File size: 14,340 Bytes
d4576ce |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 |
"""Contains classes and methods related to interpretation for components in Gradio."""
from __future__ import annotations
import copy
import math
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Any, Dict, List, Tuple
import numpy as np
from gradio import components, utils
if TYPE_CHECKING: # Only import for type checking (is False at runtime).
from gradio import Interface
class Interpretable(ABC):
def __init__(self) -> None:
self.set_interpret_parameters()
def set_interpret_parameters(self):
"""
Set any parameters for interpretation. Properties can be set here to be
used in get_interpretation_neighbors and get_interpretation_scores.
"""
pass
def get_interpretation_scores(
self, x: Any, neighbors: List[Any] | None, scores: List[float], **kwargs
) -> List:
"""
Arrange the output values from the neighbors into interpretation scores for the interface to render.
Parameters:
x: Input to interface
neighbors: Neighboring values to input x used for interpretation.
scores: Output value corresponding to each neighbor in neighbors
Returns:
Arrangement of interpretation scores for interfaces to render.
"""
return scores
class TokenInterpretable(Interpretable, ABC):
@abstractmethod
def tokenize(self, x: Any) -> Tuple[List, List, None]:
"""
Interprets an input data point x by splitting it into a list of tokens (e.g
a string into words or an image into super-pixels).
"""
return [], [], None
@abstractmethod
def get_masked_inputs(self, tokens: List, binary_mask_matrix: List[List]) -> List:
return []
class NeighborInterpretable(Interpretable, ABC):
@abstractmethod
def get_interpretation_neighbors(self, x: Any) -> Tuple[List, Dict]:
"""
Generates values similar to input to be used to interpret the significance of the input in the final output.
Parameters:
x: Input to interface
Returns: (neighbor_values, interpret_kwargs, interpret_by_removal)
neighbor_values: Neighboring values to input x to compute for interpretation
interpret_kwargs: Keyword arguments to be passed to get_interpretation_scores
"""
return [], {}
async def run_interpret(interface: Interface, raw_input: List):
"""
Runs the interpretation command for the machine learning model. Handles both the "default" out-of-the-box
interpretation for a certain set of UI component types, as well as the custom interpretation case.
Parameters:
raw_input: a list of raw inputs to apply the interpretation(s) on.
"""
if isinstance(interface.interpretation, list): # Either "default" or "shap"
processed_input = [
input_component.preprocess(raw_input[i])
for i, input_component in enumerate(interface.input_components)
]
original_output = await interface.call_function(0, processed_input)
original_output = original_output["prediction"]
if len(interface.output_components) == 1:
original_output = [original_output]
scores, alternative_outputs = [], []
for i, (x, interp) in enumerate(zip(raw_input, interface.interpretation)):
if interp == "default":
input_component = interface.input_components[i]
neighbor_raw_input = list(raw_input)
if isinstance(input_component, TokenInterpretable):
tokens, neighbor_values, masks = input_component.tokenize(x)
interface_scores = []
alternative_output = []
for neighbor_input in neighbor_values:
neighbor_raw_input[i] = neighbor_input
processed_neighbor_input = [
input_component.preprocess(neighbor_raw_input[i])
for i, input_component in enumerate(
interface.input_components
)
]
neighbor_output = await interface.call_function(
0, processed_neighbor_input
)
neighbor_output = neighbor_output["prediction"]
if len(interface.output_components) == 1:
neighbor_output = [neighbor_output]
processed_neighbor_output = [
output_component.postprocess(neighbor_output[i])
for i, output_component in enumerate(
interface.output_components
)
]
alternative_output.append(processed_neighbor_output)
interface_scores.append(
quantify_difference_in_label(
interface, original_output, neighbor_output
)
)
alternative_outputs.append(alternative_output)
scores.append(
input_component.get_interpretation_scores(
raw_input[i],
neighbor_values,
interface_scores,
masks=masks,
tokens=tokens,
)
)
elif isinstance(input_component, NeighborInterpretable):
(
neighbor_values,
interpret_kwargs,
) = input_component.get_interpretation_neighbors(x)
interface_scores = []
alternative_output = []
for neighbor_input in neighbor_values:
neighbor_raw_input[i] = neighbor_input
processed_neighbor_input = [
input_component.preprocess(neighbor_raw_input[i])
for i, input_component in enumerate(
interface.input_components
)
]
neighbor_output = await interface.call_function(
0, processed_neighbor_input
)
neighbor_output = neighbor_output["prediction"]
if len(interface.output_components) == 1:
neighbor_output = [neighbor_output]
processed_neighbor_output = [
output_component.postprocess(neighbor_output[i])
for i, output_component in enumerate(
interface.output_components
)
]
alternative_output.append(processed_neighbor_output)
