/
cal_rate.py
219 lines (187 loc) · 8.24 KB
/
cal_rate.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
import numpy as np
import torch
from torchvision import datasets, transforms
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib, math
import tqdm, sys
import pandas as pd
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import random_split
from utils import *
from trainer import *
#%matplotlib inline
# enable cuda devices
import os
def cal_gw(density_list, grids_list, R, ep, all_num, trace):
temp_gw = 0
max_ev_index = density_list.index(max(density_list))
ev_threshold = grids_list[max_ev_index]
for i in range(len(density_list)):
density = density_list[i]
eigenvalue = abs(grids_list[i])
""" if eigenvalue <= 0:
eigenvalue = 1e-30 """
radius_2 = 2*ep/eigenvalue
temp_gw += density*radius_2/(R**2+radius_2)
gw = temp_gw/sum(density_list)
print(gw)
return gw
def get_params_diff(model, s_model):
param_list = []
s_param_list = []
norm2 = 0
for param in model.parameters():
if not param.requires_grad:
continue
param_list.append(param)
for param in s_model.parameters():
if not param.requires_grad:
continue
s_param_list.append(param)
for i in range(len(param_list)):
diff = param_list[i] - s_param_list[i]
norm2 += torch.norm(diff, p=2)**2
return norm2**0.5
def sparse_list(sparse_density_list, max_density_index, sparse_density):
if max(sparse_density_list) > sparse_density:
sparse_density_list[max_density_index] = max(sparse_density_list) - sparse_density
print('big')
done = True
else:
sparse_density_list[max_density_index] = 0
sparse_density = sparse_density - max(sparse_density_list)
max_density_index = sparse_density_list.index(max(sparse_density_list))
done = False
return sparse_density_list, sparse_density, done, max_density_index
def cal_sparse_rate(index, trace, important_ev_num=0):
init_file_path()
train_loader, val_loader, actual_val_loader = get_dataset()
criterion = get_criterion()
parser_args.mode = 'sparse'
eigen_path = parser_args.file_path + '{}/hessien_eigenvalue/'.format(parser_args.mode)
check_dir(eigen_path)
model_path = parser_args.file_path + '{}/model/model_{}.pth'.format(parser_args.mode, 0)
model = torch.load(model_path, map_location='cuda:{}'.format(parser_args.gpu))
all_num, norm2, max_w = get_params_size(model)
model.eval()
ori_loss, loss_std = cal_loss(model, train_loader, criterion)
validate(model, val_loader, criterion)
ss_model_path = parser_args.file_path + '{}/model/model_{}.pth'.format(parser_args.mode, index)
ss_model = torch.load(ss_model_path, map_location='cuda:{}'.format(parser_args.gpu))
ss_model.eval()
ss_ori_loss, ss_loss_std = cal_loss(ss_model, train_loader, criterion)
loss_std = loss_std #ori_loss + loss_std - ss_ori_loss
print(loss_std)
ep = loss_std#.detach().cpu()
temp_gw = 0
R = norm2
density_list = []
density_f = open(parser_args.file_path + 'density.txt',"r")
density_lines = density_f.readlines()
for density_line in density_lines:
density = float(density_line[:-1])
density_list.append(density)
density_f.close()
grids_list = []
grids_f = open(parser_args.file_path + 'grids.txt',"r")
grids_lines = grids_f.readlines()
for grids_line in grids_lines:
grids = float(grids_line[:-1])
grids_list.append(grids)
grids_f.close()
for i in range(len(grids_list)):
d_index = len(grids_list) - i
if density_list[d_index-1] == 0:
stop_index = d_index - 1
break
csv_data = pd.read_csv(parser_args.file_path + 'sparse/acc_loss.csv')
orign_acc = csv_data['acc'][0]
for i in range(parser_args.sparse_epoch):
sparse_acc = csv_data['acc'][i]
if sparse_acc < orign_acc-0.1:
sparse_threshold = i/parser_args.sparse_epoch
print(sparse_threshold)
break
gaussian_zero_num_rate = important_ev_num
print(gaussian_zero_num_rate)
