-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathhydra_main.py
252 lines (228 loc) · 10.7 KB
/
hydra_main.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
import os
import pytorch_lightning as pl
import torch
from pytorch_lightning.callbacks import ModelCheckpoint
import hydra
import pandas as pd
from datetime import datetime, timedelta
from hydra.utils import get_class, instantiate, call
from omegaconf import OmegaConf
import hydra_config
import numpy as np
def get_profiler():
from pytorch_lightning.profiler import PyTorchProfiler
print( torch.profiler.ProfilerActivity.CPU,)
return PyTorchProfiler(
"results/profile_report",
schedule=torch.profiler.schedule(
skip_first=2,
wait=2,
warmup=2,
active=2),
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
],
# with_stack=True,
on_trace_ready=torch.profiler.tensorboard_trace_handler('./tb_profile'),
record_shapes=True,
profile_memory=True,
)
class FourDVarNetHydraRunner:
def __init__(self, params, dm, lit_mod_cls, callbacks=None, logger=None):
self.cfg = params
self.filename_chkpt = self.cfg.ckpt_name
self.callbacks = callbacks
self.logger = logger
self.dm = dm
self.lit_cls = lit_mod_cls
dm.setup()
self.dataloaders = {
'train': dm.train_dataloader(),
'val': dm.val_dataloader(),
'test': dm.test_dataloader(),
}
test_dates = np.concatenate([ \
[str(dt.date()) for dt in \
pd.date_range(dm.test_slices[i].start,dm.test_slices[i].stop)[(self.cfg.dT//2):-(self.cfg.dT//2)]] \
for i in range(len(dm.test_slices))])
#print(test_dates)
self.time = {'time_test' : test_dates}
self.setup(dm)
def setup(self, datamodule):
self.mean_Tr = datamodule.norm_stats[0]
self.mean_Tt = datamodule.norm_stats[0]
self.mean_Val = datamodule.norm_stats[0]
self.var_Tr = datamodule.norm_stats[1] ** 2
self.var_Tt = datamodule.norm_stats[1] ** 2
self.var_Val = datamodule.norm_stats[1] ** 2
self.min_lon = datamodule.dim_range['lon'].start
self.max_lon = datamodule.dim_range['lon'].stop
self.min_lat = datamodule.dim_range['lat'].start
self.max_lat = datamodule.dim_range['lat'].stop
self.ds_size_time = datamodule.ds_size['time']
self.ds_size_lon = datamodule.ds_size['lon']
self.ds_size_lat = datamodule.ds_size['lat']
self.dX = int((datamodule.slice_win['lon']-datamodule.strides['lon'])/2)
self.dY = int((datamodule.slice_win['lat']-datamodule.strides['lat'])/2)
self.swX = datamodule.slice_win['lon']
self.swY = datamodule.slice_win['lat']
self.lon, self.lat = datamodule.coordXY()
w_ = np.zeros(self.cfg.dT)
w_[int(self.cfg.dT / 2)] = 1.
self.wLoss = torch.Tensor(w_)
self.resolution = datamodule.resolution
self.original_coords = datamodule.get_original_coords()
self.padded_coords = datamodule.get_padded_coords()
def run(self, ckpt_path=None, dataloader="test", **trainer_kwargs):
"""
Train and test model and run the test suite
:param ckpt_path: (Optional) Checkpoint from which to resume
:param dataloader: Dataloader on which to run the test Checkpoint from which to resume
:param trainer_kwargs: (Optional)
"""
mod, trainer = self.train(ckpt_path, **trainer_kwargs)
self.test(dataloader=dataloader, _mod=mod, _trainer=trainer)
def _get_model(self, ckpt_path=None):
"""
Load model from ckpt_path or instantiate new model
:param ckpt_path: (Optional) Checkpoint path to load
:return: lightning module
"""
print('get_model: ', ckpt_path)
if ckpt_path:
mod = self.lit_cls.load_from_checkpoint(ckpt_path,
hparam=self.cfg,
w_loss=self.wLoss,
strict=False,
mean_Tr=self.mean_Tr,
mean_Tt=self.mean_Tt,
mean_Val=self.mean_Val,
var_Tr=self.var_Tr,
var_Tt=self.var_Tt,
var_Val=self.var_Val,
min_lon=self.min_lon, max_lon=self.max_lon,
min_lat=self.min_lat, max_lat=self.max_lat,
ds_size_time=self.ds_size_time,
ds_size_lon=self.ds_size_lon,
ds_size_lat=self.ds_size_lat,
time=self.time,
