forked from CVLAB-Unibo/Real-time-self-adaptive-deep-stereo
-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_wandb.py
213 lines (201 loc) · 8.94 KB
/
train_wandb.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
209
210
211
212
213
import os
import tensorflow as tf
import argparse
from datetime import datetime
from madnet import MADNet
from preprocessing import StereoDatasetCreator
from losses_and_metrics import Bad3, EndPointError, ReconstructionLoss, SSIMLoss
import wandb
from wandb.keras import WandbCallback
from callbacks import WandBImagesCallback, TensorboardImagesCallback
wandb.login()
parser = argparse.ArgumentParser(description='Script for training MADNet and '
'logging to Weights and Biases dashboard.')
parser.add_argument("--train_left_dir", help='path to left images folder', required=True)
parser.add_argument("--train_right_dir", help='path to right images folder', required=True)
parser.add_argument("--train_disp_dir", help='path to left disparity maps folder', default=None, required=False)
parser.add_argument("--val_left_dir", help='path to left images folder', default=None, required=False)
parser.add_argument("--val_right_dir", help='path to right images folder', default=None, required=False)
parser.add_argument("--val_disp_dir", help='path to left disparity maps folder', default=None, required=False)
parser.add_argument("--shuffle", help='shuffle training dataset', action="store_true", default=False)
parser.add_argument("--search_range", help='maximum displacement (ie. smallest disparity)',
default=2, type=int, required=False)
parser.add_argument("-o", "--output_dir",
help='path to folder for outputting tensorboard logs and saving model weights',
required=True)
parser.add_argument("--weights_path",
help='One of the following pretrained weights (will download automatically): '
'"synthetic", "kitti", "tf1_conversion_synthetic", "tf1_conversion_kitti"'
'or a path to pretrained MADNet weights file (for fine turning)',
default=None, required=False)
parser.add_argument("--lr", help="Initial value for learning rate.", default=0.0001, type=float, required=False)
parser.add_argument("--min_lr", help="Minimum learning rate cap.", default=0.0000001, type=float, required=False)
parser.add_argument("--decay", help="Exponential decay rate.", default=0.999, type=float, required=False)
parser.add_argument("--height", help='model image input height resolution', type=int, default=480)
parser.add_argument("--width", help='model image input height resolution', type=int, default=640)
parser.add_argument("--batch_size", help='batch size to use during training', type=int, default=1)
parser.add_argument("--num_epochs", help='number of training epochs', type=int, default=1000)
parser.add_argument("--epoch_steps", help='training steps per epoch', type=int, default=1000)
parser.add_argument("--save_freq", help='model saving frequency per steps', type=int, default=1000)
parser.add_argument("--epoch_evals", help='number of epochs per evaluation', type=int, default=1)
parser.add_argument("--dataset_name", help="Name of the dataset being trained on",
default="FlyingThings3D", required=False)
parser.add_argument("--log_tensorboard", help="Logs results to tensorboard events files.", action="store_true")
parser.add_argument("--use_checkpoints",
help="Saves the weights using the tensorflow checkpoints format.",
action="store_true")
parser.add_argument("--sweep", help="Creates new output sub-folders for each sweep.", action="store_true")
parser.add_argument("--augment", help="Performs augmentation on the left and right images.", action="store_true")
args = parser.parse_args()
def main(args):
self_supervised = False
if args.train_disp_dir is None:
self_supervised = True
if args.sweep:
now = datetime.now()
current_time = now.strftime("%Y%m%dT%H%MZ")
args.output_dir = args.output_dir + f"/sweep-{current_time}"
log_dir = args.output_dir + "/logs"
