-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
174 lines (144 loc) · 8.6 KB
/
train.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
import numpy as np
import argparse
import tensorflow as tf
from tensorflow.keras.callbacks import ModelCheckpoint
from data.cyclesps import CyclesDataGenerator
from misc.metrics import AvgAngleMetric, GaussAvgAngleMetrics
from models.cnnps import densenet_2D
from models.pxnet import pxnet_2D_v2, pxnet_2D_v1
from models.unet import unet
from models.cnnps_sep4D import densenet_separable4D, densent_separable4D_3x3, densent_separable4D_5x5
from models.unet_sep4D import unet_sep4d
from misc.layres import Random90Rotation, RotateVector
from misc.projections import parse_projection
from misc.losses import GaussMSE, ProjectedSoftmax2D
from misc.layres import Gauss2D
from tensorflow.keras.optimizers import RMSprop, Adam
from tensorflow.keras.optimizers.schedules import ExponentialDecay
layers_dict = {"RotateVector": RotateVector,
"Random90Rotation": Random90Rotation,
"AvgAngleMetric": AvgAngleMetric,
"GaussMSE": GaussMSE,
"SpatialGaussMSE": GaussMSE,
"GaussAvgAngleMetrics": GaussAvgAngleMetrics,
"Gauss2D": Gauss2D,
"ProjectedSoftmax2D": ProjectedSoftmax2D,
"tf": tf}
parser = argparse.ArgumentParser(description='Photometric Stereo network training')
parser.add_argument('--seed', '-s', type=int, default=0, help='Random seed')
parser.add_argument('--epochs', '-e', type=int, default=10, help='Number of epochs')
parser.add_argument('--size', '-w', type=int, default=48, help='Size of the observation map')
parser.add_argument('--neighbours', '-n', type=int, default=5, help='Size of the spatial patch')
parser.add_argument('--rotations', '-K', type=int, default=12, help='Number of rotations')
parser.add_argument('--model', '-m', type=str, default="cnnps",
help='Architecture type [cnnps, cnnps4D, unet, unet4D, pxnet]')
parser.add_argument('--batch_size', '-b', type=int, default=768, help='Batch size')
parser.add_argument('--save_every', '-v', type=int, default=1, help='Save model every [V] epochs')
parser.add_argument('--nr_features', '-f', type=int, default=16, help='Nuber of features for U-Net based models')
parser.add_argument('--nr_blocks', '-q', type=int, default=3, help='Nuber of blocks for U-Net based models')
parser.add_argument('--hm_std', type=float, default=2, help='Heat-map standard deviation for U-Net based models')
parser.add_argument('--order', type=int, default=3, help='Spline interpolation order used if neighbours > 1')
parser.add_argument('--use_BN', dest="use_BN", action="store_true", help='Use BatchNormalization layer in U-Net')
parser.add_argument('--dividemaps', dest="dividemaps", action="store_true",
help='Divide observation maps by the max value e.g. for cnnps model')
parser.add_argument('--suffix', type=str, default="", help='Suffix to the name under which the models will be saved')
parser.add_argument('--memory_limit', '-g', type=int, default=-1,
help='If > 0, GPU memory to allocate in MB on the first GPU. If <= 0 all GPUs are fully allocated')
parser.add_argument('--add_raw', dest="add_raw", action="store_true",
help="Adds RAW (non-normalized) colour channels to the observation maps")
parser.add_argument('--optimizer', type=str, default="rmsprop", help="Optimizer")
parser.add_argument('--dataset', type=str, default="cycles", help="Training dataset [cycles]")
parser.add_argument('--dataset_path', type=str, default="datasets/CyclesPS/", help="Path to the dataset")
args = parser.parse_args()
print(args)
# Set random seeds
np.random.seed(args.seed)
tf.random.set_seed(args.seed)
# Set the memory limit for the GPUs
gpus = tf.config.list_physical_devices('GPU')
if gpus and args.memory_limit > 0:
for gpu in gpus:
tf.config.experimental.set_virtual_device_configuration(gpu, [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=args.memory_limit)])
# Define strategy for Multi-GPU training
mirrored_strategy = tf.distribute.MirroredStrategy()
# Define projection from the 3D unit sphere to a 2D plane
projection = parse_projection("standard")
