-
Notifications
You must be signed in to change notification settings - Fork 0
/
explainability_funcs.py
709 lines (552 loc) · 24.1 KB
/
explainability_funcs.py
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
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
import tensorflow as tf
from tensorflow.keras import layers, losses
from tensorflow.keras.callbacks import History
from collections import defaultdict
import numpy as np
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import os
import shutil
from PIL import Image
from typing import Tuple, Dict, List, Union
def visualize_samples(data_path: str) -> Tuple[int, Dict[int, str]]:
"""
Visualizes sample images (one for each class) from a directory and displays class labels.
Parameters:
- data_path (str): The path to the directory containing class subdirectories with images.
Returns:
- Tuple[int, Dict[int, str]]: A tuple containing the total number of classes and a dictionary
mapping numerical labels to class names.
"""
labels = []
num_classes = len(os.listdir(data_path))
fig, ax = plt.subplots(4, 6, figsize=(15, 10))
fig.subplots_adjust(hspace=0.2, wspace=0.2)
for i, label_directory in enumerate(os.listdir(data_path)):
label_path = os.path.join(data_path, label_directory)
if os.path.isdir(label_path):
sample_image = os.listdir(label_path)[0]
image_path = os.path.join(label_path, sample_image)
img = Image.open(image_path)
ax[i // 6, i % 6].set_xticks([])
ax[i // 6, i % 6].set_yticks([])
ax[i // 6, i % 6].set_title(label_directory)
ax[i // 6, i % 6].imshow(img)
labels.append(label_directory)
fig.delaxes(ax[3, 5])
fig.suptitle(f'Total number of classes: {num_classes}', fontsize=13)
plt.show()
labels_dict = {i: lab for i, lab in enumerate(sorted(labels))}
return num_classes, labels_dict
def plot_performance(history: History) -> Tuple[float, float]:
"""
Plot the performance metrics of a neural network based on its training history.
Parameters:
- history (tf.keras.callbacks.History): The training history of a neural network model.
Returns:
- Tuple[float, float]: A tuple containing the test loss and test accuracy of the model.
"""
loss, acc = model.evaluate(val, verbose=0)
metrics = pd.DataFrame({
"train_accuracy": history.history["accuracy"],
"val_accuracy": history.history["val_accuracy"],
"val_loss": history.history["val_loss"],
"train_loss": history.history["loss"],
"epoch": np.arange(0, 100)
})
minimum = metrics[metrics.val_loss == metrics.val_loss.min()].epoch.values[0]
maximum = metrics.val_loss.max()
temp = pd.melt(metrics, id_vars=["epoch"],
value_vars=['train_accuracy', 'val_accuracy', 'train_loss', 'val_loss'], var_name='Set', value_name='Score')
temp['Metric'] = temp['Set'].apply(lambda x: x.split('_')[1].capitalize())
temp['Set'] = temp['Set'].apply(lambda x: x.split('_')[0].capitalize())
sns.set()
rel = sns.relplot(temp, x="epoch", y="Score", col="Metric", hue="Set", kind="line", facet_kws={'sharey': False, 'sharex': True})
rel.fig.suptitle(f'NN FINAL PERFORMANCES: \n Test accuracy: {acc:.2f} \n Test loss: {loss:.2f}', fontsize=14)
rel.fig.subplots_adjust(top=.8)
plt.axvline(minimum, color='red', linestyle="--")
plt.text(minimum+1, maximum-0.1, f'minimum eval loss\nat {minimum} epochs', color="red", fontsize=10)
plt.show()
return loss, acc
loss, acc = plot_performance(history)
def sample_bylabel(dataset: tf.data.Dataset) -> Dict[int, tf.Tensor]:
"""
Extracts one image example for each class from a batched TensorFlow dataset.
Parameters:
- dataset (tf.data.Dataset): A batched TensorFlow dataset.
Returns:
- Dict[Any, Any]: A dictionary containing one representative image for each class.
