/
main.py
176 lines (162 loc) · 6.96 KB
/
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
import sys
import argparse
from model import *
from trainer import Trainer
parser = argparse.ArgumentParser()
# Path configurations
parser.add_argument("-savepath",
type=str,
default='out',
help="path to save output")
parser.add_argument("-prefix",
type=str,
default=None,
help="experiment prefix")
parser.add_argument("-fcnpath",
type=str,
default='/home/thanard/Downloads/FCN_mse',
help="path to fcn parameters for background subtraction")
parser.add_argument("-data_dir",
type=str,
default='/home/thanard/Downloads/rope_full',
help='path to rope_full data')
parser.add_argument("-planning_data_dir",
type=str,
default='/home/thanard/Downloads/seq_data_2',
help='path to seq_data_2 data')
parser.add_argument("-loadpath",
type=str,
default='',
help="path to the previous exp to load parameters from")
parser.add_argument("-loadepoch",
type=int,
default=None,
help="epoch number to load from")
parser.add_argument("-classifier_path", type=str,
default="classifier.pkl",
help="path to classifier parameters. "
"The classifier is pretrained on real images to classify "
"image pairs that are one step apart. "
"We use it to evaluate how feasible image transitions are,"
"and to select the best k plans.")
# Training hyperparameters
parser.add_argument("-seed", type=int, default=0)
parser.add_argument("-n_epochs", type=int, default=100)
parser.add_argument("-cc", type=int, default=7,
dest="cont_code_dim",
help="continuous code dimension")
parser.add_argument("-rn", type=int, default=2,
dest="random_noise_dim",
help="dimension of random noise")
parser.add_argument("-infow", type=float, default=0.1,
help="weight of the information term.")
parser.add_argument("-transw", type=float, default=0.1,
help="weight of the transition regularization term.")
parser.add_argument("-lr_d", type=float, default=0.0002)
parser.add_argument("-lr_g", type=float, default=0.0002)
# TODO: fix gtype, dtype, and k.
parser.add_argument("-gtype", type=int, default=1,
help="which architecture of generator to use")
parser.add_argument("-dtype", type=int, default=1,
help="which architecture of discriminator to use")
parser.add_argument("-qtype", type=int, default=1,
help="which architecture of posterior to use")
parser.add_argument("-tsize", type=int, default=[64, 64], nargs="+",
help="hidden size of Transition NN.")
parser.add_argument("-k", type=int, default=1,
help="the number of timesteps apart for training")
parser.add_argument("-color", action="store_true")
parser.add_argument("-learn_mu", action="store_true")
parser.add_argument("-learn_var", action="store_true")
# Planning
parser.add_argument("-planning_epoch", type=int, default=[100], nargs="+",
help="List of epoch numbers to run planning.")
parser.add_argument("-plan_length", type=int, default=10,
help="Set to 0 if doesn't run planning.")
parser.add_argument("-traj_eval_copies", type=int, default=100,
help='the number of plans to choose from.')
parser.add_argument("-planner", type=str, default='simple_plan',
help="either simple_plan or astar_plan")
args = parser.parse_args()
kwargs = vars(args)
# Construct more arguments
if args.prefix is None:
str_list = ["continuous",
"gtype", str(args.gtype),
"rn", str(args.random_noise_dim),
"cc", str(args.cont_code_dim),
"infow", "%.2f" % args.infow,
"transw", "%.2f" % args.transw,
]
if args.plan_length > 0 and os.path.exists(args.planning_data_dir) and args.planner:
str_list.append(args.planner)
if args.fcnpath:
str_list.append("fcn")
if args.learn_mu:
str_list.append("mu")
if args.learn_var:
str_list.append("var")
args.prefix = "-".join(str_list)
print("Experiment name : ", args.prefix)
kwargs['python_cmd'] = " ".join(sys.argv)
kwargs['gray'] = not kwargs['color']
kwargs['out_dir'] = os.path.join(args.savepath, args.prefix)
if kwargs['gray'] or kwargs['fcnpath']:
kwargs['channel_dim'] = channel_dim = 1
else:
kwargs['channel_dim'] = channel_dim = 3
# Set initial seed
seed = kwargs['seed']
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
# Make output folders
out_dir = kwargs['out_dir']
for folder in ['gen', 'real', 'plans']:
if not os.path.exists(os.path.join(out_dir, folder)):
os.makedirs(os.path.join(out_dir, folder))
# Save configuration parameters
import json
with open('%s/params.json' % out_dir, 'w') as fp:
json.dump(kwargs, fp, indent=4, sort_keys=True)
# Initialize Generator, Discriminator, Posterior, and FCN networks.
c_dim = kwargs['cont_code_dim']
z_dim = kwargs['random_noise_dim']
g = G(c_dim, z_dim, kwargs['gtype'], channel_dim)
d = D(kwargs['dtype'], channel_dim)
q = GaussianPosterior(c_dim, kwargs['qtype'], channel_dim)
t = GaussianTransition(c_dim,
hidden=kwargs['tsize'],
learn_var=kwargs['learn_var'],
learn_mu=kwargs['learn_mu'])
p = UniformDistribution(kwargs['cont_code_dim'])
var_list = [g, d, q, t, p]
kwargs['classifier'] = get_causal_classifier(kwargs['classifier_path'], default=d)
if kwargs['fcnpath']:
fcn_model = FCN_mse(n_class=2).cuda()
fcn_model.load_state_dict(torch.load(os.path.join(kwargs['fcnpath'])))
fcn_model.eval()
kwargs['fcn'] = fcn_model
# Initialize or load from previously trained networks
loadpath = kwargs['loadpath']
loadepoch = kwargs['loadepoch']
for i in var_list:
print(i)
i.cuda()
i.apply(weights_init)
if loadpath:
if i not in [p]:
try:
i.load_state_dict(torch.load(os.path.join(loadpath,
'var',
'%s_%d' % (i.__class__.__name__,
loadepoch))))
print("Loaded var %s from iter %d." % (i.__class__.__name__,
loadepoch))
except FileNotFoundError as e:
print("Couldn't load var %s" % i.__class__.__name__)
pass
# Training the variables
trainer = Trainer(*var_list, **kwargs)
trainer.train()