-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcreate_pruned_net.py
194 lines (177 loc) · 8.47 KB
/
create_pruned_net.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
################################################################################################
# Create pruned/sparse net used for later finetuning/training
################################################################################################
import os
import random
import argparse
import cv2
import numpy as np
import warnings
import copy
import torch
from torch.utils.data import DataLoader
from torchvision import models
import torch.nn as nn
from torch.autograd import Variable
import matplotlib.pyplot as plt
from dataloaders import *
from scene_net import *
from prune_utils import *
from loss import SceneNetLoss, DiSparse_SceneNetLoss
import torch.nn.utils.prune as prune
# **************************************************************************************************************** #
def create_disparse_static_nyuv2(net, ratio, criterion, train_loader, num_batches, device, tasks):
if ratio == 90:
keep_ratio = 0.08
elif ratio == 70:
keep_ratio = 0.257
elif ratio == 50:
keep_ratio = 0.46
elif ratio == 30:
keep_ratio = 0.675
else:
keep_ratio = (100 - ratio) / 100
net = disparse_prune_static(net, criterion, train_loader, num_batches, keep_ratio, device, tasks)
return net
def create_disparse_static_cityscapes(net, ratio, criterion, train_loader, num_batches, device, tasks):
if ratio == 90:
keep_ratio = 0.095
elif ratio == 70:
keep_ratio = 0.3
elif ratio == 50:
keep_ratio = 0.51
elif ratio == 30:
keep_ratio = 0.71
else:
keep_ratio = (100 - ratio) / 100
net = disparse_prune_static(net, criterion, train_loader, num_batches, keep_ratio, device, tasks)
return net
def create_disparse_static_taskonomy(net, ratio, criterion, train_loader, num_batches, device, tasks):
if ratio == 90:
keep_ratio = 0.097
elif ratio == 70:
keep_ratio = 0.257
elif ratio == 50:
keep_ratio = 0.46
elif ratio == 30:
keep_ratio = 0.675
else:
keep_ratio = (100 - ratio) / 100
net = disparse_prune_static(net, criterion, train_loader, num_batches, keep_ratio, device, tasks)
return net
def create_disparse_pt_nyuv2(net, ratio, criterion, train_loader, num_batches, device, tasks, dest="/data"):
if ratio == 90:
keep_ratio = 0.1
elif ratio == 70:
keep_ratio = 0.3
elif ratio == 50:
keep_ratio = 0.5
elif ratio == 30:
keep_ratio = 0.7
else:
keep_ratio = (100 - ratio) / 100
net.load_state_dict(torch.load(f"/home/nravi/DiSparse-Multitask-Model-Compression/results_new/5000th_nyuv2_baseline.pth"))
# net.load_state_dict(torch.load(f"{dest}/final_seg_sn.pth"))
net = disparse_prune_pretrained(net, criterion, train_loader, num_batches, keep_ratio, device, tasks)
return net
def create_disparse_pt_cityscapes(net, ratio, criterion, train_loader, num_batches, device, tasks, dest="/data"):
if ratio == 90:
keep_ratio = 0.13
elif ratio == 70:
keep_ratio = 0.37
elif ratio == 50:
keep_ratio = 0.6
elif ratio == 30:
keep_ratio = 0.644
else:
keep_ratio = (100 - ratio) / 100
net.load_state_dict(torch.load(f"/data/alexsun/save_model/cityscapes/final_seg_sn.pth"))
# net.load_state_dict(torch.load(f"{dest}/final_seg_depth.pth"))
net = disparse_prune_pretrained_l1(net, criterion, train_loader, num_batches, keep_ratio, device, tasks)
return net
def create_disparse_pt_taskonomy(net, ratio, criterion, train_loader, num_batches, device, tasks, dest="/data"):
if ratio == 90:
keep_ratio = 0.2
elif ratio == 70:
keep_ratio = 0.3
elif ratio == 50:
keep_ratio = 0.5
elif ratio == 30:
keep_ratio = 0.7
else:
keep_ratio = (100 - ratio) / 100
net = torch.nn.DataParallel(net)
net.load_state_dict(torch.load(f"/data/alexsun/save_model/final_taskonomy_5task.pth"))
# net.load_state_dict(torch.load(f"{dest}/final_taskonomy_5task.pth"))
net = net.module
net = disparse_prune_pretrained(net, criterion, train_loader, num_batches, keep_ratio, device, tasks)
return net
# **************************************************************************************************************** #
################################################################################################
if __name__ == "__main__":
warnings.filterwarnings('ignore')
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, help='dataset: choose between nyuv2, cityscapes, taskonomy', default="nyuv2")
parser.add_argument('--num_batches',type=int, help='number of batches to estimate importance', default=50)
parser.add_argument('--method', type=str, help='method name', default="disparse_static")
parser.add_argument('--ratio',type=int, help='percentage of sparsity level', default=90)
parser.add_argument('--dest', default='/data/alexsun/save_model/release_test/', type=str, help='Destination Save Folder.')
