-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_model.py
212 lines (179 loc) · 8.99 KB
/
train_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
from PIESRGAN import PIESRGAN
from util import DataLoader_s3d, do_normalisation, Image_generator2, do_inverse_normalisation
from util import Image_generator
import time
import tensorflow as tf
import numpy as np
from gan_post import GAN_post
from tensorflow.keras.callbacks import CSVLogger, ReduceLROnPlateau
import pickle
#import horovod.tensorflow.keras as hvd
def generate_image(pred, imgs_lr, imgs_hr, mins, maxs, norm, idx):
pred = do_inverse_normalisation(pred, norm, mins, maxs)
pred = train_loader_hr.inverse_reshape_array(pred)
imgs_lr = train_loader_hr.inverse_reshape_array(imgs_lr)
imgs_lr = do_inverse_normalisation(imgs_lr, norm, mins, maxs)
imgs_hr = train_loader_hr.inverse_reshape_array(imgs_hr)
imgs_hr = do_inverse_normalisation(imgs_hr, norm, mins, maxs)
Image_generator2(imgs_lr[:,:,0,0], pred[:,:,0,0], imgs_hr[:,:,0,0], 'testfig_{}.png'.format(idx))
def generate_image_slice(weights, i, norm, DNS, Filt, pref):
Filt.weights=weights
pred = Filt.get_gan_slice_serial(0,48, norm, data=None)
HR = DNS.get_data_plane(0)
#Filt=DataLoader_s3d(datapath_train+'/Filt_4x/filt_s-1.5000E-05', 1536, 1536, 1536, 2, batch_size, boxsize, 1)
LR = Filt.get_data_plane(0)
Image_generator(LR[:,:,0], pred[:,:,0,0], HR[:,:,0], pref+'_{}.png'.format(i))
return
def generate_spectrum(weights, i, norm, Filt, pref):
Filt.weights=weights
spectrum = Filt.get_spectrum_slice(48, 5e-3, 0, 5e-3, 0, 0, norm, savePred=None, pref=pref+"_{}_".format(i))
return
def save_model_generator(model, savedir, idx):
model.generator.save_weights(savedir+'generator_idx_{}.h5'.format(idx))
opt_wght = model.optimizer.get_weights()
pickle.dump(opt_wght,open(savedir+'generator_opt_idx_{}.h5'.format(idx),'wb'))
return
def load_model_generator(model,savedir, idx):
print("loading model from epoch {}".format(idx))
model.generator.load_weights(savedir+'generator_idx_{}.h5'.format(idx))
opt_wght = pickle.load(open(savedir+'generator_opt_idx_{}.h5'.format(idx),'rb'))
def load_opt():
grad_vars = model.generator.trainable_weights
zero_grads = [tf.zeros_like(w) for w in grad_vars]
model.optimizer.apply_gradients(zip(zero_grads, grad_vars))
model.optimizer.set_weights(opt_wght)
model.mirrored_strategy.run(load_opt)
return
def save_model_gan(model, savedir, idx):
model.gen_gan.save_weights(savedir+'gen_gan_idx_{}.h5'.format(idx))
opt_wght = model.generator_optimizer.get_weights()
pickle.dump(opt_wght,open(savedir+'gen_gan_opt_idx_{}.h5'.format(idx),'wb'))
model.dis_gan.save_weights(savedir+'dis_gan_idx_{}.h5'.format(idx))
opt_wght = model.discriminator_optimizer.get_weights()
pickle.dump(opt_wght,open(savedir+'dis_gan_opt_idx_{}.h5'.format(idx),'wb'))
return
def load_model_gan(model,savedir, idx, Gen=True, Dis=False):
print("loading model from epoch {}".format(idx))
if(Gen):
model.gen_gan.load_weights(savedir+'gen_gan_idx_{}.h5'.format(idx))
opt_wght = pickle.load(open(savedir+'gen_gan_opt_idx_{}.h5'.format(idx),'rb'))
def load_opt():
grad_vars = model.gen_gan.trainable_weights
zero_grads = [tf.zeros_like(w) for w in grad_vars]
model.generator_optimizer.apply_gradients(zip(zero_grads, grad_vars))
model.generator_optimizer.set_weights(opt_wght)
model.mirrored_strategy.run(load_opt)
if(Dis):
model.dis_gan.load_weights(savedir+'dis_gan_idx_{}.h5'.format(idx))
opt_wght = pickle.load(open(savedir+'dis_gan_opt_idx_{}.h5'.format(idx),'rb'))
def load_opt_dis():
grad_vars = model.dis_gan.trainable_weights
zero_grads = [tf.zeros_like(w) for w in grad_vars]
model.discriminator_optimizer.apply_gradients(zip(zero_grads, grad_vars))
model.discriminator_optimizer.set_weights(opt_wght)
model.mirrored_strategy.run(load_opt_dis)
return
nxg, nyg, nzg =192, 192, 192
batch_size=8
boxsize=32
datapath_train='/scratch/w47/share/IsotropicTurb'
train_loader_hr = DataLoader_s3d(datapath_train+'/Filtered_Relambda_162_up_50_Lt_2mm/DNS_regrid_192grid/s-7.0000E-06', nxg, nyg, nzg, 2, batch_size, boxsize, 3)
train_loader_hr.get_norm_constants(48)
