-
Notifications
You must be signed in to change notification settings - Fork 0
/
binarynet.py
411 lines (377 loc) · 24.1 KB
/
binarynet.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
import torch
from myoptimizer import ALQ_optimizer
REL_NORM_THRES = 1e-6
def construct_bit_table(bit):
"""Construct a look-up-table to store bitwise values of all intergers given a bitwidth."""
bit_table = -torch.ones((2**bit, bit), dtype=torch.int8)
for i in range(1,2**bit):
for j in range(bit):
if (i & (1<<j)):
bit_table[i,j] = 1
return bit_table.to('cuda')
def binarize(input_t):
"""Binarize input tensor."""
dim = input_t.nelement()
output_t = torch.ones(dim)
output_t[input_t<0] = -1
return output_t
def transform_bin_basis(w_vec, max_dim, rel_norm_thres=REL_NORM_THRES):
"""Transform a full precision weight vector into multi-bit form, i.e. binary bases and coordiantes."""
# Reshape the coordinates vector in w domain
crd_w = w_vec.detach().view(-1,1) # 展开成一个列向量
#print(crd_w)
# Get the dimensionality in w domain
dim_w = crd_w.nelement() #列向量的长度(16)
#print(dim_w)
# Determine the max number of dimensionality in alpha domain
if dim_w <= max_dim:
max_dim_alpha = dim_w
else:
max_dim_alpha = max_dim
# Initialize binary basis matrix in alpha domain
bin_basis_alpha = torch.zeros((dim_w, max_dim_alpha))
# Initialize coordinates vector in alpha domain
crd_alpha = torch.zeros(max_dim_alpha)
res = crd_w.detach()
res_L2Norm_square = torch.sum(torch.pow(res,2))
ori_L2Norm_square = torch.sum(torch.pow(crd_w,2))
for i in range(max_dim_alpha):
if res_L2Norm_square/ori_L2Norm_square < rel_norm_thres:
break
new_bin_basis = binarize(res.view(-1))
bin_basis_alpha[:,i] = new_bin_basis
B_ = bin_basis_alpha[:,:i+1]
# Find the optimal coordinates in the space spanned by B_
alpha_ = torch.mm(torch.inverse(torch.mm(torch.t(B_),B_)),torch.mm(torch.t(B_),crd_w))
# Compute the residual (orthogonal to the space spanned by B_)
res = crd_w - torch.mm(B_, alpha_)
crd_alpha[:i+1] = alpha_.view(-1)
res_L2Norm_square = torch.sum(torch.pow(res,2))
ind_neg = crd_alpha < 0
crd_alpha[ind_neg] = -crd_alpha[ind_neg]
bin_basis_alpha[:,ind_neg] = -bin_basis_alpha[:,ind_neg]
# Get the valid indexes
ind_valid = crd_alpha != 0
# Get the valid dimensionality in alpha domain
dim_alpha = torch.sum(ind_valid)
sorted_ind = torch.argsort(crd_alpha[ind_valid])
#print(dim_alpha) #alpha 的长度 6
#print(bin_basis_alpha) # 二值化网络权重 二维tensor
#print(ind_valid)
#print(sorted_ind)
if dim_alpha == 0:
return [], [], 0
else:
return bin_basis_alpha[:,ind_valid][:,sorted_ind].to(torch.int8), crd_alpha[ind_valid][sorted_ind], dim_alpha
class ConvLayer_bin(object):
"""This class defines the multi-bit form of the weight tensor of a convolutional layer used in ALQ.
Arguments:
w_ori (float tensor): the 3-dim pretrained weight tensor of a convolutional layer.
ind_layer (int): the index of this layer in the network.
structure (string): the structure used for grouping the weights in this layer, optional values: 'kernelwise', 'pixelwise', 'channelwise'.
max_bit (int): the maximum bitwidth used in initialization.