# Define & compile the selected model
nr_channels = 4 if args.add_raw else 1
with mirrored_strategy.scope():
if args.model == "cnnps":
args.neighbours = 1
model = densenet_2D(args.size, args.size, nr_channels)
model.compile(optimizer=args.optimizer, loss='mean_squared_error', metrics=[AvgAngleMetric()])
is4D = False
elif args.model == "cnnps4D":
if args.neighbours == 3:
model = densent_separable4D_3x3((args.size, args.size), nr_channels, False)
elif args.neighbours == 5:
model = densent_separable4D_5x5((args.size, args.size), nr_channels, False)
else:
model = densenet_separable4D((args.size, args.size), (args.neighbours, args.neighbours), nr_channels, False)
model.compile(optimizer=args.optimizer, loss='mean_squared_error', metrics=[AvgAngleMetric()])
is4D = True
elif args.model == "unet":
args.neighbours = 1
model = unet((args.size, args.size, nr_channels), 1, args.nr_features, args.nr_blocks, 2, use_BN=args.use_BN)
lr_schedule = ExponentialDecay(0.001, decay_steps=(1000000/args.batch_size), decay_rate=0.985, staircase=True)
if args.optimizer == "rmsprop":
optimizer = RMSprop(learning_rate=lr_schedule)
elif args.optimizer == "adam":
optimizer = Adam(learning_rate=lr_schedule)
else:
raise ValueError("Only rmsprop and adam optimizers are supported by this model")
model.compile(optimizer=optimizer,
loss=GaussMSE(args.hm_std, args.size), # loss=ProjectedSoftmax2D(args.size),
metrics=[GaussAvgAngleMetrics(args.size, 1, spherical=False)])
is4D = False
elif args.model == "unet4D":
if args.neighbours == 3 or args.neighbours == 5 or args.neighbours == 7 or args.neighbours == 9:
model = unet_sep4d((args.neighbours, args.neighbours, args.size, args.size, nr_channels), 1, args.nr_features, args.nr_blocks, 2, use_BN=args.use_BN)
else:
raise NotImplementedError("Only spatial patches of 3x3, 5x5, 7x7 and 9x9 are supported by this model")
lr_schedule = ExponentialDecay(0.001, decay_steps=(1000000 / args.batch_size), decay_rate=0.985, staircase=True)
if args.optimizer == "rmsprop":
optimizer = RMSprop(learning_rate=lr_schedule)
elif args.optimizer == "adam":
optimizer = Adam(learning_rate=lr_schedule)
else:
raise ValueError("Only rmsprop and adam optimizers are supported by this model")
model.compile(optimizer=optimizer,
loss=GaussMSE(args.hm_std, args.size),
metrics=[GaussAvgAngleMetrics(args.size, 1, spherical=False)])
is4D = True
elif args.model == "pxnet":
# model = pxnet_2D_v1(args.size, args.size, nr_channels)
model = pxnet_2D_v2(args.size, args.size, nr_channels)
model.compile(optimizer=args.optimizer, loss='mean_squared_error', metrics=[AvgAngleMetric()])
is4D = False
else:
raise NotImplementedError("Unknown model, use: cnnps, cnnps4D, unet, unet4D, or pxnet.")
model.summary(line_length=120)
print(model.optimizer)
print(model.loss)
# Define the training & validation data generators
dataset_args = {
'batch_size': args.batch_size, 'shuffle': True, 'random_illums': True,
'spatial_patch_size': args.neighbours, 'obs_map_size': args.size, 'keep_axis': is4D, 'projection': projection,
'add_raw': args.add_raw, 'divide_maps': args.dividemaps,
'nr_rotations': args.rotations, 'order': args.order, 'rot_2D': False,
'verbose': True
}
if args.dataset == "cycles":
dg_train = CyclesDataGenerator(args.dataset_path, objlist=None, validation_split=0.1, **dataset_args)
dg_valid = dg_train.get_validation_generator()
else:
raise ValueError("Unknown dataset, only 'cycles' dataset is currently supported")
# Define the callbacks for the training
model_chackpoint = ModelCheckpoint("checkpoints/M" + args.model + "_K" + str(args.rotations) +
"_N" + str(args.neighbours) + "_B" + str(args.batch_size) +
"_s" + str(args.seed) + args.suffix + "-{epoch:02d}-{val_loss:.4f}.hdf5",
monitor='val_loss', verbose=0, save_best_only=False, save_weights_only=False,
mode='auto', period=args.save_every)
# Launch the training
hist = model.fit(dg_train, epochs=args.epochs, verbose=1, callbacks=[model_chackpoint],
validation_data=dg_valid, initial_epoch=0)