-
Notifications
You must be signed in to change notification settings - Fork 0
/
rotated_EWC_classifier.py
196 lines (174 loc) · 8.89 KB
/
rotated_EWC_classifier.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
from continual_classifier import ContinualClassifier
from utils import estimate_fisher_diagonal
from keras.losses import categorical_crossentropy
import numpy as np
from keras.models import Model
from keras.layers import Dense, Activation
import tensorflow as tf
import keras.backend as K
from sklearn.utils import shuffle
from tqdm import tqdm
class RotatedEWCClassifier(ContinualClassifier):
def __init__(self, ewc_lambda=1, fisher_n=0, rotate_n=0, empirical=False, true_lambda=True, *args, **kwargs):
self.ewc_lambda = ewc_lambda
self.means = []
self.precisions = []
self.task_count = 0
self.fisher_n=fisher_n
self.rotate_n=rotate_n
self.empirical=empirical
self.true_lambda = true_lambda
self.U1 = {}
self.U2 = {}
super().__init__(*args, **kwargs)
def save_model_method(self, objs):
objs['ewc_lambda']=self.ewc_lambda
objs['means']=self.means
objs['precisions']=self.precisions
objs['task_count']=self.task_count
objs['fisher_n']=self.fisher_n
objs['rotate_n']=self.rotate_n
objs['empirical']=self.empirical
objs['true_lambda']=self.true_lambda
def load_model_method(self, objs):
self.ewc_lambda=objs['ewc_lambda']
self.means=objs['means']
self.precisions=objs['precisions']
self.task_count=objs['task_count']
self.fisher_n=objs['fisher_n']
self.rotate_n=objs['rotate_n']
self.empirical=objs['empirical']
self.true_lambda=objs['true_lambda']
self.U1={}
self.U2={}
def _task_fit(self, X, Y, model, new_task, batch_size, epochs, validation_data=None, verbose=2):
if self.task_count == 0: #first task
model.compile(loss=self.loss,optimizer=self.optimizer,metrics=['accuracy'])
model.fit(X,Y, batch_size=batch_size, epochs=epochs, validation_data = validation_data, verbose=verbose, shuffle=True)
self.task_count+=1
elif new_task: #non-first task
transformed_model = self.transform(model)
if self.empirical:
self.update_laplace_approxiation_parameters(transformed_model,X,Y)
else:
self.update_laplace_approxiation_parameters(transformed_model,X)
self.inject_regularization(self.EWC,transformed_model)
self.task_count+=1
transformed_model.compile(loss=self.loss,optimizer=self.optimizer,metrics=['accuracy'])
transformed_model.fit(X,Y, batch_size=batch_size, epochs=epochs, validation_data = validation_data, verbose=verbose, shuffle=True)
self.combine(model,transformed_model)
else: #old task
transformed_model = self.transform(model)
self.inject_regularization(self.EWC,transformed_model)
transformed_model.compile(loss=self.loss,optimizer=self.optimizer,metrics=['accuracy'])
transformed_model.fit(X,Y, batch_size=batch_size, epochs=epochs, validation_data = validation_data, verbose=verbose, shuffle=True)
self.combine(model,transformed_model)
self.rotate(model,X,Y,self.rotate_n)
def rotate(self,model,X,Y,rotate_n=0):
if rotate_n is 0 or rotate_n>X.shape[0]:
rotate_n = X.shape[0]
rotated_layers=[]
layer_inputs=[]
layer_outputs=[]
for l in model.layers:
if ('Dense' in repr(l) or 'Conv2D' in repr(l)) and 'softmax' not in repr(l.activation) and 'Softmax' not in repr(l.output):
rotated_layers.append(l)
layer_inputs.append(l.input)
layer_outputs.append(l.output)
input_sums = [np.zeros([l.input.shape[-1]]*2) for l in rotated_layers]
output_sums = [np.zeros([l.output.shape[-1]]*2) for l in rotated_layers]
X=X[0:rotate_n]
Y=Y[0:rotate_n]
X,Y = shuffle(X,Y)
label_tensor = tf.where(tf.equal(tf.reduce_max(model.output,1,keepdims=True),model.output),tf.constant(1,shape=(1,model.output.shape[1])),tf.constant(0,shape=(1,model.output.shape[1])))
grads_tensor = K.gradients(categorical_crossentropy(label_tensor,model.output),layer_outputs)