"""
representative_examples = {}
for image, label in dataset.unbatch():
label = label.numpy() # Convert label to a numpy array for easy comparison
# check if we already have an example for this label
if label not in representative_examples.keys():
representative_examples[label] = image
# Check if we have found one example for each label
if len(representative_examples) == num_classes:
break
return representative_examples
def compute_saliency(input_: tf.Tensor, model: tf.keras.Model) -> np.ndarray:
"""
Compute the saliency map for a given input image considering the relative model's output.
Parameters:
- input_ (tf.Tensor): Input image.
- model (tf.keras.Model): A TensorFlow neural network model.
Returns:
- tf.Tensor: Saliency map.
"""
input_ = tf.expand_dims(input_, axis=0) # add a dimension to check the expected Tensor dimonsion for the model
with tf.GradientTape() as gt:
gt.watch(input_) # track the input image
logit = model(input_, training=False)
logit = tf.squeeze(logit)
class_score = logit[tf.argmax(logit)] # take only the logit for the selected class
saliency_map = gt.gradient(class_score, input_)
saliency_map = tf.squeeze(saliency_map)
saliency_map = np.max(tf.abs(saliency_map), axis=-1)
return saliency_map
def modified_saliency(input_: tf.Tensor,
model: tf.keras.Model) -> np.ndarray:
"""
Compute the modified saliency map for a given input image considering the relative model's output.
The modified saliency map is the element-wise moltiplication of an input by its saliency map.
Parameters:
- input_ (tf.Tensor): Input image.
- model (tf.keras.Model): A TensorFlow neural network model.
Returns:
- tf.Tensor: Modified saliency map.
"""
input_ = tf.expand_dims(input_, axis=0)
with tf.GradientTape() as gt:
gt.watch(input_)
logit = model(input_, training=False)
logit = tf.squeeze(logit)
class_score = logit[tf.argmax(logit)]
saliency_map = gt.gradient(class_score, input_)
saliency_map = tf.squeeze(saliency_map)
saliency_map = tf.abs(saliency_map)
squeezed_input = tf.squeeze(input_)
mod_sm = saliency_map * squeezed_input
return saliency_map
def normalize_to_image(im: tf.Tensor) -> np.ndarray:
"""
Normalize the input image tensor to the range [0, 255].
Parameters:
- im (tf.Tensor): Input image.
Returns:
- np.ndarray: Normalized image.
"""
im = 255 * (im - tf.reduce_min(im)) / (tf.reduce_max(im) - tf.reduce_min(im))
return np.array(im, dtype=np.uint8)
def capped_saliency(saliency_map: np.ndarray) -> np.ndarray:
"""
Cap the saliency map values at the 99-th percentile.
Parameters:
- saliency_map (np.ndarray): Saliency map.
Returns:
- np.ndarray: Capped saliency map.
"""
max_val = np.percentile(saliency_map, 99)
capped_saliency_map = np.array(saliency_map).copy()
capped_saliency_map[capped_saliency_map > max_val] = max_val
return capped_saliency_map
def plot_example_sm(ex: np.ndarray,
model: tf.keras.Model) -> None:
"""
Plots the original input image, its saliency map, and the saliency map capped at the 99th percentile.
Parameters:
- ex (tf.Tensor): An example tensor representing an image.
- model (tf.keras.Model): A TensorFlow neural network model.
Returns:
- None
"""
# Compute the saliency map using the 'compute_saliency' function
saliency_map = compute_saliency(ex, model)
plt.rcdefaults()
fig, ax = plt.subplots(1, 3, figsize=(12, 16))
# Plot the original input image
ax[0].imshow(np.array(ex, dtype=np.uint8))
ax[0].set_xticks([])
ax[0].set_yticks([])
ax[0].set_title("Original input")
# Plot the saliency map
ax[1].imshow(saliency_map, cmap="gray")
ax[1].set_xticks([])
ax[1].set_yticks([])
ax[1].set_title("Saliency map")
# Plot the saliency map capped at 99th percentile
ax[2].imshow(capped_saliency(saliency_map), cmap="gray")
ax[2].set_xticks([])
ax[2].set_yticks([])
ax[2].set_title("Saliency map capped at 99-th perc.")
plt.show()
def noisy_saliency(input_: tf.Tensor,
model: tf.keras.Model,
sigma: int) -> np.ndarray:
"""
Add noise to the given image, then compute the saliency map for the noisy
(smoothed) image, considering the relative model's output.
