-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_model.py
247 lines (194 loc) · 11.8 KB
/
run_model.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
import torch
import argparse
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import time
from torch.utils import data
from models import Relaxed_ConvNet, Relaxed_Rot_SteerConvNet, Relaxed_Scale_SteerCNNs
from utils import Dataset, train_epoch, eval_epoch, get_lr
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(torch.cuda.device_count())
parser = argparse.ArgumentParser(description='Approximately Equivariant CNNs')
parser.add_argument('--dataset', type=str, required=False, default="PhiFlow", help='PhiFlow or JetFlow')
parser.add_argument('--hidden_dim', type=int, required=False, default="128", help='hidden dimension')
parser.add_argument('--num_layers', type=int, required=False, default="5", help='number of layers')
parser.add_argument('--out_length', type=int, required=False, default="6", help='number of prediction losses used for backpropagation')
parser.add_argument('--num_banks', type=int, required=False, default="2", help='number of filter banks used in relaxed group convolution')
parser.add_argument('--alpha', type=float, required=False, default="0", help='coefficient of the regularizer')
parser.add_argument('--input_length', type=int, required=False, default="10", help='input length')
parser.add_argument('--batch_size', type=int, required=False, default="8", help='batch size')
parser.add_argument('--num_epoch', type=int, required=False, default="1000", help='maximum number of epochs')
parser.add_argument('--learning_rate', type=float, required=False, default="0.001", help='learning rate')
parser.add_argument('--decay_rate', type=float, required=False, default="0.95", help='learning decay rate')
parser.add_argument('--relaxed_symmetry', type=str, required=True, default="Translation", help='translation or rotation or scaling')
args = parser.parse_args()
hidden_dim = args.hidden_dim
num_layers = args.num_layers
out_length = args.out_length
num_banks = args.num_banks
alpha = args.alpha
input_length = args.input_length
batch_size = args.batch_size
num_epoch = args.num_epoch
learning_rate = args.learning_rate
decay_rate = args.decay_rate
symmetry = args.relaxed_symmetry
mid = input_length + 2
min_rmse = 1e8
# Split time ranges
train_time = list(range(0, 160))
valid_time = list(range(160, 200))
test_future_time = list(range(200, 250))
test_domain_time = list(range(0, 100))
# Split tasks
if args.dataset == "PhiFlow":
data_direc = args.dataset + "/" + symmetry
h_size, w_size = 64, 64
if symmetry == "Translation":
train_task = [(48, 10), (56, 10), (8, 20), (40, 5), (56, 25),
(48, 20), (48, 5), (16, 20), (56, 5), (32, 10),
(56, 15), (16, 5), (40, 15), (40, 25), (48, 25),
(48, 15), (24, 10), (56, 20), (32, 15), (16, 15),
(8, 10), (24, 15), (8, 15), (32, 25), (8, 5)]
test_domain_task = [(32, 20), (32, 5), (24, 20), (16, 25), (24, 5),
(16, 10), (40, 20), (8, 25), (24, 25), (40, 10)]
elif symmetry == "Rotation":
train_task = [(27, 2), (33, 0), (3, 2), (28, 3),(9, 0),
(12, 3), (22, 1), (8, 3), (30, 1), (25, 0),
(16, 3), (11, 2), (23, 2), (29, 0), (36, 3),
(26, 1), (1, 0), (35, 2), (19, 2), (34, 1),
(4, 3), (2, 1), (7, 2), (31, 2), (17, 0)]
test_domain_task = [(6, 1), (14, 1), (15, 2), (10, 1), (18, 1),
(20, 3), (24, 3), (13, 0), (21, 0), (5, 0)]
elif symmetry == "Scale":
train_task = [27, 9, 7, 11, 4, 26, 35,
2, 29, 10, 34, 12, 37, 28,
18, 24, 8, 14, 1, 31, 25,
0, 19, 15, 36, 3, 20, 13]
test_domain_task = [ 5, 30, 16, 23, 33,
6, 17, 22, 21, 32]
elif args.dataset == "JetFlow":
h_size, w_size = 62, 23
data_direc = args.dataset
train_task = [(1, 4), (3, 4), (2, 4), (1, 1), (2, 6), (3, 5),
(3, 3), (3, 1), (1, 8), (3, 8), (3, 6), (2, 1),
(1, 3), (1, 6), (2, 8), (1, 7), (1, 2), (2, 2)]
test_domain_task = [(2, 3), (3, 7), (2, 7), (2, 5), (3, 2), (1, 5)]
else:
print("Invalid dataset name entered!")