gradients = []
sess=K.get_session()
for i in tqdm(range(rotate_n), desc='Rotating'):
gradient = sess.run(grads_tensor, feed_dict={model.input:np.array([X[i]])})
for j in range(len(layer_outputs)):
if 'Dense' in repr(rotated_layers[j]):
output_sums[j]+=np.dot(gradient[j].transpose(),gradient[j])/rotate_n
inputs = sess.run(layer_inputs,feed_dict={model.input:np.array([X[i]])})
for j in range(len(layer_inputs)):
if 'Dense' in repr(rotated_layers[j]):
input_sums[j]+=np.dot(inputs[j].transpose(),inputs[j])/rotate_n
for i in tqdm(range(len(output_sums)),desc='Doing SVDs'):
if 'Dense' in repr(rotated_layers[j]):
self.U1[rotated_layers[i]]=np.linalg.svd(input_sums[i], full_matrices=False)[0]
self.U2[rotated_layers[i]]=np.linalg.svd(output_sums[i], full_matrices=False)[0]
#order of U1, W, and U2 is U1.T@W@U2.T instead of U2.T@W@U1.T because keras dense implements a FC as y=xW+b
def _replace_dense(self,dense_layer,input_tensor):
U1 = self.U1[dense_layer]
U2 = self.U2[dense_layer]
name = dense_layer.name
u1l = Dense(int(input_tensor.shape[-1]),name=name+'_U1')
x=u1l(input_tensor)
u1w,u1b = u1l.get_weights()
u1w = U1
u1b = np.zeros_like(u1b)
u1l.set_weights([u1w,u1b])
u1l.trainable = False
transformed_l = Dense(int(dense_layer.output.shape[-1]),name=name+'_transformed')
x=transformed_l(x)
W,b = dense_layer.get_weights()
W = U1.transpose() @ W @ U2.transpose()
transformed_l.set_weights([W,b])
u2l = Dense(int(dense_layer.output.shape[-1]),name=name+'_U2')
x=u2l(x)
u2w,u2b = u2l.get_weights()
u2w = U2
u2b = np.zeros_like(u2b)
u2l.set_weights([u2w,u2b])
u2l.trainable = False
x = Activation(dense_layer.activation)(x)
return x
#non-sequential layer replacement code based on https://stackoverflow.com/a/54517478
def transform(self,model):
#if not dense or conv2d then just add the layer
#if dense/conv2d: replace it with 3 layers (named appropriately), followed by Activation
network_dict = {'input_layers_of': {}, 'new_output_tensor_of': {}}
for layer in model.layers:
for node in layer._outbound_nodes:
layer_name = node.outbound_layer.name
if layer_name not in network_dict['input_layers_of']:
network_dict['input_layers_of'].update({layer_name: [layer.name]})
else:
network_dict['input_layers_of'][layer_name].append(layer.name)
network_dict['new_output_tensor_of'].update({model.layers[0].name: model.input})
for layer in model.layers[1:]:
layer_input = [network_dict['new_output_tensor_of'][layer_aux] for layer_aux in network_dict['input_layers_of'][layer.name]]
if len(layer_input) == 1:
layer_input = layer_input[0]
if layer in self.U1.keys():
if 'Dense' in repr(layer):
x = self._replace_dense(layer, layer_input)
else:
pass
else:
x = layer(layer_input)
network_dict['new_output_tensor_of'].update({layer.name: x})
return Model(inputs=model.input,outputs=x)
def combine(self,model,transformed_model):
for l in model.layers:
if l in self.U1.keys():
U1 = self.U1[l]
U2 = self.U2[l]
transformed_W = transformed_model.get_layer(l.name+'_transformed').get_weights()[0]
b = l.get_weights()[1]
W = U1 @ transformed_W @ U2
l.set_weights([W,b])
def update_laplace_approxiation_parameters(self,model,X,Y=None):
len_weights = len(K.batch_get_value(model.trainable_weights))-2*(not self.singleheaded)
fisher_estimates = estimate_fisher_diagonal(model,X,Y,self.fisher_n,len_weights,False,self.true_lambda)
self.means.append(K.batch_get_value(model.trainable_weights))
self.precisions.append(fisher_estimates)
def EWC(self,weight_no):
task_count = self.task_count
if task_count is 0:
def ewc_reg(weights):
return 0
return ewc_reg
def ewc_reg(weights):
return self.ewc_lambda*0.5*K.sum((self.precisions[-1][weight_no]) * (weights-self.means[-1][weight_no])**2)
return ewc_reg