"""
def __init__(self, w_ori, ind_layer, structure, max_bit):
# The layer type
self.layer_type = 'conv'
# The shape of the weight tensor of this layer
self.tensor_shape = w_ori.size()
# The maintained full precision weight tensor of this layer used in STE
self.w_FP = w_ori.clone().to('cuda')
# The index of this layer in the network
self.ind_layer = ind_layer
# The structure used for grouping the weights in this layer
self.structure = structure
# The maximum bitwidth used in initialization
self.max_bit = max_bit
# The binary bases, the coordinates, and the mask (only for parallel computing purposes) of each group
self.B, self.alpha, self.mask = self.structured_sketch()
# The total number of binary filters in this layer, namely the total number of (valid) alpha's
self.num_bin_filter = torch.sum(self.mask)
# The average bitwidth of this layer
self.avg_bit = self.num_bin_filter.float()/(self.mask.size(0)*self.mask.size(1))
# The total number of weights of this layer
self.num_weight = self.w_FP.nelement()
# The used look-up-table for bitwise values
self.bit_table = construct_bit_table(self.max_bit)
def structured_sketch(self):
"""Initialize the weight tensor using structured sketching.
Namely, structure the weights in groupwise, and quantize each group's weights in multi-bit form w.r.t. the reconstruction error.
Return the binary bases, the coordinates, and the mask (only for parallel computing purposes) of each group.
"""
w_cpu = self.w_FP.to('cpu')
if self.structure == 'kernelwise':
B = torch.zeros((self.tensor_shape[0],self.tensor_shape[1],self.max_bit,self.tensor_shape[2])).to(torch.int8)
alpha = torch.zeros((self.tensor_shape[0],self.tensor_shape[1],self.max_bit,1)).to(torch.float32)
mask = torch.zeros((self.tensor_shape[0],self.tensor_shape[1],self.max_bit,1)).to(torch.bool)
elif self.structure == 'pixelwise':
B = torch.zeros((self.tensor_shape[0],self.tensor_shape[2],self.max_bit,self.tensor_shape[1])).to(torch.int8)
alpha = torch.zeros((self.tensor_shape[0],self.tensor_shape[2],self.max_bit,1)).to(torch.float32)
mask = torch.zeros((self.tensor_shape[0],self.tensor_shape[2],self.max_bit,1)).to(torch.bool)
elif self.structure == 'channelwise':
B = torch.zeros((self.tensor_shape[0],1,self.max_bit,self.tensor_shape[1]*self.tensor_shape[2])).to(torch.int8)
alpha = torch.zeros((self.tensor_shape[0],1,self.max_bit)).to(torch.float32)
mask = torch.zeros((self.tensor_shape[0],1,self.max_bit)).to(torch.bool)
for k in range(self.tensor_shape[0]):
if self.structure == 'kernelwise':
for q in range(self.tensor_shape[1]):
bin_basis, crd, dim = transform_bin_basis(w_cpu[k,q,:].view(-1), self.max_bit)
#print(crd) # tensor([0.0030, 0.0093, 0.0231, 0.0371, 0.0655, 0.1454])
#print(dim) # tensor(6)
mask[k,q,:dim,0] = 1
#print(B.shape) # torch.Size([8, 1, 6, 16])
#print(t(bin_basis).shape) ## 这里我进行了改动
#print(torch.t(bin_basis).shape) # torch.Size([6, 16])
B[k,q,:dim,:] = torch.t(bin_basis)
#print(alpha.shape)
alpha[k,q,:dim,0] = crd
elif self.structure == 'pixelwise':
for h in range(self.tensor_shape[2]):
for w in range(self.tensor_shape[3]):
bin_basis, crd, dim = transform_bin_basis(w_cpu[k,:,w].view(-1), self.max_bit)
mask[k,h*self.tensor_shape[3]+w,:dim,0] = 1
B[k,h*self.tensor_shape[3]+w,:dim,:] = torch.t(bin_basis)
alpha[k,h*self.tensor_shape[3]+w,:dim,0] = crd
if self.structure == 'channelwise':
bin_basis, crd, dim = transform_bin_basis(w_cpu[k,:,:,:].view(-1), self.max_bit)
mask[k,0,:dim,0] = 1
B[k,0,:dim,:] = torch.t(bin_basis)
alpha[k,0,:dim,0] = crd
return B.to('cuda'), alpha.to('cuda'), mask.to('cuda')
def reconstruct_w(self):
"""Reconstruct the weight tensor from the current quantization.
Return the reconstructed weight tensor of this layer, i.e. \hat{w}.
