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 | class Architect(object):
"""
A learner operating on the architecture weights of a DARTS model.
This learner handles training the weights associated with mixture operations
(architecture weights).
"""
def __init__(
self,
model: Network,
arch_learning_rate_max: float,
arch_momentum: float,
arch_weight_decay: float,
arch_weight_decay_df: float = 0,
arch_weight_decay_base: float = 0,
fair_darts_loss_weight: float = 1,
):
"""
Initializes the architecture learner.
Arguments:
model: a network model implementing the full DARTS model.
arch_learning_rate_max: learning rate for the architecture weights
arch_momentum: arch_momentum used in the Adam optimizer for architecture weights
arch_weight_decay: general weight decay for the architecture weights
arch_weight_decay_df: (weight decay applied to architecture weights in proportion
to the number of parameters of an operation)
arch_weight_decay_base: (a constant weight decay applied to architecture weights)
fair_darts_loss_weight: (a regularizer that pushes architecture weights more toward
zero or one in the fair DARTS variant)
"""
# set parameters for architecture learning
self.network_arch_momentum = arch_momentum
self.network_weight_decay = arch_weight_decay
self.network_weight_decay_df = arch_weight_decay_df
self.arch_weight_decay_base = arch_weight_decay_base * model._steps
self.fair_darts_loss_weight = fair_darts_loss_weight
self.model = model
self.lr = arch_learning_rate_max
# architecture is optimized using Adam
self.optimizer = torch.optim.Adam(
self.model.arch_parameters(),
lr=arch_learning_rate_max,
betas=(0.5, 0.999),
weight_decay=arch_weight_decay,
)
# initialize weight decay matrix
self._init_decay_weights()
# initialize the logged loss
self.current_loss = 0
def _init_decay_weights(self):
"""
This function initializes the weight decay matrix. The weight decay matrix
is subtracted from the architecture weight matrix on every learning step. The matrix
specifies a weight decay which is proportional to the number of parameters used in an
operation.
"""
n_params = list()
for operation in self.model.cells._ops[0]._ops:
if isiterable(operation):
n_params_total = (
1 # any non-zero operation is counted as an additional parameter
)
for subop in operation:
for parameter in subop.parameters():
if parameter.requires_grad is True:
n_params_total += parameter.data.numel()
else:
n_params_total = 0 # no operation gets zero parameters
n_params.append(n_params_total)
self.decay_weights = Variable(
torch.zeros(self.model.arch_parameters()[0].data.shape)
)
for idx, param in enumerate(n_params):
if param > 0:
self.decay_weights[:, idx] = (
param * self.network_weight_decay_df + self.arch_weight_decay_base
)
else:
self.decay_weights[:, idx] = param
self.decay_weights = self.decay_weights
self.decay_weights = self.decay_weights.data
def _compute_unrolled_model(
self,
input: torch.Tensor,
target: torch.Tensor,
eta: float,
network_optimizer: torch.optim.Optimizer,
):
"""
Helper function used to compute the approximate architecture gradient.
Arguments:
input: input patterns
target: target patterns
eta: learning rate
network_optimizer: optimizer used to updating the architecture weights
Returns:
unrolled_model: the unrolled architecture
"""
loss = self.model._loss(input, target)
theta = _concat(self.model.parameters()).data
try:
moment = _concat(
network_optimizer.state[v]["momentum_buffer"]
for v in self.model.parameters()
).mul_(self.network_arch_momentum)
except Exception:
moment = torch.zeros_like(theta)
dtheta = (
_concat(torch.autograd.grad(loss, self.model.parameters())).data
+ self.network_weight_decay * theta
)
unrolled_model = self._construct_model_from_theta(
theta.sub(eta, moment + dtheta)
)
return unrolled_model
def step(
self,
input_valid: torch.Tensor,
target_valid: torch.Tensor,
network_optimizer: torch.optim.Optimizer,
unrolled: bool,
input_train: Optional[torch.Tensor] = None,
target_train: Optional[torch.Tensor] = None,
eta: float = 1,
):
"""
Updates the architecture parameters for one training iteration
Arguments:
input_valid: input patterns for validation set
target_valid: target patterns for validation set
network_optimizer: optimizer used to updating the architecture weights
unrolled: whether to use the unrolled architecture or not (i.e., whether to use
the approximate architecture gradient or not)
input_train: input patterns for training set
target_train: target patterns for training set
eta: learning rate for the architecture weights
"""
# input_train, target_train only needed for approximation (unrolled=True)
# of architecture gradient
# when performing a single weigh update
# initialize gradients to be zero
self.optimizer.zero_grad()
# use different backward step depending on whether to use
# 2nd order approximation for gradient update
if unrolled: # probably using eta of parameter update here
self._backward_step_unrolled(
input_train,
target_train,
input_valid,
target_valid,
eta,
network_optimizer,
)
else:
self._backward_step(input_valid, target_valid)
# move Adam one step
self.optimizer.step()
# backward step (using non-approximate architecture gradient, i.e., full training)
def _backward_step(self, input_valid: torch.Tensor, target_valid: torch.Tensor):
"""
Computes the loss and updates the architecture weights assuming full optimization
of coefficients for the current architecture.
