:
default_cmap = "Blues"
vmin, vmax = 0, 1
else:
raise AssertionError("Visualize Sign type is not valid.")
cmap = cmap if cmap is not None else default_cmap
# Show appropriate image visualization.
if ImageVisualizationMethod[method] == ImageVisualizationMethod.heat_map:
heat_map = plt_axis.imshow(norm_attr, cmap=cmap, vmin=vmin, vmax=vmax)
elif (
ImageVisualizationMethod[method]
== ImageVisualizationMethod.blended_heat_map
):
plt_axis.imshow(np.mean(original_image, axis=2), cmap="gray")
heat_map = plt_axis.imshow(
norm_attr, cmap=cmap, vmin=vmin, vmax=vmax, alpha=alpha_overlay
)
elif ImageVisualizationMethod[method] == ImageVisualizationMethod.masked_image:
assert VisualizeSign[sign] != VisualizeSign.all, (
"Cannot display masked image with both positive and negative "
"attributions, choose a different sign option."
)
plt_axis.imshow(
_prepare_image(original_image * np.expand_dims(norm_attr, 2))
)
elif ImageVisualizationMethod[method] == ImageVisualizationMethod.alpha_scaling:
assert VisualizeSign[sign] != VisualizeSign.all, (
"Cannot display alpha scaling with both positive and negative "
"attributions, choose a different sign option."
)
plt_axis.imshow(
np.concatenate(
[
original_image,
_prepare_image(np.expand_dims(norm_attr, 2) * 255),
],
axis=2,
)
)
else:
raise AssertionError("Visualize Method type is not valid.")
# Add colorbar. If given method is not a heatmap and no colormap is relevant,
# then a colormap axis is created and hidden. This is necessary for appropriate
# alignment when visualizing multiple plots, some with heatmaps and some
# without.
if show_colorbar:
axis_separator = make_axes_locatable(plt_axis)
colorbar_axis = axis_separator.append_axes("bottom", size="5%", pad=0.1)
if heat_map:
plt_fig.colorbar(heat_map, orientation="horizontal", cax=colorbar_axis)
else:
colorbar_axis.axis("off")
if title:
plt_axis.set_title(title)
if use_pyplot:
plt.show()
return plt_fig, plt_axis
def visualize_image_attr_multiple(
attr: ndarray,
original_image: Union[None, ndarray],
methods: List[str],
signs: List[str],
titles: Union[None, List[str]] = None,
fig_size: Tuple[int, int] = (8, 6),
use_pyplot: bool = True,
**kwargs: Any,
):
r"""
Visualizes attribution using multiple visualization methods displayed
in a 1 x k grid, where k is the number of desired visualizations.
Args:
attr (numpy.array): Numpy array corresponding to attributions to be
visualized. Shape must be in the form (H, W, C), with
channels as last dimension. Shape must also match that of
the original image if provided.
original_image (numpy.array, optional): Numpy array corresponding to
original image. Shape must be in the form (H, W, C), with
channels as the last dimension. Image can be provided either
with values in range 0-1 or 0-255. This is a necessary
argument for any visualization method which utilizes
the original image.
methods (list of strings): List of strings of length k, defining method
for each visualization. Each method must be a valid
string argument for method to visualize_image_attr.
signs (list of strings): List of strings of length k, defining signs for
each visualization. Each sign must be a valid
string argument for sign to visualize_image_attr.
titles (list of strings, optional): List of strings of length k, providing
a title string for each plot. If None is provided, no titles
are added to subplots.
Default: None
fig_size (tuple, optional): Size of figure created.
Default: (8, 6)
use_pyplot (boolean, optional): If true, uses pyplot to create and show
figure and displays the figure after creating. If False,
uses Matplotlib object oriented API and simply returns a
figure object without showing.
Default: True.
**kwargs (Any, optional): Any additional arguments which will be passed
to every individual visualization. Such arguments include
`show_colorbar`, `alpha_overlay`, `cmap`, etc.
Returns:
2-element tuple of **figure**, **axis**:
- **figure** (*matplotlib.pyplot.figure*):
Figure object on which visualization
is created. If plt_fig_axis argument is given, this is the
same figure provided.
- **axis** (*matplotlib.pyplot.axis*):
Axis object on which visualization
is created. If plt_fig_axis argument is given, this is the
same axis provided.
Examples::
>>> # ImageClassifier takes a single input tensor of images Nx3x32x32,
>>> # and returns an Nx10 tensor of class probabilities.
>>> net = ImageClassifier()
>>> ig = IntegratedGradients(net)
>>> # Computes integrated gradients for class 3 for a given image .
>>> attribution, delta = ig.attribute(orig_image, target=3)
>>> # Displays original image and heat map visualization of
>>> # computed attributions side by side.