dX=self.dX, dY = self.dY,
swX=self.swX, swY=self.swY,
coord_ext={'lon_ext': self.lon,
'lat_ext': self.lat},
test_domain=self.cfg.test_domain,
resolution=self.resolution,
original_coords=self.original_coords,
padded_coords=self.padded_coords
)
else:
mod = self.lit_cls(hparam=self.cfg,
w_loss=self.wLoss,
mean_Tr=self.mean_Tr,
mean_Tt=self.mean_Tt,
mean_Val=self.mean_Val,
var_Tr=self.var_Tr,
var_Tt=self.var_Tt,
var_Val=self.var_Val,
min_lon=self.min_lon, max_lon=self.max_lon,
min_lat=self.min_lat, max_lat=self.max_lat,
ds_size_time=self.ds_size_time,
ds_size_lon=self.ds_size_lon,
ds_size_lat=self.ds_size_lat,
time=self.time,
dX=self.dX, dY = self.dY,
swX=self.swX, swY=self.swY,
coord_ext = {'lon_ext': self.lon,
'lat_ext': self.lat},
test_domain=self.cfg.test_domain,
resolution=self.resolution,
original_coords=self.original_coords,
padded_coords=self.padded_coords
)
return mod
def train(self, ckpt_path=None, **trainer_kwargs):
"""
Train a model
:param ckpt_path: (Optional) Checkpoint from which to resume
:param trainer_kwargs: (Optional) Trainer arguments
:return:
"""
mod = self._get_model(ckpt_path=ckpt_path)
checkpoint_callback = ModelCheckpoint(monitor='val_loss',
filename=self.filename_chkpt,
save_top_k=3,
mode='min')
from pytorch_lightning.callbacks import LearningRateMonitor
lr_monitor = LearningRateMonitor(logging_interval='step')
num_nodes = int(os.environ.get('SLURM_JOB_NUM_NODES', 1))
gpus = trainer_kwargs.get('gpus', torch.cuda.device_count())
num_gpus = gpus if isinstance(gpus, (int, float)) else len(gpus) if hasattr(gpus, '__len__') else 0
accelerator = "ddp" if (num_gpus * num_nodes) > 1 else None
trainer_kwargs_final = {**dict(num_nodes=num_nodes, gpus=gpus, logger=self.logger, strategy=accelerator, auto_select_gpus=(num_gpus * num_nodes) > 0,
callbacks=[checkpoint_callback, lr_monitor]), **trainer_kwargs}
print(trainer_kwargs)
print(trainer_kwargs_final)
trainer = pl.Trainer(**trainer_kwargs_final)
trainer.fit(mod, self.dataloaders['train'], self.dataloaders['val'])
return mod, trainer
def test(self, ckpt_path=None, dataloader="test", _mod=None, _trainer=None, **trainer_kwargs):
"""
Test a model
:param ckpt_path: (Optional) Checkpoint from which to resume
:param dataloader: Dataloader on which to run the test Checkpoint from which to resume
:param trainer_kwargs: (Optional)
"""
if _trainer is not None:
_trainer.test(mod, dataloaders=self.dataloaders[dataloader])
return
mod = _mod or self._get_model(ckpt_path=ckpt_path)
trainer = pl.Trainer(num_nodes=1, gpus=1, accelerator=None, **trainer_kwargs)
trainer.test(mod, dataloaders=self.dataloaders[dataloader])
return mod
def profile(self):
"""
Run the profiling
:return:
"""
from pytorch_lightning.profiler import PyTorchProfiler
profiler = PyTorchProfiler(
"results/profile_report",
schedule=torch.profiler.schedule(
wait=1,
warmup=1,
active=1),
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
],
on_trace_ready=torch.profiler.tensorboard_trace_handler('./tb_profile'),
record_shapes=True,
profile_memory=True,
)
self.train(
**{
'profiler': profiler,
'max_epochs': 1,
}
)
def _main(cfg):
print(OmegaConf.to_yaml(cfg))
pl.seed_everything(seed=cfg.get('seed', None))
dm = instantiate(cfg.datamodule)
if cfg.get('callbacks') is not None:
callbacks = [instantiate(cb_cfg) for cb_cfg in cfg.callbacks]
else:
callbacks=[]
if cfg.get('logger') is not None:
print('instantiating logger')
print(OmegaConf.to_yaml(cfg.logger))
logger = instantiate(cfg.logger)
else:
logger=True
lit_mod_cls = get_class(cfg.lit_mod_cls)
runner = FourDVarNetHydraRunner(cfg.params, dm, lit_mod_cls, callbacks=callbacks, logger=logger)
call(cfg.entrypoint, self=runner)
main = hydra.main(config_path='hydra_config', config_name='main')(_main)
if __name__ == '__main__':
main()