forked from NoamRosenberg/autodeeplab
-
Notifications
You must be signed in to change notification settings - Fork 1
/
decoding_formulas.py
297 lines (259 loc) · 13.6 KB
/
decoding_formulas.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
import numpy as np
import torch
import torch.nn.functional as F
from genotypes import PRIMITIVES
from genotypes import Genotype
class Decoder(object):
def __init__(self, alphas, bottom_betas, betas8, betas16, top_betas, block_multiplier, steps):
normalized_bottom_betas = F.softmax(bottom_betas, dim=-1)
normalized_betas8 = F.softmax (betas8, dim = -1)
normalized_betas16 = F.softmax(betas16, dim=-1)
normalized_top_betas = F.softmax(top_betas, dim=-1)
self.network_space = torch.zeros(12,4,3)
self.network_space[0, 0, 1] = 1.
self.network_space[0, 1, 0] = 1.
self.network_space[1:, 0, 1:] = normalized_bottom_betas
self.network_space[1, 1, :2] = normalized_top_betas[0]
self.network_space[1, 2, 0] = 1.
self.network_space[2:, 1, :] = normalized_betas8
self.network_space[2, 2, :2] = normalized_top_betas[1]
self.network_space[2, 3, 0] = 1.
self.network_space[3:, 2, :] = normalized_betas16
self.network_space[3:, 3, :2] = normalized_top_betas[2:]
self.alphas = alphas
self.block_multiplier = block_multiplier
self.steps = steps
def viterbi_decode(self):
#TODO: consider if to set single path probabilities to a third or half some other value other
#TODO: than one in order not to bias the path decodicings
#remember best paths
paths_space = torch.zeros(self.network_space.shape)
paths_space[0, 0, 1] = 1.
paths_space[0, 1, 0] = 1.
#update total transition probability
prob_space = torch.zeros(self.network_space.shape[:2])
prob_space[0, 0] = 1.
prob_space[0, 1] = 1.
for layer in range(1, paths_space.shape[0]):
for sample in range(paths_space.shape[1]):
new_probs_ls = []
for i in range(0,3):
prev_sample = sample + (i - 1)
if prev_sample < 0 or prev_sample > 3:
new_probs_ls.append(torch.tensor(0.))
continue
prev_prob = prob_space[layer - 1, prev_sample]
curr_prob = self.network_space[layer, sample, i]
new_prob = prev_prob * curr_prob
new_probs_ls.append(new_prob)
new_probs_tls = torch.tensor(new_probs_ls)
if new_probs_tls[new_probs_tls > 0].shape[0] == 0:
continue
prob_space[layer, sample] = new_probs_tls.max()
best_index = new_probs_tls.argmax()
paths_space[layer, sample, best_index] = 1.
#pick the branch with the highest probability
all_samples = torch.tensor([i for i in range(prob_space.shape[-1])])
best_sample = torch.argmax(prob_space[-1])
other_samples = all_samples[all_samples != best_sample]
paths_space[-1, other_samples] = 0.
i = paths_space[-1, best_sample].nonzero()[0, 0]
prev_sample = best_sample + (i - 1)
#follow the branch and eliminate all other branches
for layer in range(prob_space.shape[0] - 2, -1, -1):
other_samples = all_samples[all_samples != prev_sample]
paths_space[layer, other_samples] = 0.
i = paths_space[layer, prev_sample].nonzero()[0, 0]
prev_sample = prev_sample + (i - 1)
actual_path = paths_space.nonzero()[:,1]
return actual_path, paths_space
def dfs_decode (self) :
best_result = []
max_prop = 0
def _parse (weight_network, layer, curr_value, curr_result, last) :
nonlocal best_result
nonlocal max_prop
if layer == self._num_layers :
if max_prop < curr_value :
# print (curr_result)
best_result = curr_result[:]
max_prop = curr_value
return
if layer == 0 :
print ('begin0')
num = 0
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
print ('end0-1')
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
elif layer == 1 :
print ('begin1')
num = 0
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
print ('end1-1')
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
num = 1
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][2]
curr_result.append ([num,2])
_parse (weight_network, layer + 1, curr_value, curr_result, 2)
curr_value = curr_value / weight_network[layer][num][2]
curr_result.pop ()
elif layer == 2 :
print ('begin2')
num = 0
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
print ('end2-1')
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
num = 1
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][2]
curr_result.append ([num,2])
_parse (weight_network, layer + 1, curr_value, curr_result, 2)
curr_value = curr_value / weight_network[layer][num][2]
curr_result.pop ()
num = 2
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 2)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][2]
curr_result.append ([num,2])
_parse (weight_network, layer + 1, curr_value, curr_result, 3)
curr_value = curr_value / weight_network[layer][num][2]
curr_result.pop ()
else :
num = 0
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
num = 1
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 0)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][2]
curr_result.append ([num,2])
_parse (weight_network, layer + 1, curr_value, curr_result, 2)
curr_value = curr_value / weight_network[layer][num][2]
curr_result.pop ()
num = 2
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 1)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 2)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][2]
curr_result.append ([num,2])
_parse (weight_network, layer + 1, curr_value, curr_result, 3)
curr_value = curr_value / weight_network[layer][num][2]
curr_result.pop ()
num = 3
if last == num :
curr_value = curr_value * weight_network[layer][num][0]
curr_result.append ([num,0])
_parse (weight_network, layer + 1, curr_value, curr_result, 2)
curr_value = curr_value / weight_network[layer][num][0]
curr_result.pop ()
curr_value = curr_value * weight_network[layer][num][1]
curr_result.append ([num,1])
_parse (weight_network, layer + 1, curr_value, curr_result, 3)
curr_value = curr_value / weight_network[layer][num][1]
curr_result.pop ()
network_weight = F.softmax(self.last_betas_network, dim=-1) * 5
network_weight = network_weight.data.cpu().numpy()
_parse (network_weight, 0, 1, [],0)
print (max_prop)
return best_result
def genotype_decode(self):
def _parse(alphas, steps):
gene = []
start = 0
n = 2
for i in range(steps):
end = start + n
edges = sorted(range(start, end), key=lambda x: -np.max(alphas[x,1:])) #ignore none value
top2edges = edges[:2]
for j in top2edges:
best_op_index = np.argmax(alphas[j]) #this can include none op
gene.append([j, best_op_index])
start = end
n += 1
return np.array(gene)
normalized_alphas = F.softmax(self.alphas, dim=-1).data.cpu().numpy()
gene_cell = _parse(normalized_alphas, self.steps)
return gene_cell