-
Notifications
You must be signed in to change notification settings - Fork 3
/
computecomplexityfinal.py
399 lines (337 loc) · 13.1 KB
/
computecomplexityfinal.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
"""Custom Complexity Measure based on Margin Distribution"""
import numpy as np
import tensorflow as tf
from tensorflow import keras
from collections import defaultdict
import json
import pickle
import os
import time
import sys
import random
import math
from tensorflow.keras.models import load_model
import tensorflow.keras.backend as K
from scipy.stats import *
sys.path.append('..')
from sklearn.preprocessing import StandardScaler, Normalizer
from sklearn.metrics import silhouette_score
from math import log
import matplotlib.pyplot as plt
from augment import *
import gc
import time
class CustomComplexityFinal:
"""
A class used to create margin based complexity measures
Attributes
----------
model : tf.keras.Model()
The Keras model for which the complexity measure is to be computed
dataset : tf.data.Dataset
Dataset object from PGDL data loader
rootpath : str, optional
Path to root directory
computeOver : int
The number of samples over which to compute the complexity measure
batchSize: int
The batch size
basename: str, optional
basename argument of PGDL directory structure
metric: str, optional
Metric to use to scale margin Distribution
augment : str, optional
The type of augmentation to use ('standard', 'mixup', 'adverserial', 'adverserial+standard', 'mixup+standard')
penalize : bool, optional
Whether to penalize misclassified samples
input_margin : bool, optional
Whether to compute margin on input data instead of intermediate representations
network_scale: bool, optional
Only used for auxiliary experiments involving regularizing for network topology
seed: int, optional
Random seed
"""
def __init__(self, model, ds, rootpath=None, mid=None, computeOver = 500, batchSize = 50, basename=None, metric='batch_variance', augment='standard', penalize=True, input_margin=False, network_scale = False, seed=1):
self.model = model
self.dataset = ds
self.computeOver = computeOver
if rootpath is not None:
self.rootPath = os.path.join(rootpath, 'computed_data/pickles/{}'.format(mid))
if not os.path.exists(self.rootPath):
os.makedirs(self.rootPath)
self.mid = mid
self.basename = basename
self.batchSize = batchSize
self.metric = metric
self.verbose = False
self.norm = 'l2'
self.penalize = penalize
self.augment = augment
self.input_margin = input_margin
self.network_scale = network_scale
self.seed=seed
# ====================================== Functions for Margin Based Solution =====================================
def computeMargins(self, top = 2):
'''
Fuction to compute margin distribution
Returns
-------
marginDistribution : dict
Dictionary containing lists of margin distribution for each layer
'''
it = iter(self.dataset.repeat(-1).shuffle(5000, seed=self.seed).batch(self.batchSize))
marginDistribution = {}
totalVariance = {}
totalVarianceTensor = {}
totalNorm = {}
totalNormTensor = {}
self.layers = []
ratio_list = {}
for l in range(len(self.model.layers)):
c = list(self.model.get_layer(index = l).get_config().keys())
if 'filters' in c or 'units' in c:
self.layers.append(l)
if len(self.layers) == 1:
break
if self.input_margin == True:
self.layers = [-1]
if self.verbose == True:
for l, layer in enumerate(self.model.layers):
print(self.model.get_layer(index = l).get_config())
for i in range(self.computeOver//self.batchSize):
batch = next(it)
if self.augment == 'standard':
D = DataAugmentor(batch[0], batchSize = self.batchSize)
batch_ = (D.augment(), batch[1])
elif self.augment == 'adverserial':
batch_ = (self.getAdverserialBatch(batch), batch[1])
elif self.augment == 'adverserial+standard':
D = DataAugmentor(batch[0], batchSize = self.batchSize)
batch_ = (D.augment(), batch[1])
batch_ = (self.getAdverserialBatch(batch_), batch_[1])
elif self.augment == 'mixup':
batch_ = self.batchMixupLabelwiseLinear(batch)
elif self.augment == 'mixup+standard':
D = DataAugmentor(batch[0], batchSize = self.batchSize)
batch_ = (D.augment(), batch[1])
batch_ = self.batchMixupLabelwise(batch_)
else:
batch_ = batch
for layer in self.layers:
if self.augment is not None:
grads, inter = self.distancefromMargin(batch_, layer+1, top)
else:
grads, inter = self.distancefromMargin(batch_, layer+1, top)
try:
marginDistribution[layer] += grads
totalVarianceTensor[layer] = np.vstack((totalVarianceTensor[layer], np.array(inter).reshape(inter.shape[0], -1)))
except Exception as e:
marginDistribution[layer] = grads
totalVarianceTensor[layer] = np.array(inter).reshape(inter.shape[0], -1)
if self.network_scale == True:
return marginDistribution, {}
normWidth = {}
for layer in self.layers:
totalVariance[layer] = (trim_mean(np.var(totalVarianceTensor[layer].reshape(totalVarianceTensor[layer].shape[0], -1), axis = 0), proportiontocut=0.05))**(1/2)
normWidth[layer] = math.sqrt(np.prod(totalVarianceTensor[layer].shape[1:]))
marginDistribution[layer] = np.array(marginDistribution[layer])/(np.array(totalVariance[layer])+1e-7) #/np.sqrt(m_factor) #/totalVarianceTensor[layer].shape[1:][0]
return marginDistribution, normWidth
def distancefromMargin(self, batch, layer, top = 2):
'''
Fuction to calculate margin distance for a given layer
Parameters
----------
batch : tf.data.Dataset()
The batch over which to compute the margin distance. A tuple of tf.Tensor of the form (input data, labels)
layer : int
The layer for which to compute margin distance
top : int, optional
Index for which to compute margin. For example, top = 2 will compute the margins between the class with the highes and second-highest softmax scores
Returns
-------
grads : list
A list containing the scaled margin distances
np_out : np.array
An array containing the flattened intermediate feature vector
'''
if self.network_scale == True:
batch_ = tf.ones(shape = batch[0].shape)
else:
batch_ = batch[0]
with tf.GradientTape(persistent=True) as tape:
intermediateVal = [batch_]
tape.watch(intermediateVal)
for l, layer_ in enumerate(self.model.layers):
intermediateVal.append(layer_(intermediateVal[-1]))
out_hard = tf.math.top_k(tf.nn.softmax(intermediateVal[-1], axis = 1), k = top)[1]
top_1 = out_hard[:,top-2]
misclassified = np.where(top_1 != batch[1])
if self.penalize:
top_1_og = out_hard[:,top-2]
top_2_og = out_hard[:,top-1]
mask = np.array(top_1_og == batch[1]).astype(int)
top_1 = top_1_og*mask + batch[1]*(1.- mask)
top_2 = top_2_og*mask + top_1_og*(1.- mask)
else:
top_1 = out_hard[:,top-2]
top_2 = out_hard[:,top-1]
mask = np.array(top_1 == batch[1]).astype(int)
top_2 = top_2*mask + batch[1]*(1.- mask)
index = list(range(batch[0].shape[0]))
index1 = np.array(list(zip(index, top_1)))
index2 = np.array(list(zip(index, top_2)))
preds = intermediateVal[-1]
logit1 = tf.gather_nd(preds, tf.constant(index1, tf.int32))
logit2 = tf.gather_nd(preds, tf.constant(index2, tf.int32))
if self.network_scale == True:
grad_i = tape.gradient(intermediateVal[-1], intermediateVal[layer])
grad_diff = (np.reshape(grad_i.numpy(), (self.batchSize, -1)))
denominator = np.linalg.norm(grad_diff, axis = 1, ord=2)
np_out = np.array(intermediateVal[layer])
print(denominator, np.mean(grad_i**2))
return denominator, np_out, grad_diff
else:
grad_i = tape.gradient(logit1, intermediateVal[layer])
grad_j = tape.gradient(logit2, intermediateVal[layer])
numerator = tf.gather_nd(preds, tf.constant(index1, tf.int32)) - tf.gather_nd(preds, tf.constant(index2, tf.int32)).numpy()
grad_diff = (np.reshape(grad_i.numpy(), (grad_i.numpy().shape[0], -1)) - np.reshape(grad_j.numpy(), (grad_j.numpy().shape[0], -1)))
denominator = np.linalg.norm(grad_diff, axis = 1, ord=2)
inf = np.linalg.norm(grad_diff, axis = 1, ord=np.inf)
if self.penalize == False:
numerator = np.delete(numerator, misclassified, axis = 0)
denominator = np.delete(denominator, misclassified, axis = 0)
if self.metric == 'spectral':
grads = numerator
else:
grads = numerator/(denominator+1e-7)
np_out = np.array(intermediateVal[layer])
gc.collect()
return list(grads), np_out
# ====================================== Functions for Mixup Based Solution ======================================
def batchMixup(self, batch, seed=1):
'''
Fuction to perform mixup on a batch of data
Parameters
----------
batch : tf.data.Dataset()
The batch over which to compute the margin distance. A tuple of tf.Tensor of the form (input data, labels)
seed : int, optional
