-
Notifications
You must be signed in to change notification settings - Fork 4
/
pdebench.py
212 lines (190 loc) · 5.36 KB
/
pdebench.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
import gc
import logging
import pathlib
import sys
import wandb
import torch
from torch import nn
from torch.utils import data
import data_utils
from layers.attention import TNOBlock3D
from layers.fino import SpectralConvolutionKernel3D
from models.codano import CoDANOTemporal
from neuraloperator.models import FNO
from models.get_models import get_ssl_models_codaNo, SSLWrapper, StageEnum
from train import multi_physics_trainer, test_single_physics
from utils import get_wandb_api_key, save_model
import YParams
from data_utils.visualization import (
show_data_diff,
show_multi_physics_data_diffs,
)
logging.basicConfig(level=logging.DEBUG, stream=sys.stdout, force=True)
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# SSL model
params = YParams.YParams("./config/pdebench.yaml", "codano", print_params=False,)
verbose = True
# Set up WandB logging
if params.wandb["log"]:
wandb.login(key=get_wandb_api_key())
wandb.init(
config=params,
name=params.wandb["name"],
group=params.wandb["group"],
project=params.wandb["project"],
entity=params.wandb["entity"],
)
# +
if verbose:
logger.debug(f"{params.nettype=}")
if params.nettype == "transformer":
if verbose:
logger.info(f"{params.grid_type=}")
if params.grid_type == "uniform":
encoder, decoder, contrastive, predictor = get_ssl_models_codaNo(
params,
CoDANOTemporal,
TNOBlock3D,
SpectralConvolutionKernel3D,
)
if verbose:
logger.info(f"{params.pretrain_ssl=}")
model = SSLWrapper(
params,
encoder,
decoder,
contrastive,
predictor,
stage=("ssl" if params.pretrain_ssl else "sl"),
)
model = model.cuda()
Equation = data_utils.Equation
MultiPhysicsDataset = data_utils.MultiPhysicsDataset
NSIncompressibleDataset = data_utils.NSIncompressibleDataset
# Datasets to be used in reconstructive (i.e. SSL) learning:
train_reconstructive = MultiPhysicsDataset(
swe_args=params.shallow_water,
diff_args=params.diffusion_reaction,
ns_args=params.navier_stokes,
predictive=False, # Target output will be the same as the input.
)
test_reconstructive = MultiPhysicsDataset(
swe_args=params.shallow_water,
diff_args=params.diffusion_reaction,
ns_args=params.navier_stokes,
offset=10,
predictive=False, # Target output will be the same as the input.
)
# Datasets to be used in predictive (i.e. SL) learning:
train_predictive = MultiPhysicsDataset(
swe_args=params.shallow_water,
diff_args=params.diffusion_reaction,
ns_args=params.navier_stokes,
predictive=True, # Target output will be the next time trajectory.
)
test_predictive = MultiPhysicsDataset(
swe_args=params.shallow_water,
diff_args=params.diffusion_reaction,
ns_args=params.navier_stokes,
offset=10,
predictive=True, # Target output will be the next time trajectory.
)
# +
train_reconstructive_loader = data.DataLoader(
train_reconstructive, batch_size=params.batch_size, shuffle=False,
)
test_reconstructive_loader = data.DataLoader(
test_reconstructive, batch_size=params.batch_size, shuffle=False,
)
train_predictive_loader = data.DataLoader(
train_predictive, batch_size=params.batch_size, shuffle=False,
)
test_predictive_loader = data.DataLoader(
test_predictive, batch_size=params.batch_size, shuffle=False,
)
# -
# Train/test for the reconstructive (i.e. SSL) task:
model.train()
model.stage = StageEnum.RECONSTRUCTIVE
# import pdb; pdb.set_trace()
multi_physics_trainer(
model,
train_reconstructive_loader,
test_reconstructive_loader,
nn.MSELoss(),
params,
epochs=params.epochs,
wandb_log=params.wandb["log"],
log_interval=params.wandb["log_interval"],
script=True,
)
gc.collect()
torch.cuda.empty_cache()
show_multi_physics_data_diffs(
model,
train_reconstructive_loader,
swe_index=0,
diff_index=50,
ns_index=100,
stage=StageEnum.RECONSTRUCTIVE,
logger=logger,
)
show_multi_physics_data_diffs(
model,
test_reconstructive_loader,
swe_index=0,
diff_index=50,
ns_index=100,
stage=StageEnum.RECONSTRUCTIVE,
logger=logger,
)
save_model(
model,
directory=pathlib.Path("/weight"),
stage=StageEnum.RECONSTRUCTIVE,
)
model.train()
logger.setLevel(logging.DEBUG)
params["gradient"]["threshold"] = 0.1
print(f"{params.pretrain_ssl=}")
if params.pretrain_ssl:
# Now train/test for the predictive (i.e. SL) task:
model.stage = StageEnum.PREDICTIVE
multi_physics_trainer(
model,
train_predictive_loader,
test_predictive_loader,
nn.MSELoss(), # training loss_fn
params,
epochs=10,
wandb_log=params.wandb["log"],
log_interval=params.wandb["log_interval"],
)
# -
test_single_physics(
model, test_predictive_loader, nn.MSELoss(), start=200, stop=2_000, script=False,
)
show_multi_physics_data_diffs(
model,
train_predictive_loader,
swe_index=0,
diff_index=50,
ns_index=100,
stage=StageEnum.PREDICTIVE,
logger=logger,
)
show_multi_physics_data_diffs(
model,
test_predictive_loader,
swe_index=0,
diff_index=50,
ns_index=100,
stage=StageEnum.PREDICTIVE,
logger=logger,
)
save_model(
model, directory=pathlib.Path("/weight"), stage=StageEnum.PREDICTIVE,
)
if params.wandb["log"]:
wandb.finish()