-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
285 lines (233 loc) · 10.5 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
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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
# imports
import os
import sys
import larq as lq
import tensorflow.keras as keras
import tensorflow.keras.backend as K
from tensorflow.keras.optimizers import Adam, SGD
import tensorflow.keras.preprocessing.image
from tensorflow.keras.callbacks import LearningRateScheduler
from tensorflow.keras.callbacks import ReduceLROnPlateau
import tensorflow as tf
import cv2
from datetime import date, timedelta
import time
from six import raise_from
import argparse
# import custom modules
from utils.processing_tools import *
from utils.generators import *
from utils.augmentor.misc import MiscEffect
from utils.augmentor.color import VisualEffect
# remove the try in final version
import pyximport
pyximport.install(reload_support=True)
from utils.evaluate import *
def check_args(parsed_args):
"""
Function to check for inherent contradictions within parsed arguments.
For example, epochs < 1
Intended to raise errors prior to backend initialisation.
Args
parsed_args: parser.parse_args()
Returns
parsed_args
"""
if parsed_args.epochs < 1 :
raise ValueError(
"Epochs must be equal or greater than one (received: {}) ".format(parsed_args.epochs))
arch_path = os.path.join('architectures', parsed_args.architecture + '.py')
if os.path.exists(arch_path):
pass
else:
raise ValueError(
"Architecture is not available on architectures folder (received: {}) ".format(parsed_args.architecture))
if os.path.exists(parsed_args.dataset_path):
# check if pascalvoc exists
pass
else:
raise ValueError(
"Dataset folder does not exist(received: {}) ".format(parsed_args.dataset_path))
return parsed_args
def parse_args(args):
"""
Parse the arguments.
"""
parser = argparse.ArgumentParser(description='Script for training a binary centernet network.')
parser.add_argument('--architecture', help='Architecture that will be used to train the network.', type=str)
parser.add_argument('--batch_size', help='Size of the batches.', default=32, type=int)
parser.add_argument('--num_classes', help='Number of classes (20 for PascalVOC).', default=20, type=int)
parser.add_argument('--input_size', help='Size of the input (eg.: 512, 256).', default=512, type=int)
parser.add_argument('--gpu', help='Id of the GPU to use (as reported by nvidia-smi).')
parser.add_argument('--epochs', help='Number of epochs to train.', type=int, default=200)
parser.add_argument('--train-verbose', help='Tensorflow .fit verbose: 0 = silent, 1 = progress bar, 2 = only epochs.', default=2, type=int)
parser.add_argument('--no-snapshots', help='Disable saving snapshots.', dest='snapshots', action='store_false')
parser.add_argument('--epoch-evaluation', help='Disable per epoch evaluation.', dest='epoch_evaluation',
action='store_false')
parser.add_argument('--final-evaluation', help='Disable final model evaluation.', dest='final_evaluation',
action='store_false')
parser.add_argument('--data-augmentation', help='Enable/disable data augmentation.', default=True, type=bool)
parser.add_argument('--model-summary', help='Disable model summary.', dest='model_summary', action='store_false')
parser.add_argument('--dataset-path', help='Set the dataset path.', default=os.path.join('datasets', 'PascalVOC-OD-2007-2012'), type=str)
parser.add_argument('--compute-val-loss', help='Disable validation loss during training', dest='compute_val_loss',
action='store_false')
parser.add_argument('--val-sampling', help='Set the fraction relative to all data to be sampled at each epoch when calculating the validation mAP during training (1 uses all val set)', type=float, default=0.25)
parser.add_argument('--train-sampling', help='Set the fraction relative to all data to be sampled at each epoch when calculating the train mAP during training (1 uses all train set)', type=float, default=0.05)
print(vars(parser.parse_args(args)))
return check_args(parser.parse_args(args))
def main(args=None):
# parse arguments
if args is None:
args = sys.argv[1:]
args = parse_args(args)
num_classes = args.num_classes
input_size = (args.input_size,args.input_size)
#assert input_size[0] == input_size[1], "Input shape must be the same"
batch_size = args.batch_size
# optionally choose specific GPU
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
# CHECK GPU
if tf.test.gpu_device_name():
print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))
else:
print('problem loading gpu')
import importlib
sys.path.append('architectures')
#import architecture module
architecture_module = importlib.import_module(args.architecture)
# set architecture name
architecture_name = args.architecture
# create model given an architecture
# (the centernet input must be specified by the architecture file)
model, prediction_model, debug_model = architecture_module.centernet(input_size = input_size, num_classes = num_classes)
# get resnet stem weights if architecture is correct
if architecture_name == 'qn_centernet_resnet50_stem':
def copy_resnet_weights(resnet50, centernet, verbose = False):
n_layers_stem = 19
for layer_id in range(n_layers_stem):
#print(layer.get_config(), layer.get_weights())
if verbose:
print('Copying layer weights from resnet50 to stem index = {}'.format(layer_id))
w_copy = resnet50.layers[layer_id].get_weights()
centernet.layers[layer_id].set_weights(w_copy)
centernet.layers[layer_id].trainable = False # Freeze the layer
if verbose:
try:
if tf.math.equal(centernet.layers[layer_id].get_weights(), resnet50.layers[layer_id].get_weights()):
print('Success...')