Parameters:
- input_ (tf.Tensor): Input image.
- model (tf.keras.Model): A TensorFlow neural network model.
- sigma: the noise level to apply (integer between [0,+inf) which represents a percentage)
Returns:
- tf.Tensor: Modified saliency map.
"""
input_ = tf.expand_dims(input_, axis=0)
with tf.GradientTape() as gt:
noise = tf.random.normal(input_.shape, mean=0, stddev=sigma)
noisy_input = input_ + noise
gt.watch(noisy_input)
logit = model(noisy_input, training=False)
logit = tf.squeeze(logit)
class_score = logit[tf.argmax(logit)]
saliency_map = gt.gradient(class_score, noisy_input)
saliency_map = tf.squeeze(saliency_map)
saliency_map = np.max(tf.abs(saliency_map), axis=-1)
noisy_input = tf.squeeze(noisy_input)
return saliency_map
def compute_smooth_saliency(input_: tf.Tensor,
model: tf.keras.Model,
n:int,
sigma: int) -> Tuple[np.ndarray, tf.Tensor]:
"""
Add noise to the given image n times creting n randomly smoothed versions of the image.
Then compute the smooth gradient using the n samples and their relative model's outputs.
Return the smooth gradient result as a tensor (np.array type) and an example smoothed
version of the original image just for visualization
Parameters:
- input_ (tf.Tensor): Input image.
- model (tf.keras.Model): A TensorFlow neural network model.
- sigma: the noise level to apply (integer between [0,+inf) which represents a percentage)
- n: number of samples
Returns:
- tf.Tensor: Modified saliency map.
"""
rep_tensor = tf.convert_to_tensor(tf.repeat(tf.expand_dims(input_, axis=0), n, 0))
out = tf.map_fn(lambda x: noisy_saliency(x, model, sigma), rep_tensor)
smooth_grad = np.mean(out, 0)
noise = tf.random.normal(input_.shape, mean=0, stddev=sigma)
noisy_inp = input_ + noise
return smooth_grad, noisy_inp
def plot_example_sg(ex: np.ndarray,
model: tf.keras.Model) -> None:
"""
Plots the original input image, its saliency map, and the saliency map capped at the 99th percentile,
an example of smoothed input and its corrispective
Parameters:
- ex (tf.Tensor): An example tensor representing an image.
- model (tf.keras.Model): A TensorFlow neural network model.
Returns:
- None
"""
saliency_map = compute_saliency(ex, model)
smooth_saliency_map, noisy_ex = compute_smooth_saliency(ex, model, 10, 60)
plt.rcdefaults()
fig, ax = plt.subplots(1,6, figsize=(19,16))
ax[0].imshow(np.array(ex, dtype=np.uint8))
ax[0].set_xticks([])
ax[0].set_yticks([])
ax[0].set_title("Original input")
ax[1].imshow(saliency_map, cmap="gray")
ax[1].set_xticks([])
ax[1].set_yticks([])
ax[1].set_title("Saliency map")
ax[2].imshow(capped_saliency(saliency_map), cmap="gray")
ax[2].set_xticks([])
ax[2].set_yticks([])
ax[2].set_title("\"Capped\" saliency map")
ax[3].imshow(normalize_to_image(noisy_ex))
ax[3].set_xticks([])
ax[3].set_yticks([])
ax[3].set_title("Smoothed input")
ax[4].imshow(smooth_saliency_map, cmap="gray")
ax[4].set_xticks([])
ax[4].set_yticks([])
ax[4].set_title("Smoothed saliency map")
ax[5].imshow(capped_saliency(smooth_saliency_map), cmap="gray")
ax[5].set_xticks([])
ax[5].set_yticks([])
ax[5].set_title("\"Capped\" Smoothed map")
plt.show()
def compare_noise_and_samples(ex: np.ndarray,
model: tf.keras.Model) -> None:
"""
Plot the smooth gradient for different levels of noise and number of samples
Parameters:
- ex (tf.Tensor): An example tensor representing an image.
- model (tf.keras.Model): A TensorFlow neural network model.