train_loader_lr=GAN_post(datapath_train+'/Filtered_Relambda_162_up_50_Lt_2mm/Filt_8x_regrid_192grid/s-7.0000E-06', nxg, nyg, nzg, 2, batch_size, boxsize, 3, 1)
#idx=0
#imgs_lr = do_normalisation(train_loader_lr.getTrainData_plane(48,0,idx),'minmax', mins, maxs)
#imgs_hr = do_normalisation(train_loader_hr.getTrainData_plane(48,0,idx),'minmax',mins,maxs)
epochs=50000000
print_frequency = 200
save_frequency = 50
fig_frequency = 500
savedir = './data/regridded/'
#mirrored_strategy = tf.distribute.MirroredStrategy()
mode = 1 # 1 = GAN, 0 = generator
norm='minmax'
mins = train_loader_hr.mins
maxs = train_loader_hr.maxs
if(mode==1):
logfile = open('GAN_logs_regrid','w')
logfile.write('epoch \t total_loss \t grad_loss\t gen_loss\t pixel_loss\t cont_loss\t dis_loss\n')
gan = PIESRGAN(training_mode=True,
height_lr = boxsize, width_lr=boxsize, depth_lr=boxsize,
gen_lr=1e-5, dis_lr=1.0e-5,
channels=3,
RRDB_layers=6
)
gan.build_gan(gen_weights=savedir+'generator_idx_6800.h5', disc_weights=None)
#gan.build_gan()
#load_model_gan(gan, savedir, 5700, Gen=True, Dis=True)
#out = gan.distributed_train_step(imgs_lr, imgs_hr)
idx=0
isave=0
for i in range(0,epochs):
imgs_lr = do_normalisation(train_loader_lr.getData(idx),norm, mins, maxs)
imgs_hr = do_normalisation(train_loader_hr.getData(idx),norm,mins,maxs)
idx=idx+1
if(idx>train_loader_lr.nbatches-1):
idx=0
t1=time.time()
total_loss, grad_loss, gen_loss, pixel_loss, cont_loss, disc_loss = gan.distributed_train_step(imgs_lr, imgs_hr)
logfile.write("{},{},{},{},{},{},{}\n".format(i, total_loss, grad_loss, gen_loss, pixel_loss, cont_loss, disc_loss))
logfile.flush()
if(i%print_frequency == 0):
print("EPOCH : {} TOTAL GEN LOSS : {} DISC LOSS : {} EPOCH TIME {} secs".format(i, total_loss, disc_loss, time.time()-t1))
if(i%save_frequency == 0):
save_model_gan(gan,savedir, i)
isave=i
if(i%fig_frequency==0):
if(i==0):
continue
generate_image_slice(savedir+"gen_gan_idx_{}.h5".format(isave),isave, norm, train_loader_hr, train_loader_lr, 'slice_gan_regrid_')
#generate_spectrum(savedir+"gen_gan_idx_{}.h5".format(isave),isave, norm, train_loader_lr, 'pred_gan_s-7.0000E-06_8_l12')
# pred = gan.gen_gan(imgs_lr)
# print(np.sum((pred-imgs_hr)**2))
# generate_image(pred, imgs_lr, imgs_hr, mins, maxs, i)
save_model_gan(gan,savedir, epochs)
logfile.close()
if(mode==0):
gan = PIESRGAN(training_mode=True,
height_lr = boxsize, width_lr=boxsize, depth_lr=boxsize,
gen_lr=5e-5, dis_lr=2e-6,
channels=3, RRDB_layers=6
)
load_model_generator(gan, savedir, 2800)
#gan.generator.load_weights(savedir+'generator_idx_11000.h5')
#gan.generator = tf.keras.models.load_model(savedir+'generator_idx_100')
csv_logger = CSVLogger("regridded.csv", append=True)
#reduce_lr = ReduceLROnPlateau(monitor='loss', factor=0.75, patience=5, min_lr=1e-6, verbose=1)
#hvd_callback = hvd.callbacks.BroadcastGlobalVariablesCallback(0)
idx=0
isave=700
for i in range(2801,epochs):
imgs_lr = do_normalisation(train_loader_lr.getData(idx),norm, mins, maxs)
imgs_hr = do_normalisation(train_loader_hr.getData(idx),norm,mins,maxs)
idx=idx+1
if(idx>train_loader_lr.nbatches-1):
idx=0
t1=time.time()
gan.generator.fit(
imgs_lr, imgs_hr,
steps_per_epoch=1,
epochs=1,
callbacks = [csv_logger],
)
if(i%print_frequency==0):
print("Epoch {} out of nbatches {} took {} secs".format(i,train_loader_lr.nbatches,time.time()-t1))
if(i%save_frequency == 0):
#gan.generator.save_weights(savedir+"generator_idx_{}.h5".format(i))
save_model_generator(gan, savedir, i)
#gan.generator.save(savedir+"generator_idx_{}".format(i))
isave=i
if(i%fig_frequency==0):
if(i==0):
continue
generate_image_slice(savedir+"generator_idx_{}.h5".format(isave),isave, norm, train_loader_hr, train_loader_lr, 'regredded')
#generate_spectrum(savedir+"generator_idx_{}.h5".format(isave),isave, norm, train_loader_lr, 'pred_regredded')
# pred = gan.gen_gan(imgs_lr)
# print(np.sum((pred-imgs_hr)**2))
# generate_image(pred, imgs_lr, imgs_hr, mins, maxs, i)
#gan.generator.save_weights(savedir+"generator_idx_{}.h5".format(epochs))
save_model_generator(gan, savedir, epochs)