"""
#print(self.B.shape)
#print(self.alpha.shape)
#print(self.mask.shape)
w_bin = torch.sum(self.B.float()*(self.alpha*self.mask.float()),dim=2)
#print(w_bin)
#print(tensor_shape)
if self.structure == 'kernelwise':
return w_bin.reshape(self.tensor_shape)
elif self.structure == 'pixelwise':
return torch.transpose(w_bin,1,2).reshape(self.tensor_shape)
elif self.structure == 'channelwise':
return w_bin.reshape(self.tensor_shape)
def update_w_FP(self, w_FP_new=None):
"""Update the full precision weight tensor.
In STE with loss-aware optimization, w_FP is the maintained full precision weight tensor.
In ALQ optimization, w_FP is used to store the reconstructed weight tensor from the current quantization.
"""
if w_FP_new is not None:
self.w_FP.add_(w_FP_new)
else:
self.w_FP.zero_().add_(self.reconstruct_w())
def construct_grad_alpha(self, grad_w):
"""Compute and return the gradient (or the first momentum) in alpha domain w.r.t the loss.
"""
if self.structure == 'kernelwise':
return torch.matmul(self.B.float(), grad_w.reshape((self.tensor_shape[0],self.tensor_shape[1],-1,1)))*self.mask.float()
elif self.structure == 'pixelwise':
return torch.matmul(self.B.float(), torch.transpose(grad_w.reshape((self.tensor_shape[0],self.tensor_shape[1],-1,1)), 1,2) )*self.mask.float()
elif self.structure == 'channelwise':
return torch.matmul(self.B.float(), grad_w.reshape((self.tensor_shape[0],1,-1,1)))*self.mask.float()
def construct_hessian_alpha(self, diag_hessian_w):
"""Compute and return the diagonal Hessian (or the second momentum) in alpha domain w.r.t the loss.
"""
if self.structure == 'kernelwise':
diag_hessian = torch.matmul(self.B.float()*diag_hessian_w.reshape((self.tensor_shape[0],self.tensor_shape[1],1,-1)), torch.transpose(self.B,2,3).float())
return torch.diagonal(diag_hessian,dim1=-2,dim2=-1).unsqueeze(-1)*self.mask.float()
elif self.structure == 'pixelwise':
diag_hessian = torch.matmul(self.B.float()*torch.transpose(diag_hessian_w.reshape((self.tensor_shape[0],self.tensor_shape[1],1,-1)), 1,3), torch.transpose(self.B,2,3).float())
return torch.diagonal(diag_hessian,dim1=-2,dim2=-1).unsqueeze(-1)*self.mask.float()
elif self.structure == 'channelwise':
diag_hessian = torch.matmul(self.B.float()*diag_hessian_w.reshape((self.tensor_shape[0],1,1,-1)), torch.transpose(self.B,2,3).float())
return torch.diagonal(diag_hessian,dim1=-2,dim2=-1).unsqueeze(-1)*self.mask.float()
def sort_importance_bin_filter(self, grad_alpha, diag_hessian_alpha, num_top):
"""Compute and sort the importance of binary filters (alpha's) in this layer.
The importance is defined by the modeled loss increment caused by pruning each individual alpha.
Return the selected num_top alpha's with the least importance.
"""
delta_loss_prune = -grad_alpha*self.alpha+0.5*torch.pow(self.alpha,2)*diag_hessian_alpha
sorted_ind = torch.argsort(delta_loss_prune[self.mask].view(-1))
top_importance_list = torch.tensor([[self.ind_layer, sorted_ind[i], delta_loss_prune.view(-1)[sorted_ind[i]]] for i in range(num_top)])
return top_importance_list
def prune_alpha(self, ind_prune):
"""Prune the cooresponding alpha's of this layer give the indexes.
"""
num_bin_filter_ = torch.sum(self.mask)
self.mask.view(-1)[self.mask.view(-1).nonzero().view(-1)[ind_prune]]=0
self.B *= self.mask.char()
self.alpha *= self.mask.float()
self.num_bin_filter = torch.sum(self.mask)
self.avg_bit = self.num_bin_filter.float()/(self.mask.size(0)*self.mask.size(1))
if num_bin_filter_-self.num_bin_filter != ind_prune.size(0):
print('wrong pruning')
return False
return True
def optimize_bin_basis(self, pseudo_grad, pseudo_hessian):
"""Take one optimization step on the binary bases of this layer while fixing coordinates.
