-
Notifications
You must be signed in to change notification settings - Fork 14
/
regularizers.py
427 lines (357 loc) · 16.9 KB
/
regularizers.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
import abc
import math
import torch
from torch.optim import Optimizer, SGD
from settings import args, FILL_VAL, TOKENS_WEIGHT
from utils import get_losses, get_model_dir
from parallel import DataParallelCriterion
from torch.nn import CrossEntropyLoss, MSELoss
import pickle as pkl
import os
from torch.nn.functional import softmax
class Regularizer(abc.ABC):
def __init__(self, model, parallel_model, dataloaders, task, prev_task=None):
self.model = model
self.parallel_model = parallel_model
self.dataloaders = dataloaders
self.task = task
self.prev_task = prev_task
@abc.abstractmethod
def task_start_do(self):
return NotImplemented
@abc.abstractmethod
def task_end_do(self):
return NotImplemented
def save_reg_params(self):
model_dir = get_model_dir([self.task])
reg_params_path = os.path.join(model_dir, "reg_params.pkl")
with open(reg_params_path, 'wb') as f:
pkl.dump(self.model.reg_params,f)
def load_reg_params(self):
if self.prev_task:
model_dir = get_model_dir([self.prev_task])
reg_params_path = os.path.join(model_dir, "reg_params.pkl")
with open(reg_params_path, 'rb') as f:
self.model.reg_params = pkl.load(f)
input()
class MAS(Regularizer):
def task_start_do(self,freeze_layers=[]):
#self.load_reg_params()
task_start_do(self.model, freeze_layers)
def task_end_do(self):
updater = Omega_update(self.model.parameters(), lr=0.0001, momentum=0.9)
compute_importance(self.model, self.parallel_model, updater, self.dataloaders)
accumulate_reg_params(self.model)
self.save_reg_params()
class EWC(Regularizer):
def task_start_do(self,freeze_layers=[]):
#self.load_reg_params()
task_start_do(self.model, freeze_layers)
def task_end_do(self):
updater = Omega_update(self.model.parameters(), lr=0.0001, momentum=0.9)
compute_importance(self.model, self.parallel_model, updater, self.dataloaders, loss_type="ewc")
accumulate_reg_params(self.model)
self.save_reg_params()
REG_TYPES = {
"mas": MAS,
"ewc": EWC,
}
args.REG_TYPE_KEYS = REG_TYPE_KEYS = list(REG_TYPES.keys())
def task_start_do(model, freeze_layers=[]):
if not hasattr(model,"reg_params"):
initialize_reg_params(model,freeze_layers)
else:
clean_omega_sum(model,freeze_layers)
def initialize_reg_params(model,freeze_layers=[]):
"""initialize an omega for each parameter to zero"""
reg_params={}
for name, param in model.named_parameters():
if not name in freeze_layers:
# print('initializing param',name)
omega=torch.FloatTensor(param.size()).zero_()
omega=omega.cuda()
init_val=param.data.clone()
init_val=init_val.cuda()
reg_param={}
reg_param['omega'] = omega
reg_param['omega_sum'] = omega
#initialize the initial value to that before starting training
reg_param['init_val'] = init_val
reg_params[param]=reg_param
if 'data_count' not in reg_params:
reg_params['data_count'] = 0
reg_params['lambda'] = args.reg_lambda
model.reg_params = reg_params
def clean_omega_sum(model,freeze_layers=[]):
for name, param in model.named_parameters():
if not name in freeze_layers:
omega=torch.FloatTensor(param.size()).zero_()
omega=omega.cuda()
reg_param = model.reg_params.get(param)
reg_param['omega_sum'] = omega
model.reg_params[param]=reg_param
model.reg_params['data_count'] = 0
class Weight_Regularized_AdamW(Optimizer):
""" Implements Adam algorithm with weight decay fix.
Parameters:
lr (float): learning rate. Default 1e-3.
betas (tuple of 2 floats): Adams beta parameters (b1, b2). Default: (0.9, 0.999)
eps (float): Adams epsilon. Default: 1e-6
weight_decay (float): Weight decay. Default: 0.0
correct_bias (bool): can be set to False to avoid correcting bias in Adam (e.g. like in Bert TF repository). Default True.