>>> _ = visualize_image_attr_multiple(attribution, orig_image,
>>> ["original_image", "heat_map"], ["all", "positive"])
"""
assert len(methods) == len(signs), "Methods and signs array lengths must match."
if titles is not None:
assert len(methods) == len(titles), (
"If titles list is given, length must " "match that of methods list."
)
if use_pyplot:
plt_fig = plt.figure(figsize=fig_size)
else:
plt_fig = Figure(figsize=fig_size)
plt_axis = plt_fig.subplots(1, len(methods))
# When visualizing one
if len(methods) == 1:
plt_axis = [plt_axis]
for i in range(len(methods)):
visualize_image_attr(
attr,
original_image=original_image,
method=methods[i],
sign=signs[i],
plt_fig_axis=(plt_fig, plt_axis[i]),
use_pyplot=False,
title=titles[i] if titles else None,
**kwargs,
)
plt_fig.tight_layout()
if use_pyplot:
plt.show()
return plt_fig, plt_axis
# These visualization methods are for text and are partially copied from
# experiments conducted by Davide Testuggine at Facebook.
class VisualizationDataRecord:
r"""
A data record for storing attribution relevant information
"""
__slots__ = [
"word_attributions",
"pred_prob",
"pred_class",
"true_class",
"attr_class",
"attr_score",
"raw_input_ids",
"convergence_score",
]
def __init__(
self,
word_attributions,
pred_prob,
pred_class,
true_class,
attr_class,
attr_score,
raw_input_ids,
convergence_score,
) -> None:
self.word_attributions = word_attributions
self.pred_prob = pred_prob
self.pred_class = pred_class
self.true_class = true_class
self.attr_class = attr_class
self.attr_score = attr_score
self.raw_input_ids = raw_input_ids
self.convergence_score = convergence_score
def _get_color(attr):
# clip values to prevent CSS errors (Values should be from [-1,1])
attr = max(-1, min(1, attr))
if attr > 0:
hue = 120
sat = 75
lig = 100 - int(50 * attr)
else:
hue = 0
sat = 75
lig = 100 - int(-40 * attr)
return "hsl({}, {}%, {}%)".format(hue, sat, lig)
def format_classname(classname):
return '<td><text style="padding-right:2em"><b>{}</b></text></td>'.format(classname)
def format_special_tokens(token):
if token.startswith("<") and token.endswith(">"):
return "#" + token.strip("<>")
return token
def format_tooltip(item, text):
return '<div class="tooltip">{item}\
<span class="tooltiptext">{text}</span>\
</div>'.format(
item=item, text=text
)
def format_word_importances(words, importances):
if importances is None or len(importances) == 0:
return "<td></td>"
assert len(words) <= len(importances)
tags = ["<td>"]
for word, importance in zip(words, importances[: len(words)]):
word = format_special_tokens(word)
color = _get_color(importance)
unwrapped_tag = '<mark style="background-color: {color}; opacity:1.0; \
line-height:1.75"><font color="black"> {word}\
</font></mark>'.format(
color=color, word=word
)
tags.append(unwrapped_tag)
tags.append("</td>")
return "".join(tags)
def visualize_text(
datarecords: Iterable[VisualizationDataRecord], legend: bool = True
) -> "HTML": # In quotes because this type doesn't exist in standalone mode
assert HAS_IPYTHON, (
"IPython must be available to visualize text. "
"Please run 'pip install ipython'."
)
dom = ["<table width: 100%>"]
rows = [
"<tr><th>True Label</th>"
"<th>Predicted Label</th>"
"<th>Attribution Label</th>"
"<th>Attribution Score</th>"
"<th>Word Importance</th>"
]
for datarecord in datarecords:
rows.append(
"".join(
[
"<tr>",
format_classname(datarecord.true_class),
format_classname(
"{0} ({1:.2f})".format(
datarecord.pred_class, datarecord.pred_prob
)
),
format_classname(datarecord.attr_class),
format_classname("{0:.2f}".format(datarecord.attr_score)),
format_word_importances(
datarecord.raw_input_ids, datarecord.word_attributions
),
"<tr>",
]
)
)
if legend:
dom.append(
'<div style="border-top: 1px solid; margin-top: 5px; \
padding-top: 5px; display: inline-block">'
)
dom.append("<b>Legend: </b>")
for value, label in zip([-1, 0, 1], ["Negative", "Neutral", "Positive"]):
dom.append(
'<span style="display: inline-block; width: 10px; height: 10px; \
border: 1px solid; background-color: \
{value}"></span> {label} '.format(
value=_get_color(value), label=label
)
)
dom.append("</div>")
dom.append("".join(rows))
dom.append("</table>")
html = HTML("".join(dom))
display(html)
return html