valid_task = train_task
test_future_task = train_task
train_set = Dataset(input_length = input_length,
mid = mid,
output_length = out_length,
direc = data_direc,
task_list = train_task,
sample_list = train_time,
stack = True)
valid_set = Dataset(input_length = input_length,
mid = mid,
output_length = out_length,
direc = data_direc,
task_list = valid_task,
sample_list = valid_time,
stack = True)
test_set_future = Dataset(input_length = input_length,
mid = mid,
output_length = 20,
direc = data_direc,
task_list = test_future_task,
sample_list = test_future_time,
stack = True)
test_set_domain = Dataset(input_length = input_length,
mid = mid,
output_length = 20,
direc = data_direc,
task_list = test_domain_task,
sample_list = test_domain_time,
stack = True)
train_loader = data.DataLoader(train_set, batch_size = batch_size, shuffle = True, num_workers = 8)
valid_loader = data.DataLoader(valid_set, batch_size = batch_size, shuffle = True, num_workers = 8)
test_loader_future = data.DataLoader(test_set_future, batch_size = batch_size, shuffle = False, num_workers = 8)
test_loader_domain = data.DataLoader(test_set_domain, batch_size = batch_size, shuffle = False, num_workers = 8)
if symmetry == "Translation":
model = nn.DataParallel(Relaxed_ConvNet(in_channels = input_length*2,
out_channels = 2,
hidden_dim = hidden_dim,
kernel_size = 3,
h_size = h_size,
w_size = w_size,
num_layers = num_layers,
num_banks = num_banks,
alpha = alpha).to(device))
model_name = "Relaxed_ConvNet_bz{}_pred{}_lr{}_decay{}_hid{}_layers{}_banks{}_alpha{}".format(batch_size,
out_length,
learning_rate,
decay_rate,
hidden_dim,
num_layers,
num_banks,
alpha)
elif symmetry == "Rotation":
model = nn.DataParallel(Relaxed_Rot_SteerConvNet(in_frames = input_length,
out_frames = 1,
hidden_dim = hidden_dim,
kernel_size = 3,
num_layers = num_layers,
N = 4,
alpha = alpha).to(device))
model_name = "Relaxed_Rot_SteerConvNet_bz{}_pred{}_lr{}_decay{}_hid{}_layers{}_alpha{}".format(batch_size,
out_length,
learning_rate,
decay_rate,
hidden_dim,
num_layers,
alpha)
elif symmetry == "Scale":
model = nn.DataParallel(Relaxed_Scale_SteerCNNs(in_channels = input_length*2,
out_channels = 2,
hidden_dim = hidden_dim,
kernel_size = 5,
num_layers = num_layers,
scales= [1.0,1.5,2.0,2.5],
basis_type='A',
alpha = alpha).to(device))
model_name = "Relaxed_Scale_SteerCNNs_bz{}_pred{}_lr{}_decay{}_hid{}_layers{}_alpha{}".format(batch_size,
out_length,
learning_rate,
decay_rate,
hidden_dim,
num_layers,
alpha)
print(model_name)
print("number of paramters:", sum(p.numel() for p in model.parameters() if p.requires_grad)/1e6)
optimizer = torch.optim.Adam(model.parameters(), learning_rate,betas=(0.9, 0.999), weight_decay=4e-4)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size= 1, gamma=decay_rate)
loss_fun = torch.nn.MSELoss()
########################################################################################################################################################
train_rmse = []
valid_rmse = []
for i in range(num_epoch):
start = time.time()
model.train()
train_rmse.append(train_epoch(train_loader, model, optimizer, loss_fun))
model.eval()
mse, preds, trues = eval_epoch(valid_loader, model, loss_fun)
valid_rmse.append(mse)
if valid_rmse[-1] < min_rmse:
min_rmse = valid_rmse[-1]
best_model = model
end = time.time()
# Early Stopping
if (len(train_rmse) > 100 and np.mean(valid_rmse[-5:]) >= np.mean(valid_rmse[-10:-5])):
break
scheduler.step()
print("Epoch {} | T: {:0.2f} | Train RMSE: {:0.3f} | Valid RMSE: {:0.3f}".format(i+1, (end-start)/60, train_rmse[-1], valid_rmse[-1]))
future_rmse, future_preds, future_trues = eval_epoch(test_loader_future, best_model, loss_fun)
domain_rmse, domain_preds, domain_trues = eval_epoch(test_loader_domain, best_model, loss_fun)
print("Future Test RMSE:", future_rmse, ";", "Domain Test RMSE:", domain_rmse)
torch.save({"test_future": [future_rmse, future_preds, future_trues],
"test_domain": [domain_rmse, domain_preds, domain_trues],
"model": best_model.state_dict()},
args.dataset + "_" + model_name + ".pt")