Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,400 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import numpy as np
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
from scipy import ndimage
|
5 |
+
from IPython.display import Image
|
6 |
+
|
7 |
+
import tensorflow as tf
|
8 |
+
from tensorflow import keras
|
9 |
+
from tensorflow.keras import layers
|
10 |
+
from tensorflow.keras.applications import xception
|
11 |
+
|
12 |
+
# Size of the input image
|
13 |
+
img_size = (299, 299, 3)
|
14 |
+
|
15 |
+
# Load Xception model with imagenet weights
|
16 |
+
model = xception.Xception(weights="imagenet")
|
17 |
+
|
18 |
+
# The local path to our target image
|
19 |
+
img_path = keras.utils.get_file("elephant.jpg", "https://i.imgur.com/Bvro0YD.png")
|
20 |
+
|
21 |
+
def get_gradients(img_input, top_pred_idx):
|
22 |
+
"""Computes the gradients of outputs w.r.t input image.
|
23 |
+
|
24 |
+
Args:
|
25 |
+
img_input: 4D image tensor
|
26 |
+
top_pred_idx: Predicted label for the input image
|
27 |
+
|
28 |
+
Returns:
|
29 |
+
Gradients of the predictions w.r.t img_input
|
30 |
+
"""
|
31 |
+
images = tf.cast(img_input, tf.float32)
|
32 |
+
|
33 |
+
with tf.GradientTape() as tape:
|
34 |
+
tape.watch(images)
|
35 |
+
preds = model(images)
|
36 |
+
top_class = preds[:, top_pred_idx]
|
37 |
+
|
38 |
+
grads = tape.gradient(top_class, images)
|
39 |
+
return grads
|
40 |
+
|
41 |
+
|
42 |
+
def get_integrated_gradients(img_input, top_pred_idx, baseline=None, num_steps=50):
|
43 |
+
"""Computes Integrated Gradients for a predicted label.
|
44 |
+
|
45 |
+
Args:
|
46 |
+
img_input (ndarray): Original image
|
47 |
+
top_pred_idx: Predicted label for the input image
|
48 |
+
baseline (ndarray): The baseline image to start with for interpolation
|
49 |
+
num_steps: Number of interpolation steps between the baseline
|
50 |
+
and the input used in the computation of integrated gradients. These
|
51 |
+
steps along determine the integral approximation error. By default,
|
52 |
+
num_steps is set to 50.
|
53 |
+
|
54 |
+
Returns:
|
55 |
+
Integrated gradients w.r.t input image
|
56 |
+
"""
|
57 |
+
# If baseline is not provided, start with a black image
|
58 |
+
# having same size as the input image.
|
59 |
+
if baseline is None:
|
60 |
+
baseline = np.zeros(img_size).astype(np.float32)
|
61 |
+
else:
|
62 |
+
baseline = baseline.astype(np.float32)
|
63 |
+
|
64 |
+
# 1. Do interpolation.
|
65 |
+
img_input = img_input.astype(np.float32)
|
66 |
+
interpolated_image = [
|
67 |
+
baseline + (step / num_steps) * (img_input - baseline)
|
68 |
+
for step in range(num_steps + 1)
|
69 |
+
]
|
70 |
+
interpolated_image = np.array(interpolated_image).astype(np.float32)
|
71 |
+
|
72 |
+
# 2. Preprocess the interpolated images
|
73 |
+
interpolated_image = xception.preprocess_input(interpolated_image)
|
74 |
+
|
75 |
+
# 3. Get the gradients
|
76 |
+
grads = []
|
77 |
+
for i, img in enumerate(interpolated_image):
|
78 |
+
img = tf.expand_dims(img, axis=0)
|
79 |
+
grad = get_gradients(img, top_pred_idx=top_pred_idx)
|
80 |
+
grads.append(grad[0])
|
81 |
+
grads = tf.convert_to_tensor(grads, dtype=tf.float32)
|
82 |
+
|
83 |
+
# 4. Approximate the integral using the trapezoidal rule
|
84 |
+
grads = (grads[:-1] + grads[1:]) / 2.0
|
85 |
+
avg_grads = tf.reduce_mean(grads, axis=0)
|
86 |
+
|
87 |
+
# 5. Calculate integrated gradients and return
|
88 |
+
integrated_grads = (img_input - baseline) * avg_grads
|
89 |
+
return integrated_grads
|
90 |
+
|
91 |
+
|
92 |
+
def random_baseline_integrated_gradients(
|
93 |
+
img_input, top_pred_idx, num_steps=50, num_runs=2
|
94 |
+
):
|
95 |
+
"""Generates a number of random baseline images.
