-
Notifications
You must be signed in to change notification settings - Fork 1
/
plot_metric_distribution.py
270 lines (229 loc) · 26 KB
/
plot_metric_distribution.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
import os
import sys
import time
import argparse
import matplotlib.pyplot as pyplot
import random
from easydict import EasyDict as edict
import numpy as np
import tqdm
import torch
from torch import nn
from pathlib import Path
from typing import List, Tuple, Callable
from dataclasses import dataclass
lib_dir = (Path(__file__).parent / 'lib').resolve()
if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir))
from datasets import get_datasets, get_nas_search_loaders
from procedures import prepare_seed, MetricType, get_linear_region_counter_v2, get_ntk_n, get_ntk_n_v2, \
get_nngp_n, get_nngp_n_v2, regional_division_counter, RegionDivisionScoreType, synflow, logsynflow, zen_score, \
grasp, snip, fisher
from models import get_cell_based_tiny_net, get_search_spaces
from nas_201_api import SimpleApi as API
from prune_tenas import init_model
@dataclass
class MethodDescriptor(object):
name: str
codename: str
func: Callable
METHODS_LIST = [
MethodDescriptor("Conditional number NTK", "cond_ntk_v1", lambda loader, _, net: get_ntk_n(loader, [net], recalbn=0, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Accuracy NNGP", "acc_nngp_v1", lambda train_loader, valid_loader, net: get_nngp_n(train_loader, valid_loader, [net], train_mode=True, num_batch=2)[0]),
MethodDescriptor("Accuracy of NTK", "acc_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, num_batch=1)[0]),
MethodDescriptor("MSE of NTK", "mse_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Label-Gradient Alignment of NTK", "lga_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Frobenius norm of NTK", "fro_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Mean value of NTK", "mean_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Conditional number of NTK", "cond_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Eigenvalue score of NTK", "eig_ntk", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, num_batch=1)[0]),
MethodDescriptor("Accuracy of NTK(correlation)", "acc_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("MSE of NTK(correlation)", "mse_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("Label-Gradient Alignment of NTK(correlation)", "lga_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("Frobenius norm of NTK(correlation)", "fro_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("Mean value of NTK(correlation)", "mean_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("Conditional number of NTK(correlation)", "cond_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("Eigenvalue score of NTK(correlation)", "eig_ntk_corr", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, as_correlation=True, num_batch=1)[0]),
MethodDescriptor("Accuracy of NNGP", "acc_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, num_batch=2)[0]),
MethodDescriptor("MSE of NNGP", "mse_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, num_batch=2)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP", "lga_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, num_batch=2)[0]),
MethodDescriptor("Frobenius norm of NNGP", "fro_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, num_batch=2)[0]),
MethodDescriptor("Mean value of NNGP", "mean_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, num_batch=2)[0]),
MethodDescriptor("Conditional number of NNGP", "cond_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, num_batch=2)[0]),
MethodDescriptor("Eigenvalue score of NNGP", "eig_nngp", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, num_batch=2)[0]),
MethodDescriptor("Accuracy of NNGP(correlation)", "acc_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("MSE of NNGP(correlation)", "mse_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP(correlation)", "lga_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("Frobenius norm of NNGP(correlation)", "fro_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("Mean value of NNGP(correlation)", "mean_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("Conditional number of NNGP(correlation)", "cond_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("Eigenvalue score of NNGP(correlation)", "eig_nngp_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, as_correlation=True, num_batch=2)[0]),
MethodDescriptor("Accuracy of NNGP(readout)", "acc_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("MSE of NNGP(readout)", "mse_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP(readout)", "lga_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Frobenius norm of NNGP(readout)", "fro_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Mean value of NNGP(readout)", "mean_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Condition number of NNGP(readout)", "cond_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Eigenvalue score of NNGP(readout)", "eig_nngp_read", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, as_correlation=False, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Accuracy of NNGP(readout, correlation)", "acc_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("MSE of NNGP(readout, correlation)", "mse_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP(readout, correlation)", "lga_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Frobenius norm of NNGP(readout, correlation)", "fro_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Mean value of NNGP(readout, correlation)", "mean_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Conditional number of NNGP(readout, correlation)", "cond_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Eigenvalue score of NNGP(readout, correlation)", "eig_nngp_read_corr", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, as_correlation=True, use_logits=True, num_batch=2)[0]),
