-
Notifications
You must be signed in to change notification settings - Fork 0
/
superposition.py
285 lines (235 loc) · 13.3 KB
/
superposition.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
import numpy as np
import torch
import math
from help_functions import *
def random_binary_array(size):
"""
Create an array of 'size' length consisting only of numbers -1 and 1 (approximately 50% each).
:param size: length of the created array
:return: binary numpy array with values -1 or 1
"""
vec = np.random.uniform(-1, 1, size)
vec[vec < 0] = -1
vec[vec >= 0] = 1
return vec
def create_context_vectors(model, num_tasks, element_wise, use_PSP=False):
"""
Create random binary context vectors for all model layers.
Return together with layer dimension side, which is a list of 0 (first dimension taken for context size)
and 1 (second dimension taken for context size).
:param model: torch model instance
:param num_tasks: number of tasks
:param element_wise: boolean - if True, the number of context values in self attention part is the same as number of parameters
:param use_PSP: boolean - if True, PSP method is used, meaning we need set of contexts for each task (including the first)
:return: context_vectors (shape=(num_tasks-1, num of model layers)), layer_dimension (length=num of model layers)
"""
context_vectors = []
layer_dimension = []
n = num_tasks if use_PSP else num_tasks - 1
for t in range(n): # our contexts only needed between tasks, i.e. len(contexts)=num_task-1
task_contexts = []
for name, params in model.named_parameters():
if name.endswith('weight'): # only weight, not bias
if 'conv' in name:
vector_size = math.prod(params.size())
if t == 0:
layer_dimension.append(2) # to apply element-wise product
elif 'self_attn' not in name: # FC layer
vector_size = params.size()[1]
if t == 0:
layer_dimension.append(1)
else: # not FC layer (e.g., Wq, Wk, Wv in multi-head attention)
if element_wise:
vector_size = params.size()[0] * params.size()[1]
if t == 0:
layer_dimension.append(2)
else:
vector_size = params.size()[0]
if t == 0:
layer_dimension.append(0)
binary_context_vector = random_binary_array(vector_size)
task_contexts.append(binary_context_vector)
context_vectors.append(task_contexts)
# # Ablate a single W inside a multi-head attention layer
# for con_vec in context_vectors:
# con_vec[0][1024:2048] = np.ones(1024)
return context_vectors, layer_dimension
def context_multiplication(model, contexts, layer_dimension, task_index):
"""
Perform context multiplication of parameters in model.
:param model: torch model instance
:param contexts: binary context vectors (shape=(num_tasks-1, num of model layers))
:param layer_dimension: list of 0 (first dimension taken for context size), 1 (second dimension taken), 2 (element-wise)
:param task_index: index of the current task, which has finished learning
:return: None (but model parameters are updated)
"""
# Ablation study (layers 0 - 5)
first_ablated_layer = -1 # -1 means no ablation
last_ablated_layer = -1 # -1 means no ablation
ablated_layers = list(range(first_ablated_layer, last_ablated_layer + 1))
layer_index = 0
for name, params in model.named_parameters():
if name.endswith('weight'): # only weight, not bias
if layer_index not in ablated_layers: # if layer is not ablated
with torch.no_grad():
if layer_dimension[layer_index] == 0:
context_matrix = torch.from_numpy(np.diag(contexts[task_index][layer_index]).astype(np.float32)).cuda()
new_params = torch.matmul(context_matrix, params)
elif layer_dimension[layer_index] == 1:
context_matrix = torch.from_numpy(np.diag(contexts[task_index][layer_index]).astype(np.float32)).cuda()
new_params = torch.matmul(params, context_matrix)
elif layer_dimension[layer_index] == 2: # element-wise multiplication
context_matrix = torch.from_numpy(np.reshape(contexts[task_index][layer_index],
newshape=params.size()).astype(np.float32)).cuda()
new_params = params * context_matrix
else:
raise ValueError('Layer dimension must be 0, 1 or 2.')
params.copy_(new_params)
layer_index += 1
def evaluate_results(model, contexts, layer_dimension, all_tasks_test_data, superposition, task_index, first_average, use_MLP, batch_size, use_PSP=False):
"""
Evaluate the results on test data with or without using superposition. Return accuracy, AUROC and AUPRC.
:param model: torch model instance
:param contexts: binary context vectors (shape=(num_tasks-1, num of model layers))
:param layer_dimension: list of 0 (first dimension taken for context size) and 1 (second dimension taken)
:param all_tasks_test_data: list of all test data [X_test, y_test, mask_test] until the current task index
:param superposition: boolean - True, if superposition is used
:param task_index: index of the current task, which is being learned
:param first_average: string - show results on 'first' task only or the 'average' results until current task index
:param use_MLP: boolean - if True use MLP, else use Transformer
:param batch_size: batch size
:param use_PSP: boolean - if True, PSP method is used, meaning we need set of contexts for each task (including the first)
:return: accuracy, AUROC, AUPRC
"""
if superposition: # superposition used
if first_average == 'first': # not implemented for PSP
# unfold network parameters to the first task
for task_i in range(task_index - 1, -1, -1):
context_multiplication(model, contexts, layer_dimension, task_i)
# evaluate the model on the first task
acc, auroc, auprc = evaluate_current_task(model, all_tasks_test_data, 0, use_MLP)
# restore model parameters to the old ones (before context multiplication)
for task_i in range(task_index):
context_multiplication(model, contexts, layer_dimension, task_i)
return acc, auroc, auprc
elif first_average == 'average':
if use_PSP:
return evaluate_tasks_average_PSP(model, all_tasks_test_data, contexts, layer_dimension, task_index, use_MLP)
else:
return evaluate_tasks_average(model, all_tasks_test_data, contexts, layer_dimension,
superposition, task_index, use_MLP, batch_size)
else:
raise ValueError('The value of "first_average" has to be string "first" or "average".')
else: # superposition not used
if first_average == 'first':
return evaluate_current_task(model, all_tasks_test_data, 0, use_MLP)
elif first_average == 'average':
return evaluate_tasks_average(model, all_tasks_test_data, contexts, layer_dimension,
superposition, task_index, use_MLP, batch_size)
else:
raise ValueError('The value of "first_average" has to be string "first" or "average".')
def evaluate_current_task(model, all_tasks_test_data, task_index, use_MLP):
"""
Evaluate results on the first task, using the current model.