|
96 |
+
|
97 |
+
Args:
|
98 |
+
img_input (ndarray): 3D image
|
99 |
+
top_pred_idx: Predicted label for the input image
|
100 |
+
num_steps: Number of interpolation steps between the baseline
|
101 |
+
and the input used in the computation of integrated gradients. These
|
102 |
+
steps along determine the integral approximation error. By default,
|
103 |
+
num_steps is set to 50.
|
104 |
+
num_runs: number of baseline images to generate
|
105 |
+
|
106 |
+
Returns:
|
107 |
+
Averaged integrated gradients for `num_runs` baseline images
|
108 |
+
"""
|
109 |
+
# 1. List to keep track of Integrated Gradients (IG) for all the images
|
110 |
+
integrated_grads = []
|
111 |
+
|
112 |
+
# 2. Get the integrated gradients for all the baselines
|
113 |
+
for run in range(num_runs):
|
114 |
+
baseline = np.random.random(img_size) * 255
|
115 |
+
igrads = get_integrated_gradients(
|
116 |
+
img_input=img_input,
|
117 |
+
top_pred_idx=top_pred_idx,
|
118 |
+
baseline=baseline,
|
119 |
+
num_steps=num_steps,
|
120 |
+
)
|
121 |
+
integrated_grads.append(igrads)
|
122 |
+
|
123 |
+
# 3. Return the average integrated gradients for the image
|
124 |
+
integrated_grads = tf.convert_to_tensor(integrated_grads)
|
125 |
+
return tf.reduce_mean(integrated_grads, axis=0)
|
126 |
+
|
127 |
+
class GradVisualizer:
|
128 |
+
"""Plot gradients of the outputs w.r.t an input image."""
|
129 |
+
|
130 |
+
def __init__(self, positive_channel=None, negative_channel=None):
|
131 |
+
if positive_channel is None:
|
132 |
+
self.positive_channel = [0, 255, 0]
|
133 |
+
else:
|
134 |
+
self.positive_channel = positive_channel
|
135 |
+
|
136 |
+
if negative_channel is None:
|
137 |
+
self.negative_channel = [255, 0, 0]
|
138 |
+
else:
|
139 |
+
self.negative_channel = negative_channel
|
140 |
+
|
141 |
+
def apply_polarity(self, attributions, polarity):
|
142 |
+
if polarity == "positive":
|
143 |
+
return np.clip(attributions, 0, 1)
|
144 |
+
else:
|
145 |
+
return np.clip(attributions, -1, 0)
|
146 |
+
|
147 |
+
def apply_linear_transformation(
|
148 |
+
self,
|
149 |
+
attributions,
|
150 |
+
clip_above_percentile=99.9,
|
151 |
+
clip_below_percentile=70.0,
|
152 |
+
lower_end=0.2,
|
153 |
+
):
|
154 |
+
# 1. Get the thresholds
|
155 |
+
m = self.get_thresholded_attributions(
|
156 |
+
attributions, percentage=100 - clip_above_percentile
|
157 |
+
)
|
158 |
+
e = self.get_thresholded_attributions(
|
159 |
+
attributions, percentage=100 - clip_below_percentile
|
160 |
+
)
|
161 |
+
|
162 |
+
# 2. Transform the attributions by a linear function f(x) = a*x + b such that
|
163 |
+
# f(m) = 1.0 and f(e) = lower_end
|
164 |
+
transformed_attributions = (1 - lower_end) * (np.abs(attributions) - e) / (
|
165 |
+
m - e
|
166 |
+
) + lower_end
|
167 |
+
|
168 |
+
# 3. Make sure that the sign of transformed attributions is the same as original attributions
|
169 |
+
transformed_attributions *= np.sign(attributions)
|
170 |
+
|
171 |
+
# 4. Only keep values that are bigger than the lower_end
|
172 |
+
transformed_attributions *= transformed_attributions >= lower_end
|
173 |
+
|
174 |
+
# 5. Clip values and return
|
175 |
+
transformed_attributions = np.clip(transformed_attributions, 0.0, 1.0)
|
176 |
+
return transformed_attributions
|
177 |
+
|
178 |
+
def get_thresholded_attributions(self, attributions, percentage):
|
179 |
+
if percentage == 100.0:
|
180 |
+
return np.min(attributions)
|
181 |
+
|
182 |
+
# 1. Flatten the attributions
|
183 |
+
flatten_attr = attributions.flatten()
|
184 |
+
|
185 |
+
# 2. Get the sum of the attributions
|
186 |
+
total = np.sum(flatten_attr)
