-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtraining.py
165 lines (138 loc) · 6.15 KB
/
training.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
import xarray as xr
import numpy as np
import importlib
from matplotlib import pyplot as plt
import DA_core as DA
from glob import glob
import torch.utils.data as Data
from torch import optim
import torch
from torchsummary import summary
import ML_core as ML
from numpy.random import default_rng
import os
from torch.profiler import profile, record_function, ProfilerActivity
rng = default_rng()
DA.read_data_dir='/scratch/cimes/feiyul/PyQG/data/training'
DA.save_data_dir='/scratch/cimes/feiyul/PyQG/data/training'
data_dir='/scratch/cimes/feiyul/PyQG/data'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
DA_paras={'nens':80,
'DA_method':'EnKF',
'Nx_DA':32,
'Nx_truth':128,
'obs_freq':10,
'obs_err':[1,-5,5,-7],
'DA_freq':10,
'save_B':False,
'nobs':[50,50],
'R_W':100,
'inflate':[1,0.5]}
DA_exp=DA.DA_exp(**DA_paras)
print(DA_exp.file_name())
# obs_ds=DA_exp.read_obs()
in_ch=[0,1]
out_ch=[0,1,2]
print(in_ch,out_ch)
### Make direcotry for storing networks
os.makedirs('./ML/{}'.format(DA_exp.file_name()),exist_ok=True)
### Find time indices for the DA steps to select proper q and B data
DA_days=slice(369,3650,DA_exp.DA_freq)
DA_it=slice(int((DA_days.start-DA_exp.DA_freq+1)/DA_exp.DA_freq),int((DA_days.stop-DA_exp.DA_freq+1)/DA_exp.DA_freq)+1)
print(DA_days,DA_it)
### range of indices to select from q and B data
i_x=slice(0,DA_exp.Nx_DA)
i_y=slice(0,DA_exp.Nx_DA)
### size and starting index for the B data in U-Nets
B_size=16
B_start=0
### Read the saved covariance matrices from previous EnKF experiments
B_ens_ds=xr.open_dataset('{}/training/{}/B_ens.nc'.format(data_dir,DA_exp.file_name()))
B_ens=B_ens_ds.B_ens.isel(time=DA_it,y=i_y,x=i_x)
print(B_ens.shape)
### Read the saved ensemble-mean analysis q from previous EnKF experiments
mean_ds=DA_exp.read_mean().load()
q_full=mean_ds.q.isel(time=DA_days,y=i_y,x=i_x)
print(q_full.shape)
### Read or calculate standard deviations for normalization
if os.path.exists('./ML/{0}/std_{0}.nc'.format(DA_exp.file_name())):
ml_std_ds=xr.open_dataset('./ML/{0}/std_{0}.nc'.format(DA_exp.file_name()))
else:
B_std=np.empty((2,2))
B_std[0,0]=np.std(B_ens.isel(lev=0,lev_d=0))
B_std[0,1]=np.std(B_ens.isel(lev=0,lev_d=1))
B_std[1,0]=B_std[0,1]
B_std[1,1]=np.std(B_ens.isel(lev=1,lev_d=1))
q_std=np.zeros((2,1))
q_std[0]=np.std(q_full.isel(time=DA_days,lev=0))
q_std[1]=np.std(q_full.isel(time=DA_days,lev=1))
ml_std_ds=xr.Dataset({'B_std':xr.DataArray(B_std,coords=[mean_ds.lev,mean_ds.lev]),
'q_std':xr.DataArray(q_std.squeeze(),coords=[mean_ds.lev])})
ml_std_ds.to_netcdf('./ML/{0}/std_{0}.nc'.format(DA_exp.file_name()))
print(ml_std_ds)
### Process B data for training
B_stacked=B_ens.stack(sample=('time','y','x')).transpose('sample',...)