Returns:
- None
"""
n_samples = np.arange(10, 81, 10)
sigmas = np.arange(10, 150, 10)
fig, ax = plt.subplots(len(n_samples), len(sigmas), figsize=(21,12), gridspec_kw={'wspace': 0, 'hspace': 0})
ax = ax.flatten()
for i, n in enumerate(n_samples):
for j, s in enumerate(sigmas):
idx = i * len(sigmas) + j
smooth_saliency_map, _ = compute_smooth_saliency(ex, model, s, n)
ax[idx].imshow(smooth_saliency_map)
ax[idx].set_xticks([])
ax[idx].set_yticks([])
if j == 0:
ax[idx].set_ylabel(f'{n} samples')
if i == 0:
ax[idx].set_title(f'noise {s/255*100:.2f}%')
plt.show()
def plot_sm_vs_sg(representative_examples: Dict[int, tf.Tensor],
model: tf.keras.Model,
n: int,
sigma: int) -> None:
"""
Plots the original input image, its saliency map, and the saliency map capped at the 99th percentile,
its smooth gradient and capped smooth gradient for given level of noise sigma and number of samples
for each example image in the given dictionary. The keys of the dictionary are the labels and the values
are the corresponding sampled examples
Parameters:
- representative_examples Dict[int, tf.Tensor]: a dictionary of example images (one for each class)
- model (tf.keras.Model): A TensorFlow neural network model.
- sigma: the noise level to apply (integer between [0,+inf) which represents a percentage)
- n: number of samples
Returns:
- None
"""
fig, ax = plt.subplots(23, 5, figsize=(10,50), gridspec_kw={'wspace': 0, 'hspace': 0})
ax = ax.flatten()
for i, ex in enumerate(representative_examples.items()):
img = ex[1]
label = ex[0]
idx1 = i * 5
idx2 = i * 5 + 1
idx3 = i * 5 + 2
idx4 = i * 5 + 3
idx5 = i * 5 + 4
saliency_map = compute_saliency(img, model)
smooth_saliency, _ = compute_smooth_saliency(img, model, n, sigma)
ax[idx1].imshow(np.array(img, dtype=np.uint8))
ax[idx1].set_xticks([])
ax[idx1].set_yticks([])
ax[idx1].set_ylabel(f'{labels[int(label)]}')
ax[idx2].imshow(saliency_map, cmap="gray")
ax[idx2].set_xticks([])
ax[idx2].set_yticks([])
ax[idx3].imshow(capped_saliency(saliency_map), cmap="gray")
ax[idx3].set_xticks([])
ax[idx3].set_yticks([])
ax[idx4].imshow(smooth_saliency, cmap="gray")
ax[idx4].set_xticks([])
ax[idx4].set_yticks([])
ax[idx5].imshow(capped_saliency(smooth_saliency), cmap="gray")
ax[idx5].set_xticks([])
ax[idx5].set_yticks([])
if i == 0:
ax[idx1].set_title("Input")
ax[idx2].set_title("Saliency Map")
ax[idx3].set_title("Capped s.m.")
ax[idx4].set_title("Smooth Gradient")
ax[idx5].set_title("Capped s.g.")
plt.show()
def parallel_saliency(dataset: tf.data.Dataset,
labels: Dict[int, str],
by_class: bool = False) -> Union[List[tf.Tensor], tf.Tensor]:
"""
Compute the saliency map for each example in the training (or whatever) set.
If by_class is set to True, the saliency maps are stacked in different tensors,
divided by class, and then stored in a list. Otherwise, if by_class is set to
False, every saliency map is stacked in a unique tensor.
Parameters:
- dataset (tf.data.Dataset): Input dataset containing examples.
- labels (Dict[int, str]): Dictionary mapping class labels to their corresponding names.
- by_class (bool): If True, stack saliency maps by class; if False, stack all maps together.
Returns:
- Union[List[tf.Tensor], tf.Tensor]: Stacked saliency maps, either as a list of tensors (if by_class=True)
or as a single tensor (if by_class=False).