"""
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-6, weight_decay=0.0, correct_bias=True):
if lr < 0.0:
raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[1]))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(eps))
defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay,
correct_bias=correct_bias)
super(Weight_Regularized_AdamW, self).__init__(params, defaults)
def step(self, reg_params, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
reg_lambda=reg_params.get('lambda')
for group in self.param_groups:
for p in group['params']:
if p.grad is None:
continue
grad = p.grad.data
if grad.is_sparse:
raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p.data)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
# Decay the first and second moment running average coefficient
# In-place operations to update the averages at the same time
exp_avg.mul_(beta1).add_(1.0 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1.0 - beta2, grad, grad)
denom = exp_avg_sq.sqrt().add_(group['eps'])
step_size = group['lr']
if group['correct_bias']: # No bias correction for Bert
bias_correction1 = 1.0 - beta1 ** state['step']
bias_correction2 = 1.0 - beta2 ** state['step']
step_size = step_size * math.sqrt(bias_correction2) / bias_correction1
p.data.addcdiv_(-step_size, exp_avg, denom)
# Just adding the square of the weights to the loss function is *not*
# the correct way of using L2 regularization/weight decay with Adam,
# since that will interact with the m and v parameters in strange ways.
#
# Instead we want to decay the weights in a manner that doesn't interact
# with the m/v parameters. This is equivalent to adding the square
# of the weights to the loss with plain (non-momentum) SGD.
# Add weight decay at the end (fixed version)
#Regularize PART CODE GOES HERE
if p in reg_params:
reg_param=reg_params.get(p)
#get omega for this parameter
omega=reg_param.get('omega')
#initial value when the training start
init_val=reg_param.get('init_val')
curr_weight_val=p.data
#get the difference
weight_dif=curr_weight_val.add(-1,init_val)
#compute the MAS penalty
regulizer=weight_dif.mul(2*reg_lambda*omega)
del weight_dif
del curr_weight_val
del omega
del init_val
#add the MAS regulizer to the gradient
# grad.add_(regulizer)
p.data.add_(-group['lr'], regulizer)
del regulizer
#Regularize PART CODE ENDS
if group['weight_decay'] > 0.0:
p.data.add_(-group['lr'] * group['weight_decay'], p.data)
return loss
# update omega for one task; use in compute_importance
class Omega_update(SGD):
"""
Update the paramerter importance using the gradient of the function output norm. To be used at deployment time.
reg_params:parameters omega to be updated
batch_index,batch_size:used to keep a running average over the seen samples
"""
def __init__(self, params, lr=0.001, momentum=0, dampening=0, weight_decay=0, nesterov=False):
super(Omega_update, self).__init__(params,lr,momentum,dampening,weight_decay,nesterov)
def __setstate__(self, state):
super(Omega_update, self).__setstate__(state)
def step(self, reg_params, batch_size, closure=None):
"""
Performs a single parameters importance update setp
"""
#print('************************DOING A STEP************************')
reg_params['data_count'] += batch_size
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
#if the parameter has an omega to be updated
for p in group['params']:
#print('************************ONE PARAM************************')
if p.grad is None:
continue
if p in reg_params:
#HERE MAS IMPOERANCE UPDATE GOES
#get the gradient
unreg_dp = p.grad.data.clone()
reg_param = reg_params.get(p)
#get parameter omega
omega = reg_param.get('omega_sum')
if args.seq_train_type == "ewc":
omega = omega.add((unreg_dp)**2)
else:
omega = omega.add(unreg_dp.abs_())
reg_param['omega_sum'] = omega
reg_params[p] = reg_param
#HERE MAS IMPOERANCE UPDATE ENDS
return loss#HAS NOTHING TO DO
# update omega for one task
def compute_importance(model, parallel_model, updater, dataloaders, loss_type="l2"):
"""Mimic the depoloyment setup where the model is applied on some samples and those are used to update the importance params
Uses the L2norm of the function output. This is what we MAS uses as default
"""
# model.eval() # Set model to training mode so we get the gradient
# train_loss_fct = DataParallelCriterion(CrossEntropyLoss(ignore_index=FILL_VAL), args.device_ids)
softmax = torch.nn.Softmax(dim=-1)
if loss_type == "l2":
loss_fct = DataParallelCriterion(torch.nn.MSELoss(reduction='mean'), args.device_ids)
elif loss_type == "l1":
loss_fct = DataParallelCriterion(torch.nn.L1Loss(reduction='mean'), args.device_ids)
elif loss_type == "ewc":
CELoss = CrossEntropyLoss(ignore_index=FILL_VAL, reduction='mean', weight=TOKEN_WEIGHT)
loss_fct = DataParallelCriterion(CELoss, args.device_ids)