sparse_density_sum = gaussian_zero_num_rate/(parser_args.sparse_epoch-gaussian_zero_num_rate) * sum(density_list)
new_density_list, new_grids_list = [], []
for w in range(len(density_list)):
if density_list[w] > 0:
new_density_list.append(density_list[w])
new_grids_list.append(grids_list[w])
density_list, grids_list = new_density_list, new_grids_list
density_list.append(sparse_density_sum)
grids_list.append(1e-30)
print('calculate gaussian width')
gw_list = []
for i in range(int(norm2)+1):
R = i
gw = cal_gw(density_list, grids_list, R, ep, all_num, trace)
gw_list.append(gw)
model_path = parser_args.file_path + '{}/model/model_{}.pth'.format(parser_args.mode, index)
model = torch.load(model_path, map_location='cuda:{}'.format(parser_args.gpu))
print('calculate R & saprse rate')
R_list, sparse_rate_list, acc_list, sparse_rate_list1, R_list1 = [], [], [], [], []
for i in range(parser_args.sparse_epoch):
print(i)
s_model_path = parser_args.file_path + '{}/model/model_{}.pth'.format(parser_args.mode, i)
s_model = torch.load(s_model_path, map_location='cuda:{}'.format(parser_args.gpu))
sparsity_rate = i/parser_args.sparse_epoch
s_model = get_sparse_model(model, sparsity_rate)
s_all_num, s_norm2, s_max_w = get_params_size(s_model)
#R = (norm2 - s_norm2).cpu().detach()
R = get_params_diff(model, s_model).cpu().detach()
R_list.append(R)
acc_list.append(csv_data['acc'][i])
sparse_rate_list.append(i/parser_args.sparse_epoch)
sparse_rate_pr_gt_diff_list = []
for i in range(len(sparse_rate_list)):
R = int(R_list[i].round())
print(R)
sparse_rate_gt = sparse_rate_list[i]
sparse_rate_pr = gw_list[R]
sparse_rate_pr_gt_diff = abs(sparse_rate_pr - sparse_rate_gt)
sparse_rate_pr_gt_diff_list.append(sparse_rate_pr_gt_diff)
pr_sparse_rate_index = sparse_rate_pr_gt_diff_list.index(min(sparse_rate_pr_gt_diff_list))
sparse_rate = pr_sparse_rate_index/parser_args.sparse_epoch
print(sparse_rate)
sparse_index = pr_sparse_rate_index
acc_loss_dict = {}
acc_loss_dict["R_list"] = R_list
acc_loss_dict["sparse_rate_list"] = sparse_rate_list
acc_loss_dict["gw_list"] = gw_list
acc_loss_dict["sparse_index"] = sparse_index
acc_loss_dict["norm2"] = norm2
np.save(parser_args.file_path + 'acc_loss_dict.npy', acc_loss_dict)
plt.clf()
plt.figure(dpi=400,figsize=(6,4))
plt.plot(R_list, sparse_rate_list, label='Network Actual Value', c='b')
plt.plot(range(0,int(norm2)+1), gw_list, label='Theory Prediction', c='g')
plt.xlabel("R")
plt.ylabel("Pruning Ratio)")
plt.grid()
plt.legend()
plt.savefig(parser_args.file_path + 'predict.jpg', bbox_inches='tight')
plt.close('all')
train_csv_data = pd.read_csv(parser_args.file_path + 'sparse_train/acc_loss.csv')
plt.clf()
plt.figure(dpi=400,figsize=(6,4))
if ((max(csv_data['loss'])/csv_data['loss'][0])>5) or ((max(train_csv_data['loss'])/train_csv_data['loss'][0])) > 5:
plt.plot(sparse_rate_list, [math.log10(item/csv_data['loss'][0])+2 for item in csv_data['loss']], label='test_sparse_loss_log10', c='b')
plt.plot(sparse_rate_list, [math.log10(item/train_csv_data['loss'][0])+1.5 for item in train_csv_data['loss']], label='train_sparse_loss_log10', c='g')
else:
plt.plot(sparse_rate_list, [item/csv_data['loss'][0] for item in csv_data['loss']], label='test_sparse_loss', c='b')
plt.plot(sparse_rate_list, [item/train_csv_data['loss'][0] for item in train_csv_data['loss']], label='train_sparse_loss', c='g')
plt.plot(sparse_rate_list, [item/100 for item in csv_data['acc']], label='test_sparse_acc', c='r')
plt.plot(sparse_rate_list, [item/100 for item in train_csv_data['acc']], label='train_sparse_acc', c='k')
plt.axvline(sparse_index/parser_args.sparse_epoch)
plt.xlabel("Pruning Ratio")
plt.ylabel("Accuracy & Loss")
plt.grid()
plt.legend()
plt.savefig(parser_args.file_path + 'acc_loss.jpg', bbox_inches='tight')
plt.close('all')
return sparse_index