|
187 |
+
|
188 |
+
# 3. Sort the attributions from largest to smallest.
|
189 |
+
sorted_attributions = np.sort(np.abs(flatten_attr))[::-1]
|
190 |
+
|
191 |
+
# 4. Calculate the percentage of the total sum that each attribution
|
192 |
+
# and the values about it contribute.
|
193 |
+
cum_sum = 100.0 * np.cumsum(sorted_attributions) / total
|
194 |
+
|
195 |
+
# 5. Threshold the attributions by the percentage
|
196 |
+
indices_to_consider = np.where(cum_sum >= percentage)[0][0]
|
197 |
+
|
198 |
+
# 6. Select the desired attributions and return
|
199 |
+
attributions = sorted_attributions[indices_to_consider]
|
200 |
+
return attributions
|
201 |
+
|
202 |
+
def binarize(self, attributions, threshold=0.001):
|
203 |
+
return attributions > threshold
|
204 |
+
|
205 |
+
def morphological_cleanup_fn(self, attributions, structure=np.ones((4, 4))):
|
206 |
+
closed = ndimage.grey_closing(attributions, structure=structure)
|
207 |
+
opened = ndimage.grey_opening(closed, structure=structure)
|
208 |
+
return opened
|
209 |
+
|
210 |
+
def draw_outlines(
|
211 |
+
self, attributions, percentage=90, connected_component_structure=np.ones((3, 3))
|
212 |
+
):
|
213 |
+
# 1. Binarize the attributions.
|
214 |
+
attributions = self.binarize(attributions)
|
215 |
+
|
216 |
+
# 2. Fill the gaps
|
217 |
+
attributions = ndimage.binary_fill_holes(attributions)
|
218 |
+
|
219 |
+
# 3. Compute connected components
|
220 |
+
connected_components, num_comp = ndimage.measurements.label(
|
221 |
+
attributions, structure=connected_component_structure
|
222 |
+
)
|
223 |
+
|
224 |
+
# 4. Sum up the attributions for each component
|
225 |
+
total = np.sum(attributions[connected_components > 0])
|
226 |
+
component_sums = []
|
227 |
+
for comp in range(1, num_comp + 1):
|
228 |
+
mask = connected_components == comp
|
229 |
+
component_sum = np.sum(attributions[mask])
|
230 |
+
component_sums.append((component_sum, mask))
|
231 |
+
|
232 |
+
# 5. Compute the percentage of top components to keep
|
233 |
+
sorted_sums_and_masks = sorted(component_sums, key=lambda x: x[0], reverse=True)
|
234 |
+
sorted_sums = list(zip(*sorted_sums_and_masks))[0]
|
235 |
+
cumulative_sorted_sums = np.cumsum(sorted_sums)
|
236 |
+
cutoff_threshold = percentage * total / 100
|
237 |
+
cutoff_idx = np.where(cumulative_sorted_sums >= cutoff_threshold)[0][0]
|
238 |
+
if cutoff_idx > 2:
|
239 |
+
cutoff_idx = 2
|
240 |
+
|
241 |
+
# 6. Set the values for the kept components
|
242 |
+
border_mask = np.zeros_like(attributions)
|
243 |
+
for i in range(cutoff_idx + 1):
|
244 |
+
border_mask[sorted_sums_and_masks[i][1]] = 1
|
245 |
+
|
246 |
+
# 7. Make the mask hollow and show only the border
|
247 |
+
eroded_mask = ndimage.binary_erosion(border_mask, iterations=1)
|
248 |
+
border_mask[eroded_mask] = 0
|
249 |
+
|
250 |
+
# 8. Return the outlined mask
|
251 |