MethodDescriptor("Accuracy of NTK(1 train epoch)", "acc_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("MSE of NTK(1 train epoch)", "mse_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("Label-Gradient Alignment of NTK(1 train epoch)", "lga_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("Frobenius norm of NTK(1 train epoch)", "fro_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("Mean value of NTK(1 train epoch)", "mean_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("Conditional number of NTK(1 train epoch)", "cond_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("Eigenvalue score of NTK(1 train epoch)", "eig_ntk_train", lambda train_loader, valid_loader, net: get_ntk_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, train_iters=1, num_batch=1)[0]),
MethodDescriptor("Accuracy of NNGP(1 train epoch)", "acc_nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("MSE of NNGP(1 train epoch)", "mse_nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP(1 train epoch)", "lga_nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("Frobenius norm of NNGP(1 train epoch)", "fro_nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("Mean value of NNGP(1 iterations epoch)", "nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("Conditional number of NNGP(1 train epoch)", "cond_nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("Eigenvalue score of NNGP(1 train epoch)", "eig_nngp_train", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, train_iters=1, num_batch=2)[0]),
MethodDescriptor("ReLU regions distance", "regs_dist_full", lambda train_loader, _, net: regional_division_counter(train_loader, [net], train_mode=False, score_type=RegionDivisionScoreType.FULL, num_batch=1, verbose=False)[0]),
MethodDescriptor("ReLU regions distance(max over batch)", "regs_dist_max", lambda train_loader, _, net: regional_division_counter(train_loader, [net], train_mode=False, score_type=RegionDivisionScoreType.MAX, num_batch=3, verbose=False)[0]),
MethodDescriptor("ReLU regions distance(mean over batch)", "regs_dist_mean", lambda train_loader, _, net: regional_division_counter(train_loader, [net], train_mode=False, score_type=RegionDivisionScoreType.MEAN, num_batch=3, verbose=False)[0]),
MethodDescriptor("ReLU regions distance(1 train epoch)", "regs_dist_full_train", lambda train_loader, _, net: regional_division_counter(train_loader, [net], train_mode=False, score_type=RegionDivisionScoreType.FULL, train_iters=1, num_batch=10, verbose=False)[0]),
MethodDescriptor("ReLU regions distance(max over batch, 1 train epoch)", "regs_dist_max_train", lambda train_loader, _, net: regional_division_counter(train_loader, [net], train_mode=False, score_type=RegionDivisionScoreType.MAX, train_iters=1, num_batch=10, verbose=False)[0]),
MethodDescriptor("ReLU regions distance(mean over batch, 1 train epoch)", "regs_dist_mean_train", lambda train_loader, _, net: regional_division_counter(train_loader, [net], train_mode=False, score_type=RegionDivisionScoreType.MEAN, train_iters=1, num_batch=10, verbose=False)[0]),
MethodDescriptor("Accuracy of NNGP(more batches)", "acc_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, num_batch=40)[0]),
MethodDescriptor("MSE of NNGP(more batches)", "mse_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, num_batch=40)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP(more batches)", "lga_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, num_batch=40)[0]),
MethodDescriptor("Frobenius norm of NNGP(more batches)", "fro_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, num_batch=40)[0]),
MethodDescriptor("Mean value of NNGP(more batches)", "mean_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, num_batch=40)[0]),
MethodDescriptor("Conditional number of NNGP(more batches)", "cond_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, num_batch=40)[0]),
MethodDescriptor("Eigenvalue score of NNGP(more batches)", "eig_nngp_bb", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, num_batch=40)[0]),
MethodDescriptor("Expected number of ReLU regions", "regs_num", lambda train_loader, _, net: get_linear_region_counter_v2(train_loader, [net], train_mode=False, num_batch=3)[0]),
MethodDescriptor("Accuracy of NNGP(20 train iterations )", "acc_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.ACC, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("MSE of NNGP(20 train iterations )", "mse_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MSE, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("Label-Gradient Alignment of NNGP(20 train iterations )", "lga_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.LGA, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("Frobenius norm of NNGP(20 train iterations )", "fro_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.FRO, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("Mean value of NNGP(20 iterations)", "mean_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.MEAN, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("Conditional number of NNGP(20 train iterations )", "cond_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.COND, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("Eigenvalue score of NNGP(20 train iterations )", "eig_nngp_train_it", lambda train_loader, valid_loader, net: get_nngp_n_v2(train_loader, valid_loader, [net], metric=MetricType.EIG, train_mode=True, train_iters=20, num_batch=2)[0]),
