-
Notifications
You must be signed in to change notification settings - Fork 14
/
experiments.py
131 lines (117 loc) · 6.09 KB
/
experiments.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
import sys
import h5py
from keras.preprocessing.image import ImageDataGenerator
from pix2pix import Pix2Pix
from lasagne.nonlinearities import *
from lasagne.updates import *
from lasagne.utils import floatX
from util import Hdf5Iterator
def get_iterators(dataset, batch_size, is_a_grayscale, is_b_grayscale, da=True):
dataset = h5py.File(dataset,"r")
if da:
imgen = ImageDataGenerator(horizontal_flip=True, vertical_flip=True, rotation_range=360, fill_mode="reflect")
else:
imgen = ImageDataGenerator()
it_train = Hdf5Iterator(dataset['xt'], dataset['yt'], batch_size, imgen, is_a_grayscale=is_a_grayscale, is_b_grayscale=is_b_grayscale)
it_val = Hdf5Iterator(dataset['xv'], dataset['yv'], batch_size, imgen, is_a_grayscale=is_a_grayscale, is_b_grayscale=is_b_grayscale)
return it_train, it_val
if __name__ == '__main__':
def test1_nobn(mode):
assert mode in ["train", "interp", "gen"]
from architectures import p2p, dcgan
model = Pix2Pix(
gen_fn_dcgan=dcgan.default_generator,
disc_fn_dcgan=dcgan.default_discriminator,
gen_params_dcgan={'num_repeats':0, 'div':[2,2,4,4,8,8,8]},
disc_params_dcgan={'num_repeats':0, 'bn':False, 'nonlinearity':linear, 'div':[8,4,4,4,2,2,2]},
gen_fn_p2p=p2p.g_unet,
disc_fn_p2p=p2p.discriminator,
gen_params_p2p={'nf':64, 'act':tanh, 'num_repeats':0},
disc_params_p2p={'nf':64, 'bn':False, 'num_repeats':0, 'act':linear, 'mul_factor':[1,2,4,8]},
in_shp=512,
latent_dim=1000,
is_a_grayscale=True,
is_b_grayscale=False,
lsgan=True,
opt=rmsprop,
opt_args={'learning_rate':theano.shared(floatX(1e-4))}
)
desert_h5 = "/data/lisa/data/cbeckham/textures_v2_brown500.h5"
bs = 4
it_train, it_val = get_iterators(desert_h5, bs, True, False, True)
name = "test1_repeatnod_fixp2p_nobn"
if mode == "train":
model.train(it_train, it_val, batch_size=bs, num_epochs=1000, out_dir="output/%s" % name, model_dir="models/%s" % name)
elif mode == "interp":
model.load_model("models/%s/600.model.bak" % name)
zs = model.sampler(2, model.latent_dim)
z1, z2 = floatX(zs[0]), floatX(zs[1])
model.generate_interpolation(z1, z2, "/tmp/test.png", mode='matrix')
elif mode == "gen":
model.load_model("models/%s/600.model.bak" % name)
model.generate_gz(100, 10, "deleteme")
def test1_nobn_finetunep2p_bilin(mode):
assert mode in ["train", "interp","gen"]
from architectures import p2p, dcgan
# change the p2p discriminator
model = Pix2Pix(
gen_fn_dcgan=dcgan.default_generator,
disc_fn_dcgan=dcgan.default_discriminator,
gen_params_dcgan={'num_repeats':0, 'div':[2,2,4,4,8,8,8]},
disc_params_dcgan={'num_repeats':0, 'bn':False, 'nonlinearity':linear, 'div':[8,4,4,4,2,2,2]},
gen_fn_p2p=p2p.g_unet,
disc_fn_p2p=p2p.discriminator,
gen_params_p2p={'nf':64, 'act':tanh, 'num_repeats':0, 'bilinear_upsample':True},
disc_params_p2p={'nf':64, 'bn':False, 'num_repeats':0, 'act':linear, 'mul_factor':[1,2,4,8]},
in_shp=512,
latent_dim=1000,
is_a_grayscale=True,
is_b_grayscale=False,
lsgan=True,
opt=rmsprop,
opt_args={'learning_rate':theano.shared(floatX(1e-4))},
train_mode='p2p'
)
model.load_model("models/test1_repeatnod_fixp2p_nobn/1000.model.bak", mode='dcgan') # only load the dcgan
desert_h5 = "/data/lisa/data/cbeckham/textures_v2_brown500.h5"
bs = 4
it_train, it_val = get_iterators(desert_h5, bs, True, False, True)
name = "test1_repeatnod_fixp2p_nobn_finetunep2p_bilin"
# models/test1_repeatnod_fixp2p_nobn/600.model.bak --> good DCGAN?? (not good p2p!!)
# models/test1_repeatnod_fixp2p_nobn_finetunep2p_bilin/1000.model.bak --> good p2p (not good DCGAN??)
if mode == "train":
model.train(it_train, it_val, batch_size=bs, num_epochs=1000, out_dir="output/%s" % name, model_dir="models/%s" % name)
elif mode == "interp":
model.load_model("models/test1_repeatnod_fixp2p_nobn/1000.model.bak", mode='dcgan')
model.load_model("models/test1_repeatnod_fixp2p_nobn_finetunep2p_bilin/1000.model.bak", mode='p2p')
model.generate_interpolation_clip(100, 4, "output/%s/interp_clip_600_concat_bothdet/" % name, concat=True, deterministic=True)
# let's use bilinear upsampling with the dcgan
def test1_nobn_bilin_both(mode):
assert mode in ["train", "interp","gen"]
from architectures import p2p, dcgan
# change the p2p discriminator
model = Pix2Pix(
gen_fn_dcgan=dcgan.default_generator,
disc_fn_dcgan=dcgan.default_discriminator,
gen_params_dcgan={'num_repeats':0, 'div':[2,2,4,4,8,8,8]},
disc_params_dcgan={'num_repeats':0, 'bn':False, 'nonlinearity':linear, 'div':[8,4,4,4,2,2,2]},
gen_fn_p2p=p2p.g_unet,
disc_fn_p2p=p2p.discriminator,
gen_params_p2p={'nf':64, 'act':tanh, 'num_repeats':0, 'bilinear_upsample':True},
disc_params_p2p={'nf':64, 'bn':False, 'num_repeats':0, 'act':linear, 'mul_factor':[1,2,4,8]},
in_shp=512,
latent_dim=1000,
is_a_grayscale=True,
is_b_grayscale=False,
lsgan=True,
opt=rmsprop,
opt_args={'learning_rate':theano.shared(floatX(1e-4))},
train_mode='both'
)
desert_h5 = "/data/lisa/data/cbeckham/textures_v2_brown500.h5"
bs = 4
it_train, it_val = get_iterators(desert_h5, bs, True, False, True)
name = "test1_nobn_bilin_both_deleteme"
if mode == "train":
model.train(it_train, it_val, batch_size=bs, num_epochs=1000, out_dir="output/%s" % name, model_dir="models/%s" % name)
locals()[ sys.argv[1] ]( sys.argv[2] )