"""
if by_class:
stacked_smaps = []
for i, l in labels.items():
inputs = np.array([x[0] for x in dataset.unbatch().as_numpy_iterator() if x[1] == int(i)], dtype=np.float32)
inputs = tf.convert_to_tensor(inputs)
stacked_smaps.append(tf.map_fn(lambda x: compute_saliency(x, model), inputs))
else:
inputs = np.array([x[0] for x in train.unbatch().as_numpy_iterator()], dtype=np.float32)
inputs = tf.convert_to_tensor(inputs)
stacked_smaps = tf.map_fn(lambda x: compute_saliency(x, model), inputs)
return stacked_smaps
def global_saliency(dataset: tf.data.Dataset,
labels: Dict[int, str],
by_class: bool = False) -> Union[List[np.ndarray], np.ndarray]:
"""
Compute the global saliency map for a dataset.
Parameters:
- dataset (tf.data.Dataset): Input dataset containing examples.
- labels (Dict[int, str]): Dictionary mapping class labels to their corresponding names.
- by_class (bool): If True, compute global saliency maps by class; if False, compute overall global saliency map.
Returns:
- Union[List[np.ndarray], np.ndarray]: Global saliency map(s), either as a list of numpy arrays (if by_class=True)
or as a single numpy array (if by_class=False).
"""
stacked_smaps = parallel_saliency(dataset, labels, by_class)
if by_class:
return [np.mean(i, axis=0) for i in stacked_smaps]
return np.mean(stacked_smaps, axis=0)
def show_global_sm(train: tf.data.Dataset,
labels: Dict[int, str]) -> np.ndarray:
"""
Show global saliency maps for each class and overall dataset.
Parameters:
- train (tf.data.Dataset): Input training dataset containing examples.
- labels (Dict[int, str]): Dictionary mapping class labels to their corresponding names.
Returns:
- np.ndarray: Overall global saliency map for the entire dataset.
"""
fig, ax = plt.subplots(4, 6, figsize=(16, 12), gridspec_kw={'wspace': 0, 'hspace': 0})
ax = ax.flatten()
gs = global_saliency(train, labels, by_class=True)
for i, saliency in enumerate(gs):
ax[i].imshow(saliency)
ax[i].set_xticks([])
ax[i].set_yticks([])
ax[i].set_title(f'{labels[i]}')
glob = global_saliency(train, labels)
ax[-1].imshow(glob)
ax[-1].set_title("All data set", color="red")
ax[-1].set_xticks([])
ax[-1].set_yticks([])
plt.show()
return glob
def show_weights_importance(weights: tf.Tensor,
labels: Dict[int, str]) -> np.ndarray:
"""
Show the importance given by the linear model to each feature (pixel) of any
training input (image). In this case plot the weghts for each neuron corresponding
to a single class and also the mean of the weights w.r.t. all the classes.
Parameters:
- weights (tf.Tensor): The weights associated to a linear model (biases excluded)
- labels (Dict[int, str]): Dictionary mapping class labels to their corresponding names.
Returns:
- None
"""
weights = np.reshape(weights, (-1,180,180,3))
fig, ax = plt.subplots(4, 6, figsize=(16, 12), gridspec_kw={'wspace': 0, 'hspace': 0})
ax = ax.flatten()
for i, w in enumerate(weights):
importance = np.max(w, axis=-1)
ax[i].imshow(normalize_to_image(importance))
ax[i].set_xticks([])
ax[i].set_yticks([])
ax[i].set_title(f'{labels[i]}')
glob_importance = np.max(np.mean(weights, axis=0), axis=2)
ax[-1].imshow(normalize_to_image(glob_importance))
ax[-1].set_title("All classes", color="red")
ax[-1].set_xticks([])
ax[-1].set_yticks([])
plt.show()
def step_grad(input_: tf.Tensor,
baseline: tf.Tensor,
model: tf.keras.Model,
k: int,
m: int) -> np.ndarray:
"""
Using the baseline image and the input image, compute the intermediate result
for one step of the approximated Integrated Gradient algorithm.
Parameters:
- input_ (tf.Tensor): Input image.
- baseline (tf.Tensor): Baseline image (with a saliency map of all 0s)
- model (tf.keras.Model): A TensorFlow neural network model.
- k (int): k-th iteration for the approximated integrated gradient algorithm
- m (int): total number of iterations of the approximated integrated gradient algorithm
Returns:
- np.array: the intermediate result for one iteration of the algorithm.