MethodDescriptor("SynFlow", "synflow", lambda train_loader, _, net: synflow(train_loader, [net], train_mode=True)[0]),
MethodDescriptor("LogSynFlow", "logsynflow", lambda train_loader, _, net: logsynflow(train_loader, [net], train_mode=True)[0]),
MethodDescriptor("Zen-Score", "zen_score", lambda train_loader, _, net: zen_score(train_loader, [net], train_mode=False)[0]),
MethodDescriptor("Grasp", "grasp", lambda train_loader, _, net: grasp(train_loader, [net], train_mode=True)[0]),
MethodDescriptor("Snip", "snip", lambda train_loader, _, net: snip(train_loader, [net], train_mode=True)[0]),
MethodDescriptor("Fisher", "fisher", lambda train_loader, _, net: fisher(train_loader, [net], train_mode=True)[0]),
]
class RandomArchSampler(object):
INFINITY = -1000
def __init__(self, shapes: List[Tuple[int,...]]):
self.shapes = shapes
def sample(self):
arch_parameters = []
for sh in self.shapes:
alpha = torch.rand(sh)
alpha[:, 0] = RandomArchSampler.INFINITY
alpha = torch.argmax(alpha, dim=1).long()
alpha = nn.functional.one_hot(alpha, sh[1])
arch_parameters.append(alpha)
return arch_parameters
def is_expected_regions_number(method: MethodDescriptor):
if method == "regs_num":
return True
return False
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
prepare_seed(xargs.rand_seed)
os.makedirs(args.save_path, exist_ok=True)
os.makedirs(os.path.join(xargs.save_path, "raw_data"), exist_ok=True)
# choosing test method
for m in METHODS_LIST:
if m.codename == args.method:
method = m
print("testing {} on {}".format(m.name, xargs.dataset))
break
else:
print(f"No such method as {args.method}")
exit(1)
batch_size = args.batch_size
input_size = None
if is_expected_regions_number(args.method):
input_size = (batch_size, 1, 3, 3)
train_data, valid_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, input_size=input_size, cutout=-1)
_, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/dataset_split/', batch_size, xargs.workers)
search_space = get_search_spaces('cell', 'nas-bench-201')
model_config = edict({'name': 'DARTS-V1',
'C': 3, 'N': 1, 'depth': -1, 'use_stem': True,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
})
if is_expected_regions_number(args.method):
model_config = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 1, 'use_stem': False,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
})
network = get_cell_based_tiny_net(model_config).cuda().train()
init_model(network, xargs.init)
sampler = RandomArchSampler([a.shape for a in network.get_alphas()])
api = API(xargs.arch_nas_dataset)
values = []
accuracies = []
genotypes = []
times = []
for _ in tqdm.trange(xargs.num_samples):
alphas = sampler.sample()
network.set_alphas(alphas)
genotype = network.genotype()
start_time = time.time()
value = 0
for _ in range(xargs.repeat):
# random reinit
init_model(network, xargs.init + "_fanout" if xargs.init.startswith('kaiming') else xargs.init)
value += method.func(train_loader, valid_loader, network)
value /= xargs.repeat
times.append((time.time() - start_time) / xargs.repeat)
idx = api.query_index_by_arch(genotype)
acc = api.query_by_index(idx)[xargs.dataset]
values.append(value)
accuracies.append(acc)
genotypes.append(genotype)
time_per_iteration = np.mean(times)
print(f"Time per iteration: {time_per_iteration} sec")
## making plot
fig = pyplot.figure()
pyplot.scatter(values, accuracies, s=5.5, c="b", marker="o")
pyplot.xlabel(method.name, fontsize=15)
pyplot.ylabel(f"Test accuracy ({xargs.dataset.upper()})", fontsize=15)
pyplot.savefig(os.path.join(xargs.save_path, f"{method.codename}_{xargs.dataset}.png"))
pyplot.close(fig)
# saving results
np.savez(os.path.join(xargs.save_path, "raw_data", f"{method.codename}_{xargs.dataset}.npz"),
metric=values,
accuracy=accuracies,
times=times,
genotypes=genotypes)
if __name__ == '__main__':
parser = argparse.ArgumentParser("Metric plotter for NAS-bench-201")
parser.add_argument('--data_path', type=str, help='Path to dataset')
parser.add_argument('--dataset', type=str, choices=['cifar10', 'cifar100', 'ImageNet16-120'], help='Choose between cifar10/100/ImageNet16-120/imagenet-1k')
parser.add_argument('--max_nodes', type=int, help='The maximum number of nodes.')
parser.add_argument('--track_running_stats', type=int, choices=[0, 1], help='Whether use track_running_stats or not in the BN layer.')
parser.add_argument('--workers', type=int, default=0, help='number of data loading workers (default: 0)')
parser.add_argument('--batch_size', type=int, default=16, help='batch size for ntk')
parser.add_argument('--save_path', type=str, help='Folder to save plots')
parser.add_argument('--arch_nas_dataset', type=str, help='The path to load the nas-bench-201 architecture dataset (tiny-nas-benchmark).')
parser.add_argument('--rand_seed', type=int, help='manual seed')
parser.add_argument('--precision', type=int, default=3, help='precision for % of changes of cond(NTK) and #Regions')
parser.add_argument('--repeat', type=int, default=3, help='repeat calculation of NTK and Regions')
parser.add_argument('--init', default='kaiming_norm', choices=['kaiming_norm', 'kaiming_norm_fanin', 'kaiming_norm_fanout'], help='choose init')
parser.add_argument('--num_samples', type=int, default=100, help='number of samples for the graph')
parser.add_argument('--method', type=str, choices=[m.codename for m in METHODS_LIST], default="cond_ntk", help="choose method to test")
args = parser.parse_args()
if args.rand_seed is None or args.rand_seed < 0:
args.rand_seed = random.randint(1, 100000)
main(args)