save_extension = ".h5"
if args.use_checkpoints:
save_extension = ".ckpt"
# Initialize wandb with your project
run = wandb.init(project='madnet-keras',
sync_tensorboard=True,
config={
"learning_rate": args.lr,
"exponential_decay": args.decay,
"epochs": args.num_epochs,
"batch_size": args.batch_size,
"search_range": args.search_range,
"self_supervised_training": self_supervised,
"loss_function": "adam",
"architecture": "MADNet",
"dataset": args.dataset_name
})
config = wandb.config
perform_val = False
if args.val_left_dir is not None and args.val_right_dir is not None and args.val_disp_dir is not None:
perform_val = True
# Create output folder if it doesn't already exist
os.makedirs(args.output_dir, exist_ok=True)
# Initialise the model
model = MADNet(
input_shape=(args.height, args.width, 3),
weights=args.weights_path,
search_range=args.search_range
)
optimizer = tf.keras.optimizers.Adam(learning_rate=args.lr)
# If no train groundtruth is available, then the reprojection error
# from warping is used to calculate the loss
if self_supervised:
model.compile(
optimizer=optimizer,
loss=SSIMLoss(),
metrics=[EndPointError(), Bad3()],
run_eagerly=True if perform_val else False
)
else:
model.compile(
optimizer=optimizer,
loss=ReconstructionLoss(),
metrics=[EndPointError(), Bad3()],
run_eagerly=False
)
# Get dataset for training
train_dataset = StereoDatasetCreator(
left_dir=args.train_left_dir,
right_dir=args.train_right_dir,
batch_size=args.batch_size,
height=args.height,
width=args.width,
shuffle=args.shuffle,
disp_dir=args.train_disp_dir,
augment=args.augment
)
train_ds = train_dataset().repeat()
# Get datasets for training and callbacks
train_callback_dataset = StereoDatasetCreator(
left_dir=args.train_left_dir,
right_dir=args.train_right_dir,
batch_size=1,
height=args.height,
width=args.width,
shuffle=args.shuffle,
disp_dir=args.train_disp_dir,
augment=args.augment
)
train_callback_ds = train_callback_dataset().repeat()
val_ds = None
if perform_val:
val_dataset = StereoDatasetCreator(
left_dir=args.val_left_dir,
right_dir=args.val_right_dir,
batch_size=1,
height=args.height,
width=args.width,
shuffle=args.shuffle,
disp_dir=args.val_disp_dir
)
val_ds = val_dataset().repeat()
# Create callbacks
def scheduler(epoch, lr):
min_lr = args.min_lr
if epoch > 100:
# learning_rate * decay_rate ^ (global_step / decay_steps)
lr = lr * args.decay ** (epoch // 100)
lr = max(min_lr, lr)
tf.summary.scalar('learning rate', data=lr, step=epoch)
return lr
schedule_callback = tf.keras.callbacks.LearningRateScheduler(scheduler)
save_callback = tf.keras.callbacks.ModelCheckpoint(
filepath=args.output_dir + "/epoch-{epoch:04d}" + save_extension,
save_freq=args.save_freq,
save_weights_only=True,
verbose=0
)
wandb_callback = WandbCallback(
monitor="loss",
mode="min",
save_model=False, # Keep False, Hangs for a long time and doesn't finish the run
save_graph=False, # Keep False, Crashes script
save_weights_only=True # Keep True, full model is very large, 300MB. With just weights its 45MB.
)
wandb_images_callback = WandBImagesCallback(
training_data=train_callback_ds,
validation_data=val_ds,
val_epochs=args.epoch_evals
)
nan_callback = tf.keras.callbacks.TerminateOnNaN()
all_callbacks = [
save_callback,
schedule_callback,
wandb_callback,
wandb_images_callback,
nan_callback
]
if args.log_tensorboard:
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=log_dir,
histogram_freq=1,
write_steps_per_second=True,
update_freq="batch"
)
all_callbacks.append(tensorboard_callback)
tensorboard_images_callback = TensorboardImagesCallback(
training_data=train_callback_ds,
validation_data=val_ds,
val_epochs=args.epoch_evals
)
all_callbacks.append(tensorboard_images_callback)
# Fit the model
history = model.fit(
x=train_ds,
epochs=args.num_epochs,
verbose=1,
steps_per_epoch=args.epoch_steps,
callbacks=all_callbacks
)
run.finish()
if __name__ == "__main__":
main(args)