print(B_stacked.shape)
B_data=np.empty((len(B_ens.time)*len(B_ens.y)*len(B_ens.x),3,len(B_ens.y_d),len(B_ens.x_d)))
B_data[:,0,...]=B_stacked[:,0,0,...]/ml_std_ds.B_std[0,0].data
B_data[:,1,...]=B_stacked[:,0,1,...]/ml_std_ds.B_std[0,1].data
B_data[:,2,...]=B_stacked[:,1,1,...]/ml_std_ds.B_std[1,1].data
print(B_data.shape)
### Process q data for training
q_local=np.empty((len(q_full.time),len(q_full.lev),len(q_full.y),len(q_full.x),len(B_ens.y_d),len(B_ens.x_d)))
for i in range(len(q_full.x)):
for j in range(len(q_full.y)):
q_local[:,:,j,i,:,:]=DA.localize_q(q_full,j,i,DA_exp.Nx_DA,int(len(B_ens.x_d)/2))
q_local=q_local.transpose([0,2,3,1,4,5])
print(q_local.shape)
q_data=q_local.reshape((len(q_full.time)*len(q_full.y)*len(q_full.x),len(q_full.lev),len(B_ens.y_d),len(B_ens.x_d)))
print(q_data.shape)
q_data[:,0,...]=q_data[:,0,...]/ml_std_ds.q_std[0].data
q_data[:,1,...]=q_data[:,1,...]/ml_std_ds.q_std[1].data
q_unet=q_data[...,B_start:B_start+B_size,B_start:B_start+B_size]
B_unet=B_data[...,B_start:B_start+B_size,B_start:B_start+B_size]
B_shape=B_unet.shape
q_shape=q_unet.shape
print(B_shape,q_shape)
n_total=B_shape[0]
n_train=int(n_total*0.8)
train_ds=ML.Dataset(q_unet[0:n_train,...],B_unet[0:n_train,...],device)
valid_ds=ML.Dataset(q_unet[n_train:,...],B_unet[n_train:,...],device)
params = {'batch_size':64000,'num_workers':16,'shuffle':True}
training_generator = torch.utils.data.DataLoader(train_ds, **params)
validation_generator = torch.utils.data.DataLoader(valid_ds, **params)
features=16
Ulevels=2
if Ulevels==3:
model=ML.Unet(in_ch=len(in_ch),out_ch=len(out_ch),features=features)
elif Ulevels==2:
model=ML.Unet_2L(in_ch=len(in_ch),out_ch=len(out_ch),features=features)
model=model.to(device)
os.makedirs('./ML/{}/{}L{}f'.format(DA_exp.file_name(),Ulevels,features),exist_ok=True)
# check keras-like model summary using torchsummary
summary(model, input_size=q_shape[1:])
criterion = torch.nn.MSELoss() # MSE loss function
optimizer = optim.Adam(model.parameters(), lr=0.002)
model=model.double()
n_epochs = 200 #Number of epocs
validation_loss = list()
train_loss = list()
start_epoch=0
if start_epoch>0:
model_file='./ML/{}/{}L{}f/unet_epoch{}_in{}_out{}_B{}_{}.pt'.format(
DA_exp.file_name(),Ulevels,features,start_epoch,''.join(map(str,in_ch)),
''.join(map(str,out_ch)),B_size,DA_exp.file_name())
print(model_file)
model.load_state_dict(torch.load(model_file,map_location=torch.device('cpu')))
# time0 = time()
with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
profile_memory=True,record_shapes=True) as prof:
with record_function("model_training"):
for epoch in range(start_epoch+1, n_epochs + 1):
train_loss.append(ML.train_model(model,criterion,training_generator,optimizer,device))
validation_loss.append(ML.test_model(model,criterion,validation_generator,optimizer,device))
torch.save(model.state_dict(), './ML/{}/{}L{}f/unet_epoch{}_in{}_out{}_B{}_{}.pt'.\
format(DA_exp.file_name(),Ulevels,features,epoch,''.join(map(str,in_ch)),
''.join(map(str,out_ch)),B_size,DA_exp.file_name()))
print(prof.key_averages(group_by_input_shape=True).table(sort_by="cpu_time_total", row_limit=20))
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))