"""
# Compute the target weight tensor, i.e. the optimal point in w domain according to the quadratic model function
target_w = self.w_FP-pseudo_grad/pseudo_hessian
if self.structure == 'kernelwise':
all_disc_w = torch.matmul(self.bit_table.view((1,1,self.bit_table.size(0),self.bit_table.size(1))).float(),self.alpha)
ind_opt = torch.argmin(torch.abs(target_w.view((self.tensor_shape[0],self.tensor_shape[1],1,-1)) - all_disc_w), dim=2)
self.B = torch.transpose((self.bit_table[ind_opt.view(-1),:]).view(self.tensor_shape[0],self.tensor_shape[1],self.tensor_shape[2],self.max_bit), 2,3)
self.B *= self.mask.char()
elif self.structure == 'pixelwise':
all_disc_w = torch.matmul(self.bit_table.view((1,1,self.bit_table.size(0),self.bit_table.size(1))).float(),self.alpha)
ind_opt = torch.argmin(torch.abs(torch.transpose(target_w.view((self.tensor_shape[0],self.tensor_shape[1],1,-1)), 1,3) - all_disc_w), dim=2)
self.B = torch.transpose((self.bit_table[ind_opt.view(-1),:]).view(self.tensor_shape[0],self.tensor_shape[2]*self.tensor_shape[3],self.tensor_shape[1],self.max_bit), 2,3)
self.B *= self.mask.char()
elif self.structure == 'channelwise':
all_disc_w = torch.matmul(self.bit_table.view((1,1,self.bit_table.size(0),self.bit_table.size(1))).float(),self.alpha)
ind_opt = torch.argmin(torch.abs(target_w.view((self.tensor_shape[0],1,1,-1)) - all_disc_w), dim=2)
self.B = torch.transpose((self.bit_table[ind_opt.view(-1),:]).view(self.tensor_shape[0],1,self.tensor_shape[1]*self.tensor_shape[2]*self.tensor_shape[3],self.max_bit), 2,3)
self.B *= self.mask.char()
return True
def speedup(self, pseudo_grad, pseudo_hessian):
"""Speed up the optimization on binary bases, i.e. take a following optimization step on coordinates while fixing binary bases.
"""
revised_grad_w = -pseudo_hessian*self.w_FP+pseudo_grad
if self.structure == 'kernelwise':
revised_hessian = torch.matmul(self.B.float()*pseudo_hessian.view((self.tensor_shape[0],self.tensor_shape[1],1,-1)),torch.transpose(self.B,2,3).float())
revised_hessian += torch.diag_embed(1+1e-6-(self.mask.float().squeeze(-1)))
revised_grad = torch.matmul(self.B.float(),revised_grad_w.view((self.tensor_shape[0],self.tensor_shape[1],-1,1)))
self.alpha = -torch.matmul(torch.inverse(revised_hessian),revised_grad)
elif self.structure == 'pixelwise':
revised_hessian = torch.matmul(self.B.float()*torch.transpose(pseudo_hessian.view((self.tensor_shape[0],self.tensor_shape[1],1,-1)),1,3),torch.transpose(self.B,2,3).float())
revised_hessian += torch.diag_embed(1+1e-6-(self.mask.float().squeeze(-1)))
revised_grad = torch.matmul(self.B.float(),torch.transpose(revised_grad_w.view((self.tensor_shape[0],self.tensor_shape[1],-1,1)),1,2))
self.alpha = -torch.matmul(torch.inverse(revised_hessian),revised_grad)
elif self.structure == 'channelwise':
revised_hessian = torch.matmul(self.B.float()*pseudo_hessian.view((self.tensor_shape[0],1,1,-1)),torch.transpose(self.B,2,3).float())
revised_hessian += torch.diag_embed(1+1e-6-(self.mask.float().squeeze(-1)))
revised_grad = torch.matmul(self.B.float(),revised_grad_w.view((self.tensor_shape[0],1,-1,1)))
self.alpha = -torch.matmul(torch.inverse(revised_hessian),revised_grad)
self.alpha *= self.mask.float()
ind_neg = self.alpha<0
self.alpha[ind_neg] *= -1
self.B.contiguous().view(-1,self.B.size(-1))[ind_neg.view(-1),:] *= -1
self.num_bin_filter = torch.sum(self.mask)
self.avg_bit = self.num_bin_filter.float()/(self.mask.size(0)*self.mask.size(1))
return True
class FCLayer_bin(object):
"""This class defines the multi-bit form of the weight tensor of a convolutional layer used in ALQ.