"""
input_ = tf.expand_dims(input_, axis=0)
baseline = tf.expand_dims(baseline, axis=0)
with tf.GradientTape() as gt:
gt.watch(input_)
num = baseline + tf.cast((k/m), dtype=input_.dtype) * tf.cast((input_ - baseline), dtype=tf.float32)
logit = model(num, training=False)
logit = tf.squeeze(logit)
class_score = logit[tf.argmax(logit)]
grad = gt.gradient(class_score, input_)
grad = tf.squeeze(grad)
return grad
def compute_integrated(input_: tf.Tensor,
baseline: tf.Tensor,
model: tf.keras.Model,
m: int) -> np.ndarray:
"""
Call m times the 'step_grad' function. Return the result og the approximated
Integrated Gradient algorihtm.
Parameters:
- input_ (tf.Tensor): Input image.
- baseline (tf.Tensor): Baseline image (with a saliency map of all 0s)
- model (tf.keras.Model): A TensorFlow neural network model.
- m (int): total number of iterations of the approximated integrated gradient algorithm
Returns:
- np.array: the intermediate result for one iteration of the algorithm.
"""
rep_tensor = tf.convert_to_tensor(np.arange(1,m+1,1), dtype=tf.float32)
out = np.mean(tf.map_fn(lambda x: step_grad(input_, baseline, model, x, m), rep_tensor), 0)
coeff = np.array(input_ - baseline)
int_grad = coeff * out
int_grad = np.max(tf.abs(int_grad), axis=-1)
return int_grad
def plot_sm_vs_ig(representative_examples: Dict[int, tf.Tensor],
baseline: tf.Tensor,
model: tf.keras.Model,
m: int,
n: int,
sigma: int) -> None:
"""
Plots the original input image, its saliency map, and the saliency map capped at the 99th percentile,
its smooth gradient and capped smooth gradient for given level of noise sigma and number of samples
for each example image in the given dictionary. The keys of the dictionary are the labels and the values
are the corresponding sampled examples
Parameters:
- representative_examples Dict[int, tf.Tensor]: a dictionary of example images (one for each class).
- baseline (tf.Tensor): baseline example.
- model (tf.keras.Model): A TensorFlow neural network model.
- m (int): total number of iterations of the approximated integrated gradient algorithm
- sigma: the noise level to apply (integer between [0,+inf) which represents a percentage).
- n: number of samples
Returns:
- None
"""
fig, ax = plt.subplots(23, 5, figsize=(10,50), gridspec_kw={'wspace': 0, 'hspace': 0})
ax = ax.flatten()
for i, ex in enumerate(representative_examples.items()):
img = ex[1]
label = ex[0]
idx1 = i * 5
idx2 = i * 5 + 1
idx3 = i * 5 + 2
idx4 = i * 5 + 3
idx5 = i * 5 + 4
smooth_saliency, _ = compute_smooth_saliency(img, model, n, sigma)
integr_saliency = compute_integrated(img, baseline, model, m)
ax[idx1].imshow(np.array(img, dtype=np.uint8))
ax[idx1].set_xticks([])
ax[idx1].set_yticks([])
ax[idx1].set_ylabel(f'{labels[int(label)]}')
ax[idx2].imshow(smooth_saliency, cmap="gray")
ax[idx2].set_xticks([])
ax[idx2].set_yticks([])
ax[idx3].imshow(capped_saliency(smooth_saliency), cmap="gray")
ax[idx3].set_xticks([])
ax[idx3].set_yticks([])
ax[idx4].imshow(integr_saliency, cmap="gray")
ax[idx4].set_xticks([])
ax[idx4].set_yticks([])
ax[idx5].imshow(capped_saliency(integr_saliency), cmap="gray")
ax[idx5].set_xticks([])
ax[idx5].set_yticks([])
if i == 0:
ax[idx1].set_title("Input")
ax[idx2].set_title("Smooth gradient")
ax[idx3].set_title("Capped s.g.")
ax[idx4].set_title("Integrated gradient")
ax[idx5].set_title("Capped i.g.")
plt.show()