args = parser.parse_args()
dataset = args.dataset
ratio = args.ratio
num_batches = args.num_batches
method = args.method
dest = args.dest
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
################################################################################################
if dataset == "nyuv2":
from config_nyuv2 import *
train_dataset = NYU_v2(DATA_ROOT, 'train', crop_h=CROP_H, crop_w=CROP_W)
train_loader = DataLoader(train_dataset, batch_size = BATCH_SIZE, num_workers = 8, shuffle=True, pin_memory=True)
test_dataset = NYU_v2(DATA_ROOT, 'test')
test_loader = DataLoader(test_dataset, batch_size = 1, num_workers = 8, shuffle=True, pin_memory=True)
elif dataset == "cityscapes":
from config_cityscapes import *
train_dataset = CityScapes(DATA_ROOT, 'train', crop_h=CROP_H, crop_w=CROP_W)
train_loader = DataLoader(train_dataset, batch_size = BATCH_SIZE, num_workers = 8, shuffle=True, pin_memory=True)
test_dataset = CityScapes(DATA_ROOT, 'test')
test_loader = DataLoader(test_dataset, batch_size = 1, num_workers = 8, shuffle=True, pin_memory=True)
elif dataset == "taskonomy":
from config_taskonomy import *
train_dataset = Taskonomy(DATA_ROOT, 'train', crop_h=CROP_H, crop_w=CROP_W)
train_loader = DataLoader(train_dataset, batch_size = BATCH_SIZE//4, num_workers = 8, shuffle=True, pin_memory=True)
test_dataset = Taskonomy(DATA_ROOT, 'test')
test_loader = DataLoader(test_dataset, batch_size = 1, num_workers = 8, shuffle=True, pin_memory=True)
else:
print("Unrecognized Dataset Name.")
exit()
################################################################################################
network_name = f"{dataset}_{method}_{ratio}"
save_path = f"{dest}/{network_name}.pth"
################################################################################################
net = SceneNet(TASKS_NUM_CLASS).to(device)
if method == "disparse_static":
criterion = DiSparse_SceneNetLoss(dataset, TASKS, TASKS_NUM_CLASS, LAMBDAS, device, DATA_ROOT)
if dataset == "nyuv2":
net = create_disparse_static_nyuv2(net, ratio, criterion, train_loader, num_batches, device, tasks=TASKS)
elif dataset == "cityscapes":
net = create_disparse_static_cityscapes(net, ratio, criterion, train_loader, num_batches, device, tasks=TASKS)
elif dataset == "taskonomy":
net = create_disparse_static_taskonomy(net, ratio, criterion, train_loader, num_batches, device, tasks=TASKS)
else:
print("Unrecognized Dataset Name.")
exit()
elif method == "disparse_pt":
criterion = DiSparse_SceneNetLoss(dataset, TASKS, TASKS_NUM_CLASS, LAMBDAS, device, DATA_ROOT)
if dataset == "nyuv2":
net = create_disparse_pt_nyuv2(net, ratio, criterion, train_loader, num_batches, device, tasks=TASKS, dest=dest)
elif dataset == "cityscapes":
net = create_disparse_pt_cityscapes(net, ratio, criterion, train_loader, num_batches, device, tasks=TASKS, dest=dest)
elif dataset == "taskonomy":
net = create_disparse_pt_taskonomy(net, ratio, criterion, train_loader, num_batches, device, tasks=TASKS, dest=dest)
else:
print("Unrecognized Dataset Name.")
exit()
print_sparsity(net)
print(f"Saving the pruned model to {save_path}")
torch.save(net.state_dict(), save_path)