Arguments:
w_ori (float tensor): the 4-dim pretrained weight tensor of a convolutional layer.
ind_layer (int): the index of this layer in the network.
structure (string): the structure used for grouping the weights in this layer, optional values: 'subchannelwise'.
max_bit (int): the maximum bitwidth used in initialization.
"""
def __init__(self, w_ori, ind_layer, structure, num_subchannel, max_bit):
# The layer type
self.layer_type = 'fc'
# The shape of the weight tensor of this layer
self.tensor_shape = w_ori.size()
# The maintained full precision weight tensor of this layer used in STE
self.w_FP = w_ori.clone().to('cuda')
# The index of this layer in the network
self.ind_layer = ind_layer
# The structure used for grouping the weights in this layer
self.structure = structure
# The maximum bitwidth used in initialization
self.max_bit = max_bit
# The number of groups in each channel, i.e. the number of subchannels
self.num_subchannel = num_subchannel
# The number of weights in each subchannel
self.num_w_subc = int(self.tensor_shape[1]/self.num_subchannel)
# The binary bases, the coordinates, and the mask (only for parallel computing purposes) of each group
self.B, self.alpha, self.mask = self.structured_sketch()
# The total number of binary filters in this layer, namely the total number of (valid) alpha's
self.num_bin_filter = torch.sum(self.mask)
# The average bitwidth of this layer
self.avg_bit = self.num_bin_filter.float()/(self.mask.size(0)*self.mask.size(1))
# The total number of weights of this layer
self.num_weight = self.w_FP.nelement()
# The used look-up-table for bitwise values
self.bit_table = construct_bit_table(self.max_bit)
def structured_sketch(self):
"""Initialize the weight tensor using structured sketching.
Namely, structure the weights in groupwise, and quantize each group's weights in multi-bit form w.r.t. the reconstruction error.
Return the binary bases, the coordinates, and the mask (only for parallel computing purposes) of each group.
"""
w_cpu = self.w_FP.to('cpu')
B = torch.zeros((self.tensor_shape[0],self.num_subchannel,self.max_bit,self.num_w_subc)).to(torch.int8)
alpha = torch.zeros((self.tensor_shape[0],self.num_subchannel,self.max_bit,1)).to(torch.float32)
mask = torch.zeros((self.tensor_shape[0],self.num_subchannel,self.max_bit,1)).to(torch.bool)
for k in range(self.tensor_shape[0]):
for (q,i) in enumerate(range(0,self.tensor_shape[1],self.num_w_subc)):
bin_basis, crd, dim = transform_bin_basis(w_cpu[k,i:i+self.num_w_subc].view(-1), self.max_bit)
mask[k,q,:dim,0] = 1
B[k,q,:dim,:] = torch.t(bin_basis)
alpha[k,q,:dim,0] = crd
return B.to('cuda'), alpha.to('cuda'), mask.to('cuda')
def reconstruct_w(self):
"""Reconstruct the weight tensor from the current quantization.
Return the reconstructed weight tensor of this layer, i.e. \hat{w}.
"""
w_bin = torch.sum(self.B.float()*(self.alpha*self.mask.float()),dim=2)
return w_bin.reshape(self.tensor_shape)
def update_w_FP(self, w_FP_new=None):
"""Update the full precision weight tensor.
In STE with loss-aware optimization, w_FP is the maintained full precision weight tensor.
In ALQ optimization, w_FP is used to store the reconstructed weight tensor from the current quantization.
"""
if w_FP_new is not None:
self.w_FP.add_(w_FP_new)
else:
self.w_FP.zero_().add_(self.reconstruct_w())
def construct_grad_alpha(self, grad_w):
"""Compute and return the gradient (or the first momentum) in alpha domain w.r.t the loss.