# Iterate over data.
for dataloader in dataloaders:
for cq, len_cq, cqa, len_cqa, Y, _, _ in dataloader:
# get the inputs
n_inputs = sum(len(_cq) for _cq in cq)
for i in range(len(cqa)):
cq[i] = (cq[i].to(args.device_ids[i]),)
len_cq[i] = len_cq[i].to(args.device_ids[i])
cqa[i] = (cqa[i].to(args.device_ids[i]),)
len_cqa[i] = len_cqa[i].to(args.device_ids[i])
Y[i] = Y[i].to(args.device_ids[i])
# zero the parameter gradients
updater.zero_grad()
# forward
if loss_type != "ewc":
logits = parallel_model(cq)
logits = [logit[range(len(logit)), len_cq[i]-1, :] for i, logit in enumerate(logits)]
#logits = [softmax(logit, dim=-1) for logit in logits]
target_zeros = [torch.zeros(logit.size()).to(args.device_ids[i]) for i, logit in enumerate(logits)]
logits = [softmax(logit) for logit in logits]
if loss_type == "l2":
targets = loss_fct(logits, target_zeros)
elif loss_type == "l1":
targets = loss_fct(logits, target_zeros)
else:
targets, _ = get_losses(parallel_model, cqa, Y, None, None, loss_fct)
targets /= n_inputs
#compute the gradients
targets.backward()
#update the parameters importance
updater.step(model.reg_params, n_inputs)
# omega of task1 + omega of task2 ...
# new_omega=omega_sum/data_count; omega=new_omega+prev_omega
def accumulate_reg_params(model, freeze_layers=[]):
"""accumelate the newly computed omega with the previously stroed one from the old previous tasks"""
for name, param in model.named_parameters():
if not name in freeze_layers:
if param in model.reg_params:
reg_param=model.reg_params.get(param)
# print('restoring previous omega',name)
prev_omega=reg_param.get('omega')
new_omega=reg_param.get('omega_sum') / model.reg_params["data_count"]
acc_omega=torch.add(prev_omega,new_omega)
del reg_param['omega_sum']
reg_param['omega'] = acc_omega
model.reg_params[param]=reg_param
del prev_omega
del new_omega
del acc_omega
else:
if param in model.reg_params:
reg_param=model.reg_params.get(param)
# print('removing unused omega',name)
del reg_param['omega']
del model.reg_params[param]
class Weight_Regularized_SGD(SGD):
r"""Implements SGD training with importance params regulization. IT inherents stochastic gradient descent (optionally with momentum).
Nesterov momentum is based on the formula from
"""
def __init__(self, params, lr=0.001, momentum=0, dampening=0, weight_decay=0, nesterov=False):
super(Weight_Regularized_SGD, self).__init__(params, lr,momentum,dampening,weight_decay,nesterov)
def __setstate__(self, state):
super(Weight_Regularized_SGD, self).__setstate__(state)
def step(self, reg_params,closure=None):
loss = None
if closure is not None:
loss = closure()
reg_lambda=reg_params.get('lambda')
for group in self.param_groups:
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
for p in group['params']:
if p.grad is None:
continue
d_p = p.grad.data
#MAS PART CODE GOES HERE
#if this param has an omega to use for regulization
if p in reg_params:
reg_param=reg_params.get(p)
#get omega for this parameter
omega=reg_param.get('omega')
#initial value when the training start
init_val=reg_param.get('init_val')
curr_wegiht_val=p.data
#move the tensors to cuda
init_val=init_val.cuda()
omega=omega.cuda()
#get the difference
weight_dif=curr_wegiht_val.add(-1,init_val)
#compute the MAS penalty
regulizer=weight_dif.mul(2*reg_lambda*omega)
del weight_dif
del curr_wegiht_val
del omega
del init_val
#add the MAS regulizer to the gradient
d_p.add_(regulizer)
del regulizer
#MAS PARAT CODE ENDS
if weight_decay != 0:
d_p.add_(weight_decay,p.data.sign())
if momentum != 0:
param_state = self.state[p]
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = d_p.clone()
else:
buf = param_state['momentum_buffer']
buf.mul_(momentum).add_(1 - dampening, d_p)
if nesterov:
d_p = d_p.add(momentum, buf)
else:
d_p = buf
p.data.add_(-group['lr'], d_p)
return loss