interface_scores.append(
quantify_difference_in_label(
interface, original_output, neighbor_output
)
)
alternative_outputs.append(alternative_output)
interface_scores = [-score for score in interface_scores]
scores.append(
input_component.get_interpretation_scores(
raw_input[i],
neighbor_values,
interface_scores,
**interpret_kwargs,
)
)
else:
raise ValueError(
f"Component {input_component} does not support interpretation"
)
elif interp == "shap" or interp == "shapley":
try:
import shap # type: ignore
except (ImportError, ModuleNotFoundError):
raise ValueError(
"The package `shap` is required for this interpretation method. Try: `pip install shap`"
)
input_component = interface.input_components[i]
if not isinstance(input_component, TokenInterpretable):
raise ValueError(
"Input component {} does not support `shap` interpretation".format(
input_component
)
)
tokens, _, masks = input_component.tokenize(x)
# construct a masked version of the input
def get_masked_prediction(binary_mask):
assert isinstance(input_component, TokenInterpretable)
masked_xs = input_component.get_masked_inputs(tokens, binary_mask)
preds = []
for masked_x in masked_xs:
processed_masked_input = copy.deepcopy(processed_input)
processed_masked_input[i] = input_component.preprocess(masked_x)
new_output = utils.synchronize_async(
interface.call_function, 0, processed_masked_input
)
new_output = new_output["prediction"]
if len(interface.output_components) == 1:
new_output = [new_output]
pred = get_regression_or_classification_value(
interface, original_output, new_output
)
preds.append(pred)
return np.array(preds)
num_total_segments = len(tokens)
explainer = shap.KernelExplainer(
get_masked_prediction, np.zeros((1, num_total_segments))
)
shap_values = explainer.shap_values(
np.ones((1, num_total_segments)),
nsamples=int(interface.num_shap * num_total_segments),
silent=True,
)
assert shap_values is not None, "SHAP values could not be calculated"
scores.append(
input_component.get_interpretation_scores(
raw_input[i],
None,
shap_values[0].tolist(),
masks=masks,
tokens=tokens,
)
)
alternative_outputs.append([])
elif interp is None:
scores.append(None)
alternative_outputs.append([])
else:
raise ValueError("Unknown intepretation method: {}".format(interp))
return scores, alternative_outputs
elif interface.interpretation: # custom interpretation function
processed_input = [
input_component.preprocess(raw_input[i])
for i, input_component in enumerate(interface.input_components)
]
interpreter = interface.interpretation
interpretation = interpreter(*processed_input)
if len(raw_input) == 1:
interpretation = [interpretation]
return interpretation, []
else:
raise ValueError("No interpretation method specified.")
def diff(original: Any, perturbed: Any) -> int | float:
try: # try computing numerical difference
score = float(original) - float(perturbed)
except ValueError: # otherwise, look at strict difference in label
score = int(not (original == perturbed))
return score
def quantify_difference_in_label(
interface: Interface, original_output: List, perturbed_output: List
) -> int | float:
output_component = interface.output_components[0]
post_original_output = output_component.postprocess(original_output[0])
post_perturbed_output = output_component.postprocess(perturbed_output[0])
if isinstance(output_component, components.Label):
original_label = post_original_output["label"]
perturbed_label = post_perturbed_output["label"]
# Handle different return types of Label interface
if "confidences" in post_original_output:
original_confidence = original_output[0][original_label]
perturbed_confidence = perturbed_output[0][original_label]
score = original_confidence - perturbed_confidence
else:
score = diff(original_label, perturbed_label)
return score
elif isinstance(output_component, components.Number):
score = diff(post_original_output, post_perturbed_output)
return score
else:
raise ValueError(
"This interpretation method doesn't support the Output component: {}".format(
output_component
)
)
def get_regression_or_classification_value(
interface: Interface, original_output: List, perturbed_output: List
) -> int | float:
"""Used to combine regression/classification for Shap interpretation method."""
output_component = interface.output_components[0]
post_original_output = output_component.postprocess(original_output[0])
post_perturbed_output = output_component.postprocess(perturbed_output[0])
if isinstance(output_component, components.Label):
original_label = post_original_output["label"]
perturbed_label = post_perturbed_output["label"]
# Handle different return types of Label interface
if "confidences" in post_original_output:
if math.isnan(perturbed_output[0][original_label]):
return 0
return perturbed_output[0][original_label]
else:
score = diff(
perturbed_label, original_label
) # Intentionally inverted order of arguments.
return score
else:
raise ValueError(
"This interpretation method doesn't support the Output component: {}".format(
output_component
)
)
|