Random seed
Returns
-------
tf.tensor
The mixed-up batch
'''
np.random.seed(seed)
x = batch[0]
mix_x = batch[0].numpy()
for i in range(np.max(batch[1])):
lam = (np.random.randint(0, 3, size=x[batch[1] == i].shape[0])/10)[..., None, None, None]
mix_x[(batch[1] == i)] = x[batch[1] == i]*lam + tf.random.shuffle(x[batch[1] == i], seed=seed)*(1-lam)
if self.augment == 'mixup':
mix_x = mix_x[np.random.randint(0, mix_x.shape[0], size=self.batchSize//3)]
target = batch[1].numpy()[np.random.randint(0, mix_x.shape[0], size=self.batchSize//3)]
return (tf.convert_to_tensor(mix_x), tf.convert_to_tensor(target))
else:
return tf.convert_to_tensor(mix_x)
def batchMixupLabelwiseLinear(self, batch, seed=2):
np.random.seed(seed)
labels = batch[1]
sorted_indices = np.argsort(batch[1])
sorted_labels = batch[1].numpy()[sorted_indices]
sorted_images = batch[0].numpy()[sorted_indices]
edges = np.array([len(sorted_labels[sorted_labels == i]) for i in range(max(sorted_labels)+1)])
edges = [0] + list(np.cumsum(edges))
shuffled_indices = []
for i in range(len(edges)-1):
# print(sorted_indices[edges[i]:edges[i+1]], sorted_labels[edges[i]:edges[i+1]])
shuffled_indices += list(np.random.choice(list(range(edges[i], edges[i+1])), replace=False, size=edges[i+1] - edges[i]))
# print(shuffled_indices[edges[i]:edges[i+1]])
intrapolateImages = (sorted_images + sorted_images[shuffled_indices])/2
return (tf.convert_to_tensor(intrapolateImages), tf.convert_to_tensor(sorted_labels))
def batchMixupLabelwise(self, batch, seed=1):
'''
Fuction to perform label-wise mixup on a batch of data
Parameters
----------
batch : tf.data.Dataset()
The batch over which to compute the margin distance. A tuple of tf.Tensor of the form (input data, labels)
seed : int, optional
Random seed
Returns
-------
tf.tensor
The label-wise mixed-up batch
'''
np.random.seed(seed)
def intrapolateImages(img, alpha=0.5):
temp = np.stack([img]*img.shape[0])
try:
tempT = np.transpose(temp, axes = (1,0,2,3,4))
except:
tempT = np.transpose(temp, axes = (1,0,2,3))
ret = alpha*temp + (1-alpha)*tempT
mask = np.triu_indices(img.shape[0], 1)
return ret[mask]
def randomSample(batch, size):
indices = np.random.randint(0, batch.shape[0], size=size)
return batch[indices]
for label in range(1+np.max(batch[1].numpy())):
try:
img = batch[0][batch[1]==label]
lbl = batch[1][batch[1]==label]
try:
mixedBatch = np.vstack((mixedBatch, randomSample(intrapolateImages(img), img.shape[0])))
labels = np.concatenate((labels, lbl))
except Exception as e:
mixedBatch = randomSample(intrapolateImages(img), img.shape[0])
labels = lbl
except:
img = batch[0][batch[1]==label]
lbl = batch[1][batch[1]==label]
try:
mixedBatch = np.vstack((mixedBatch, img))
labels = np.concatenate((labels, lbl))
except:
mixedBatch = img
labels = lbl
return (tf.convert_to_tensor(mixedBatch), tf.convert_to_tensor(labels))
# ====================================== Utility Functions ======================================
def intermediateOutputs(self, batch, layer=None, mode=None):
'''
Fuction to get intermadiate feature vectors
Parameters
----------
batch : tf.Tensor
A batch of data
layer : int, optional
The layer for which to get intermediate features
mode : str, optional
'pre' to create a pre-model which takes in the input data and gives out intermediate activations,
'post' to take in intermediate activations and give out the model predictions
Returns
-------
tf.keras.Model()
An extractor model
'''
model_ = keras.Sequential()
model_.add(keras.Input(shape=(batch[0][0].shape)))
for layer_ in self.model.layers:
model_.add(layer_)
if layer is not None and mode=='pre':
if layer >= 0:
extractor = keras.Model(inputs=self.model.layers[0].input,
outputs=self.model.layers[layer].output)
else:
extractor = keras.Model(inputs=self.model.layers[0].input,
outputs=self.model.layers[0].input)
elif layer is not None and mode=='post':
input_ = keras.Input(shape = (self.model.layers[layer].input.shape[1:]))
next_layer = input_
for layer in self.model.layers[layer:layer+2]:
next_layer = layer(next_layer)
extractor = keras.Model(input_, next_layer)
else:
extractor = keras.Model(inputs=self.model.layers[0].input,
outputs=[layer.output for layer in self.model.layers])
return extractor