:param model: torch model instance
:param all_tasks_test_data: list of all test dataloaders ([X_test, y_test, mask_test]) until the current task index
:param task_index: index of the current task
:param use_MLP: boolean - if True use MLP, else use Transformer
:return: accuracy, AUROC, AUPRC
"""
curr_test_loader = all_tasks_test_data[task_index]
y = curr_test_loader.dataset.tensors[1].cuda()
model.eval()
with torch.no_grad():
test_outputs = []
for batch_X, batch_y, batch_mask in curr_test_loader:
if torch.cuda.is_available():
batch_X = batch_X.cuda()
batch_mask = batch_mask.cuda()
if use_MLP:
outputs = model.forward(batch_X)
else:
outputs = model.forward(batch_X, batch_mask)
test_outputs.append(outputs)
acc, auroc, auprc = get_stats(test_outputs, y)
return acc * 100, auroc * 100, auprc * 100
def evaluate_current_task_PSP(model, all_tasks_test_data, task_index, use_MLP, contexts, layer_dimension):
"""
Evaluate results on the first task, using the current PSP model.
:param model: torch model instance
:param all_tasks_test_data: list of all test dataloaders ([X_test, y_test, mask_test]) until the current task index
:param task_index: index of the current task
:param use_MLP: boolean - if True use MLP, else use Transformer
:param contexts: binary context vectors (shape=(num_tasks, num of model layers))
:param layer_dimension: list of 0 (first dimension taken for context size) and 1 (second dimension taken)
:return: accuracy, AUROC, AUPRC
"""
curr_test_loader = all_tasks_test_data[task_index]
y = curr_test_loader.dataset.tensors[1].cuda()
model.eval()
with torch.no_grad():
test_outputs = []
for batch_X, batch_y, batch_mask in curr_test_loader:
if torch.cuda.is_available():
batch_X = batch_X.cuda()
batch_mask = batch_mask.cuda()
if use_MLP:
outputs = model.forward(batch_X, True, contexts, task_index)
else:
outputs = model.forward(batch_X, batch_mask, True, contexts, task_index)
test_outputs.append(outputs)
acc, auroc, auprc = get_stats(test_outputs, y)
return acc * 100, auroc * 100, auprc * 100
def evaluate_tasks_average(model, all_tasks_test_data, contexts, layer_dimension, superposition, task_index, use_MLP, batch_size):
"""
Evaluate average results until the current task, using the current model.
:param model: torch model instance
:param all_tasks_test_data: list of all test data [X_test, y_test, mask_test] until the current task index
:param contexts: binary context vectors (shape=(num_tasks-1, num of model layers))
:param layer_dimension: list of 0 (first dimension taken for context size) and 1 (second dimension taken)
:param superposition: boolean - True, if superposition is used
:param task_index: index of the current task, which is being learned
:param use_MLP: boolean - if True use MLP, else use Transformer
:param batch_size: batch size
:return: mean accuracy, mean AUROC, mean AUPRC (across tasks)
"""
accs, aurocs, auprcs = [], [], []
if superposition:
for task_i in range(task_index, -1, -1): # iterate across tasks backwards
# evaluate results on the current task
acc, auroc, auprc = evaluate_current_task(model, all_tasks_test_data, task_i, use_MLP)
accs.append(acc)
aurocs.append(auroc)
auprcs.append(auprc)
# context multiplication to the previous task
if task_i > 0: # because we do not perform multiplication before the first task
context_multiplication(model, contexts, layer_dimension, task_i - 1) # task_i - 1, because contexts are only used between tasks
# restore model parameters to the old ones (before context multiplication)
for task_i in range(task_index): # iterate across tasks forward
context_multiplication(model, contexts, layer_dimension, task_i)
else:
for i in range(len(all_tasks_test_data)):
acc, auroc, auprc = evaluate_current_task(model, all_tasks_test_data, i, use_MLP)
accs.append(acc)
aurocs.append(auroc)
auprcs.append(auprc)
return np.mean(accs), np.mean(aurocs), np.mean(auprcs)
def evaluate_tasks_average_PSP(model, all_tasks_test_data, contexts, layer_dimension, task_index, use_MLP):
"""
Evaluate average results until the current task, using the current PSP model.
:param model: torch model instance
:param all_tasks_test_data: list of all test data [X_test, y_test, mask_test] until the current task index
:param contexts: binary context vectors (shape=(num_tasks, num of model layers))
:param layer_dimension: list of 0 (first dimension taken for context size) and 1 (second dimension taken)
:param task_index: index of the current task, which is being learned
:param use_MLP: boolean - if True use MLP, else use Transformer
:return: mean accuracy, mean AUROC, mean AUPRC (across tasks)
"""
accs, aurocs, auprcs = [], [], []
for task_i in range(task_index, -1, -1): # iterate across tasks backwards
# evaluate results on the current task
acc, auroc, auprc = evaluate_current_task_PSP(model, all_tasks_test_data, task_i, use_MLP, contexts, layer_dimension)
accs.append(acc)
aurocs.append(auroc)
auprcs.append(auprc)
return np.mean(accs), np.mean(aurocs), np.mean(auprcs)