+
return border_mask
|
252 |
+
|
253 |
+
def process_grads(
|
254 |
+
self,
|
255 |
+
image,
|
256 |
+
attributions,
|
257 |
+
polarity="positive",
|
258 |
+
clip_above_percentile=99.9,
|
259 |
+
clip_below_percentile=0,
|
260 |
+
morphological_cleanup=False,
|
261 |
+
structure=np.ones((3, 3)),
|
262 |
+
outlines=False,
|
263 |
+
outlines_component_percentage=90,
|
264 |
+
overlay=True,
|
265 |
+
):
|
266 |
+
if polarity not in ["positive", "negative"]:
|
267 |
+
raise ValueError(
|
268 |
+
f""" Allowed polarity values: 'positive' or 'negative'
|
269 |
+
but provided {polarity}"""
|
270 |
+
)
|
271 |
+
if clip_above_percentile < 0 or clip_above_percentile > 100:
|
272 |
+
raise ValueError("clip_above_percentile must be in [0, 100]")
|
273 |
+
|
274 |
+
if clip_below_percentile < 0 or clip_below_percentile > 100:
|
275 |
+
raise ValueError("clip_below_percentile must be in [0, 100]")
|
276 |
+
|
277 |
+
# 1. Apply polarity
|
278 |
+
if polarity == "positive":
|
279 |
+
attributions = self.apply_polarity(attributions, polarity=polarity)
|
280 |
+
channel = self.positive_channel
|
281 |
+
else:
|
282 |
+
attributions = self.apply_polarity(attributions, polarity=polarity)
|
283 |
+
attributions = np.abs(attributions)
|
284 |
+
channel = self.negative_channel
|
285 |
+
|
286 |
+
# 2. Take average over the channels
|
287 |
+
attributions = np.average(attributions, axis=2)
|
288 |
+
|
289 |
+
# 3. Apply linear transformation to the attributions
|
290 |
+
attributions = self.apply_linear_transformation(
|
291 |
+
attributions,
|
292 |
+
clip_above_percentile=clip_above_percentile,
|
293 |
+
clip_below_percentile=clip_below_percentile,
|
294 |
+
lower_end=0.0,
|
295 |
+
)
|
296 |
+
|
297 |
+
# 4. Cleanup
|
298 |
+
if morphological_cleanup:
|
299 |
+
attributions = self.morphological_cleanup_fn(
|
300 |
+
attributions, structure=structure
|
301 |
+
)
|
302 |
+
# 5. Draw the outlines
|
303 |
+
if outlines:
|
304 |
+
attributions = self.draw_outlines(
|
305 |
+
attributions, percentage=outlines_component_percentage
|
306 |
+
)
|
307 |
+
|
308 |
+
# 6. Expand the channel axis and convert to RGB
|
309 |
+
attributions = np.expand_dims(attributions, 2) * channel
|
310 |
+
|
311 |
+
# 7.Superimpose on the original image
|
312 |
+
if overlay:
|
313 |
+
attributions = np.clip((attributions * 0.8 + image), 0, 255)
|
314 |
+
return attributions
|
315 |
+
|
316 |
+
def visualize(
|
317 |
+
self,
|
318 |
+
image,
|
319 |
+
gradients,
|
320 |
+
integrated_gradients,
|
321 |
+
polarity="positive",
|
322 |
+
clip_above_percentile=99.9,
|
323 |
+
clip_below_percentile=0,
|
324 |
+
morphological_cleanup=False,
|
325 |
+
structure=np.ones((3, 3)),
|
326 |
+
outlines=False,
|
327 |
+
outlines_component_percentage=90,
|
328 |
+
overlay=True,
|
329 |
+
figsize=(15, 8),
|
330 |
+
):
|
331 |
+
# 1. Make two copies of the original image
|