Arguments:
input_valid: input patterns for validation set
target_valid: target patterns for validation set
"""
if self.model.DARTS_type == DARTSType.ORIGINAL:
loss = self.model._loss(input_valid, target_valid)
elif self.model.DARTS_type == DARTSType.FAIR:
loss1 = self.model._loss(input_valid, target_valid)
loss2 = -F.mse_loss(
torch.sigmoid(self.model.alphas_normal),
0.5 * torch.ones(self.model.alphas_normal.shape, requires_grad=False),
) # torch.tensor(0.5, requires_grad=False)
loss = loss1 + self.fair_darts_loss_weight * loss2
else:
raise Exception(
"DARTS Type " + str(self.model.DARTS_type) + " not implemented"
)
loss.backward()
self.current_loss = loss.item()
# weight decay proportional to degrees of freedom
for p in self.model.arch_parameters():
p.data.sub_((self.decay_weights * self.lr)) # weight decay
# backward pass using second order approximation
def _backward_step_unrolled(
self,
input_train: torch.Tensor,
target_train: torch.Tensor,
input_valid: torch.Tensor,
target_valid: torch.Tensor,
eta: float,
network_optimizer: torch.optim.Optimizer,
):
"""
Computes the loss and updates the architecture weights using the approximate architecture
gradient.
Arguments:
input_train: input patterns for training set
target_train: target patterns for training set
input_valid: input patterns for validation set
target_valid: target patterns for validation set
eta: learning rate
network_optimizer: optimizer used to updating the architecture weights
"""
# gets the model
unrolled_model = self._compute_unrolled_model(
input_train, target_train, eta, network_optimizer
)
if self.model.DARTS_type == DARTSType.ORIGINAL:
unrolled_loss = unrolled_model._loss(input_valid, target_valid)
elif self.model.DARTS_type == DARTSType.FAIR:
loss1 = self.model._loss(input_valid, target_valid)
loss2 = -F.mse_loss(
torch.sigmoid(self.model.alphas_normal),
torch.tensor(0.5, requires_grad=False),
)
unrolled_loss = loss1 + self.fair_darts_loss_weight * loss2
else:
raise Exception(
"DARTS Type " + str(self.model.DARTS_type) + " not implemented"
)
unrolled_loss.backward()
dalpha = [v.grad for v in unrolled_model.arch_parameters()]
vector = [v.grad.data for v in unrolled_model.parameters()]
implicit_grads = self._hessian_vector_product(vector, input_train, target_train)
for g, ig in zip(dalpha, implicit_grads):
g.data.sub_(eta, ig.data)
for v, g in zip(self.model.arch_parameters(), dalpha):
if v.grad is None:
v.grad = Variable(g.data)
else:
v.grad.data.copy_(g.data)
def _construct_model_from_theta(self, theta: torch.Tensor):
"""
Helper function used to compute the approximate gradient update
for the architecture weights.
Arguments:
theta: term used to compute approximate gradient update
"""
model_new = self.model.new()
model_dict = self.model.state_dict()
params, offset = {}, 0
for k, v in self.model.named_parameters():
v_length = np.prod(v.size())
params[k] = theta[offset : (offset + v_length)].view(v.size())
offset += v_length
assert offset == len(theta)
model_dict.update(params)
model_new.load_state_dict(model_dict)
return model_new # .cuda() # Edit SM 10/26/19: uncommented for cuda
# second order approximation of architecture gradient (see Eqn. 8 from Liu et al, 2019)
def _hessian_vector_product(
self, vector: torch.Tensor, input: torch.Tensor, target: torch.Tensor, r=1e-2
):
"""
Helper function used to compute the approximate gradient update
for the architecture weights. It computes the hessian vector product outlined in Eqn. 8
from Liu et al, 2019.
Arguments:
vector: input vector
input: input patterns
target: target patterns
r: coefficient used to compute the hessian vector product
"""
R = r / _concat(vector).norm()
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
loss = self.model._loss(input, target)
grads_p = torch.autograd.grad(loss, self.model.arch_parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.sub_(2 * R, v)
loss = self.model._loss(input, target)
grads_n = torch.autograd.grad(loss, self.model.arch_parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
# this implements Eqn. 8 from Liu et al. (2019)
return [(x - y).div_(2 * R) for x, y in zip(grads_p, grads_n)]
|