except:
print('Something happened. Check.')
input_size=(512,512)
image_input = tensorflow.keras.layers.Input(shape=(input_size[0], input_size[1], 3))
resnet50 = tf.keras.applications.resnet50.ResNet50(
include_top=False,
weights='imagenet',
input_tensor=image_input,
input_shape=None,
pooling=None,
)
copy_resnet_weights(resnet50, model, verbose = True)
if args.model_summary:
lq.models.summary(model)
# Loading Data
# create random transform objects for augmenting training data
if args.data_augmentation:
misc_effect = MiscEffect(border_value=0)
visual_effect = VisualEffect()
else:
misc_effect = None
visual_effect = None
multi_scale = True
validation_generator = PascalVocGenerator(
args.dataset_path,
# val of PascalVOC is the test of VOC2007
'val',
skip_difficult=True,
shuffle_groups=False,
input_size = input_size[0],
batch_size = batch_size
)
train_generator = PascalVocGenerator(
args.dataset_path,
# train of PascalVOC is the union of trainval VOC2007 and trainval VOC2012
'train',
skip_difficult=True,
multi_scale=multi_scale,
misc_effect=misc_effect,
visual_effect=visual_effect,
input_size = input_size[0],
batch_size = batch_size
)
# TRAINING:
# create callbacks
callbacks = []
def lr_schedule(epoch, lr):
"""Learning Rate Schedule
Learning rate is scheduled to increase after 100 epochs back to 1e-4.
Called automatically every epoch as part of callbacks during training.
# Arguments
epoch (int): The number of epochs
lr (float32): learning rate
# Returns
lr (float32): learning rate
"""
if epoch == 100:
lr = 1e-4
else:
pass
print('Learning rate: ', lr)
return lr
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(
monitor='loss',
factor=0.1,
patience=4,
verbose=1,
mode='auto',
min_delta=0.0001,
cooldown=0,
min_lr=1e-7
)
if args.epoch_evaluation:
evaluation = Evaluate(validation_generator, prediction_model, model_name = architecture_name,
sample_ratio = args.val_sampling, verbose = 2,
save_best=True, save_thr=0.005,
save_metric = True, val_type = 'val')
evaluation_train = Evaluate(train_generator, prediction_model, model_name = architecture_name,
sample_ratio = args.train_sampling, verbose = 2,
save_metric = True, val_type = 'train')
save_loss = Save_Loss(model_name = architecture_name)
callbacks.append(evaluation)
callbacks.append(evaluation_train)
callbacks.append(save_loss)
callbacks.append(lr_reducer)
callbacks.append(lr_scheduler)
# compile model
opt = tf.keras.optimizers.Adam(learning_rate=0.001, decay=0, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
model.compile(optimizer=opt, loss={'centernet_loss': lambda y_true, y_pred: y_pred})
history = model.fit(
train_generator,
epochs=args.epochs,
verbose=args.train_verbose,
callbacks=callbacks
)
if args.final_evaluation:
print(history.history)
average_precisions = evaluate(validation_generator, prediction_model,
flip_test=False,
keep_resolution=False,
score_threshold=0.01)
# compute per class average precision
total_instances = []
precisions = []
for label, (average_precision, num_annotations) in average_precisions.items():
print('{:.0f} instances of class'.format(num_annotations), validation_generator.label_to_name(label),
'with average precision: {:.4f}'.format(average_precision))
total_instances.append(num_annotations)
precisions.append(average_precision)
mean_ap = sum(precisions) / sum(x > 0 for x in total_instances)
print('mAP: {:.4f}'.format(mean_ap))
return
if __name__ == '__main__':
main()