332 |
+
img1 = np.copy(image)
|
333 |
+
img2 = np.copy(image)
|
334 |
+
|
335 |
+
# 2. Process the normal gradients
|
336 |
+
grads_attr = self.process_grads(
|
337 |
+
image=img1,
|
338 |
+
attributions=gradients,
|
339 |
+
polarity=polarity,
|
340 |
+
clip_above_percentile=clip_above_percentile,
|
341 |
+
clip_below_percentile=clip_below_percentile,
|
342 |
+
morphological_cleanup=morphological_cleanup,
|
343 |
+
structure=structure,
|
344 |
+
outlines=outlines,
|
345 |
+
outlines_component_percentage=outlines_component_percentage,
|
346 |
+
overlay=overlay,
|
347 |
+
)
|
348 |
+
|
349 |
+
# 3. Process the integrated gradients
|
350 |
+
igrads_attr = self.process_grads(
|
351 |
+
image=img2,
|
352 |
+
attributions=integrated_gradients,
|
353 |
+
polarity=polarity,
|
354 |
+
clip_above_percentile=clip_above_percentile,
|
355 |
+
clip_below_percentile=clip_below_percentile,
|
356 |
+
morphological_cleanup=morphological_cleanup,
|
357 |
+
structure=structure,
|
358 |
+
outlines=outlines,
|
359 |
+
outlines_component_percentage=outlines_component_percentage,
|
360 |
+
overlay=overlay,
|
361 |
+
)
|
362 |
+
|
363 |
+
return igrads_attr.astype(np.uint8)
|
364 |
+
|
365 |
+
def classify_image(image):
|
366 |
+
img = np.expand_dims(image, axis=0)
|
367 |
+
orig_img = np.copy(img[0]).astype(np.uint8)
|
368 |
+
img_processed = tf.cast(xception.preprocess_input(img), dtype=tf.float32)
|
369 |
+
preds = model.predict(img_processed)
|
370 |
+
top_pred_idx = tf.argmax(preds[0])
|
371 |
+
print("Predicted:", top_pred_idx, xception.decode_predictions(preds, top=1)[0])
|
372 |
+
grads = get_gradients(img_processed, top_pred_idx=top_pred_idx)
|
373 |
+
igrads = random_baseline_integrated_gradients(
|
374 |
+
np.copy(orig_img), top_pred_idx=top_pred_idx, num_steps=50, num_runs=2)
|
375 |
+
vis = GradVisualizer()
|
376 |
+
img_grads = vis.visualize(
|
377 |
+
image=orig_img,
|
378 |
+
gradients=grads[0].numpy(),
|
379 |
+
integrated_gradients=igrads.numpy(),
|
380 |
+
clip_above_percentile=99,
|
381 |
+
clip_below_percentile=0,
|
382 |
+
)
|
383 |
+
return {labels[i]: float(prediction[i]) for i in range(100)}
|
384 |
+
|
385 |
+
image = gr.inputs.Image(shape=(299,299))
|
386 |
+
label = gr.outputs.Image()
|
387 |
+
|
388 |
+
iface = gr.Interface(classify_image,image,label,
|
389 |
+
#outputs=[
|
390 |
+
# gr.outputs.Textbox(label="Engine issue"),
|
391 |
+
# gr.outputs.Textbox(label="Engine issue score")],
|
392 |
+
examples=["elephant.jpg.jpg"],
|
393 |
+
title="Model interpretability with Integrated Gradients",
|
394 |
+
description = "Model for classifying images from the CIFAR dataset using a vision transformer trained with small data.",
|
395 |
+
article = "Author: <a href=\"https://huggingface.co/joheras\">Jónathan Heras</a>"
|
396 |
+
# examples = ["sample.csv"],
|
397 |
+
)
|
398 |
+
|
399 |
+
|
400 |
+
iface.launch()
|