File size: 11,291 Bytes
443d045 |
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 |
import copy
import math
import numpy as np
from gradio import utils
from gradio.components import Label, Number
async def run_interpret(interface, raw_input):
"""
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 input_component.interpret_by_tokens:
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,
)
)
else:
(
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
)
)
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 (input_component.interpret_by_tokens):
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):
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,
)
scores.append(
input_component.get_interpretation_scores(
raw_input[i], None, shap_values[0], 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
else: # 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, []
def diff(original, perturbed):
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, original_output, perturbed_output):
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, 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, 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, original_output, perturbed_output
):
"""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 type(output_component) == 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
)
)
|