"""
return torch.matmul(self.B.float(), grad_w.reshape((self.tensor_shape[0],self.num_subchannel,self.num_w_subc,1)))*self.mask.float()
def construct_hessian_alpha(self, diag_hessian_w):
"""Compute and return the diagonal Hessian (or the second momentum) in alpha domain w.r.t the loss.
"""
diag_hessian_alpha = torch.matmul(self.B.float()*diag_hessian_w.reshape((self.tensor_shape[0],self.num_subchannel,1,self.num_w_subc)), torch.transpose(self.B,2,3).float())
return torch.diagonal(diag_hessian_alpha,dim1=-2,dim2=-1).unsqueeze(-1)*self.mask.float()
def sort_importance_bin_filter(self, grad_alpha, diag_hessian_alpha, num_top):
"""Compute and sort the importance of binary filters (alpha's) in this layer.
The importance is defined by the modeled loss increment caused by pruning each individual alpha.
Return the selected num_top alpha's with the least importance.
"""
delta_loss_prune = -grad_alpha*self.alpha+0.5*torch.pow(self.alpha,2)*diag_hessian_alpha
sorted_ind = torch.argsort(delta_loss_prune[self.mask].view(-1))
top_importance_list = torch.tensor([[self.ind_layer, sorted_ind[i], delta_loss_prune.view(-1)[sorted_ind[i]]] for i in range(num_top)])
return top_importance_list
def prune_alpha(self, ind_prune):
"""Prune the cooresponding alpha's of this layer give the indexes.
"""
num_bin_filter_ = torch.sum(self.mask)
self.mask.view(-1)[self.mask.view(-1).nonzero().view(-1)[ind_prune]]=0
self.B *= self.mask.char()
self.alpha *= self.mask.float()
self.num_bin_filter = torch.sum(self.mask)
self.avg_bit = self.num_bin_filter.float()/(self.mask.size(0)*self.mask.size(1))
if num_bin_filter_-self.num_bin_filter != ind_prune.size(0):
print('wrong pruning')
return False
return True
def optimize_bin_basis(self, pseudo_grad, pseudo_hessian):
"""Take one optimization step on the binary bases of this layer while fixing coordinates.
"""
# Compute the target weight tensor, i.e. the optimal point in w domain according to the quadratic model function
target_w = self.w_FP-pseudo_grad/pseudo_hessian
all_disc_w = torch.matmul(self.bit_table.view((1,1,self.bit_table.size(0),self.bit_table.size(1))).float(),self.alpha)
ind_opt = torch.argmin(torch.abs(target_w.view((self.tensor_shape[0],self.num_subchannel,1,-1)) - all_disc_w), dim=2)
self.B = torch.transpose((self.bit_table[ind_opt[:],:]).view(self.tensor_shape[0],self.num_subchannel,self.num_w_subc,self.max_bit), 2,3)
self.B *= self.mask.char()
return True
def speedup(self, pseudo_grad, pseudo_hessian):
"""Speed up the optimization on binary bases, i.e. take a following optimization step on coordinates while fixing binary bases.
"""
revised_grad_w = -pseudo_hessian*self.w_FP+pseudo_grad
revised_hessian = torch.matmul(self.B.float()*pseudo_hessian.view((self.tensor_shape[0],self.num_subchannel,1,-1)),torch.transpose(self.B,2,3).float())
revised_hessian += torch.diag_embed(1+1e-6-(self.mask.float().squeeze(-1)))
revised_grad = torch.matmul(self.B.float(),revised_grad_w.view((self.tensor_shape[0],self.num_subchannel,-1,1)))
self.alpha = -torch.matmul(torch.inverse(revised_hessian),revised_grad)
self.alpha *= self.mask.float()
ind_neg = self.alpha<0
self.alpha[ind_neg] *= -1
self.B.contiguous().view(-1,self.B.size(-1))[ind_neg.view(-1),:] *= -1
self.num_bin_filter = torch.sum(self.mask)
self.avg_bit = self.num_bin_filter.float()/(self.mask.size(0)*self.mask.size(1))
return True