-
Notifications
You must be signed in to change notification settings - Fork 137
/
main.py
executable file
·195 lines (150 loc) · 6.9 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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
#!/usr/bin/python
#
# Copyright 2017 Robert Csordas. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# ==============================================================================
from Utils.ArgSave import *
import sys
import os
parser = StorableArgparse(description='RFCN trainer.')
parser.add_argument('-learningRate', type=float, default=0.0001, help='Learning rate')
parser.add_argument('-adamEps', type=float, default=1e-8, help='Adam epsilon')
parser.add_argument('-dataset', type=str, default="/data/Datasets/COCO", help="Path to COCO dataset")
parser.add_argument('-name', type=str, default="save", help="Directory to save checkpoints")
parser.add_argument('-saveInterval', type=int, default=10000, help='Save model for this amount of iterations')
parser.add_argument('-reportInterval', type=int, default=30, help='Repeat after this amount of iterations')
parser.add_argument('-displayInterval', type=int, default=60, help='Display after this amount of iterations')
parser.add_argument('-optimizer', type=str, default='adam', help='sgd/adam/rmsprop')
parser.add_argument('-resume', type=str, help='Resume from this file', save=False)
parser.add_argument('-report', type=str, default="", help='Create report here', save=False)
parser.add_argument('-trainFrom', type=str, default="-1", help='Train from this layer. Use 0 for all, -1 for just the added layers')
parser.add_argument('-hardMining', type=int, default=1, help="Enable hard example mining.")
parser.add_argument('-gpu', type=str, default="0", help='Train on this GPU(s)')
parser.add_argument('-mergeValidationSet', type=int, default=1, help='Merge validation set to training set.')
parser.add_argument('-profile', type=int, default=0, help='Enable profiling', save=False)
parser.add_argument('-randZoom', type=int, default=1, help='Enable box aware random zooming and cropping')
parser.add_argument('-freezeBatchNorm', type=int, default=1, help='Freeze batch normalization during finetuning.')
parser.add_argument('-export', type=str, help='Export model here.', save=False)
parser.add_argument('-cocoVariant', type=str, default="2014", help='Coco variant to load. 2014 or 2017')
opt=parser.parse_args()
if not os.path.isdir(opt.name):
os.makedirs(opt.name)
opt = parser.load(opt.name+"/args.json")
parser.save(opt.name+"/args.json")
if not os.path.isdir(opt.name+"/log"):
os.makedirs(opt.name+"/log")
if not os.path.isdir(opt.name+"/save"):
os.makedirs(opt.name+"/save")
if not os.path.isdir(opt.name+"/preview"):
os.makedirs(opt.name+"/preview")
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpu
import tensorflow as tf
from tensorflow.python.ops import control_flow_ops
from Dataset.CocoDataset import *
from Dataset.BoxLoader import *
from Utils.RunManager import *
from Utils.CheckpointLoader import *
from BoxInceptionResnet import *
from Dataset import Augment
from Visualize import VisualizeOutput
from Utils import Model
from Utils import Export
from tensorflow.python.client import timeline
import re
globalStep = tf.Variable(0, name='globalStep', trainable=False)
globalStepInc=tf.assign_add(globalStep,1)
Model.download()
dataset = BoxLoader()
dataset.add(CocoDataset(opt.dataset, randomZoom=opt.randZoom==1, set="train"+opt.cocoVariant))
if opt.mergeValidationSet==1:
dataset.add(CocoDataset(opt.dataset, set="val"+opt.cocoVariant))
images, boxes, classes = Augment.augment(*dataset.get())
print("Number of categories: "+str(dataset.categoryCount()))
print(dataset.getCaptionMap())
net = BoxInceptionResnet(images, dataset.categoryCount(), name="boxnet", trainFrom=opt.trainFrom, hardMining=opt.hardMining==1, freezeBatchNorm=opt.freezeBatchNorm==1)
tf.losses.add_loss(net.getLoss(boxes, classes))
def createUpdateOp(gradClip=1):
with tf.name_scope("optimizer"):
optimizer=tf.train.AdamOptimizer(learning_rate=opt.learningRate, epsilon=opt.adamEps)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
totalLoss = tf.losses.get_total_loss()
grads = optimizer.compute_gradients(totalLoss, var_list=net.getVariables())
if gradClip is not None:
cGrads = []
for g, v in grads:
if g is None:
print("WARNING: no grad for variable "+v.op.name)
continue
cGrads.append((tf.clip_by_value(g, -float(gradClip), float(gradClip)), v))
grads = cGrads
update_ops.append(optimizer.apply_gradients(grads))
return control_flow_ops.with_dependencies([tf.group(*update_ops)], totalLoss, name='train_op')
trainOp=createUpdateOp()
saver=tf.train.Saver(keep_checkpoint_every_n_hours=4, max_to_keep=100)
if opt.profile==1:
runOptions = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
runMetadata = tf.RunMetadata()
iterationsSinceStart=0
else:
runOptions=None
runMetadata=None
with tf.Session(config=tf.ConfigProto(intra_op_parallelism_threads=8)) as sess:
if not loadCheckpoint(sess, opt.name+"/save/", opt.resume):
print("Loading GoogleNet")
net.importWeights(sess, "./inception_resnet_v2_2016_08_30.ckpt")
#net.importWeights(sess, "initialWeights/", permutateRgb=False)
print("Done.")
if opt.export is not None:
Export.exportModel(sess, opt.export, [lambda name: name.split("/")[0]=="boxnet" and not re.match("^[Aa]dam(_.*)?$",name.split("/")[-1])])
sys.exit(-1)
dataset.startThreads(sess)
runManager = RunManager(sess, options=runOptions, run_metadata=runMetadata)
runManager.add("train", [globalStepInc,trainOp], modRun=1)
visualizer = VisualizeOutput.OutputVisualizer(opt, runManager, dataset, net, images, boxes, classes)
i=1
cycleCnt=0
lossSum=0
while True:
#run various parts of the network
res = runManager.modRun(i)
if opt.profile==1:
print("Profiling step %d" % iterationsSinceStart)
iterationsSinceStart+=1
if iterationsSinceStart==5:
print("Writing profile data...")
tl = timeline.Timeline(runMetadata.step_stats)
ctf = tl.generate_chrome_trace_format()
with open('timeline.json', 'w') as f:
f.write(ctf)
print("Done.")
sys.exit(0)
i, loss=res["train"]
lossSum+=loss
cycleCnt+=1
visualizer.draw(res)
if i % opt.reportInterval == 0:
if cycleCnt>0:
loss=lossSum/cycleCnt
# lossS=sess.run(trainLossSum, feed_dict={
# trainLossFeed: loss
# })
# log.add_summary(lossS, global_step=samplesSeen)
epoch="%.2f" % (float(i) / dataset.count())
print("Iteration "+str(i)+" (epoch: "+epoch+"): loss: "+str(loss))
lossSum=0
cycleCnt=0
if i % opt.saveInterval == 0:
print("Saving checkpoint "+str(i))
saver.save(sess, opt.name+"/save/model_"+str(i), write_meta_graph=False)