-
Notifications
You must be signed in to change notification settings - Fork 23
/
Copy pathG1D19_ResHB_ResHB_1LSTM.py
203 lines (166 loc) · 8.22 KB
/
G1D19_ResHB_ResHB_1LSTM.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
import tensorflow as tf
import os
CUDA_VISIBLE_DEVICES=0
os.environ["CUDA_VISIBLE_DEVICES"] = "0" # set gpu number
from src.net.NetBase import *
from src.layers.LayerHelper import *
from src.layers.BasicLayers import *
from src.layers.ResidualLayers import *
from src.layers.RNN import *
import settings.LayerSettings as layerSettings
import settings.DataSettings as dataSettings
import numpy as np
DARKNET19_MODEL_PATH = 'data/pretrainModels/darknet19/darknet19.pb'
class Net(NetworkBase):
def __init__(self, inputImage_, batchSize_, unrolledSize_, isTraining_, trainingStep_):
self._inputImage = inputImage_
self._batchSize = batchSize_
self._unrolledSize = unrolledSize_
self._isTraining = isTraining_
self._trainingStep = trainingStep_
self._DROPOUT_PROB = 0.5
self._NUMBER_OF_NEURONS_IN_LSTM = 1024
self._dictOfInterestedActivations = {}
if dataSettings.GROUPED_SIZE != 1:
errorMessage = __name__ + " only take GROUPED_SIZE = 1;\n"
errorMessage += "However, DataSettings.GROUPED_SIZE = " + str(dataSettings.GROUPED_SIZE)
raise ValueError(errorMessage)
def Build(self):
darknet19_GraphDef = tf.GraphDef()
'''
The CNN only take input shape [..., w, h, c]. Thus, move the UNROLLED_SIZE dimension
to merged with BATCH_SIZE, and form the shape: [b*u, w, h, c].
'''
convInput = tf.reshape(self._inputImage, [-1,
dataSettings.IMAGE_SIZE, dataSettings.IMAGE_SIZE, dataSettings.IMAGE_CHANNELS])
with tf.name_scope("DarkNet19"):
with open(DARKNET19_MODEL_PATH, 'rb') as modelFile:
darknet19_GraphDef.ParseFromString(modelFile.read())
listOfOperations = tf.import_graph_def(darknet19_GraphDef,
input_map={"input": convInput},
# return_elements=["2-maxpool", "32-leaky"])
return_elements=["2-maxpool", "36-leaky"])
pool2 = listOfOperations[0].outputs[0]
lastOp = listOfOperations[-1]
out = lastOp.outputs[0]
out, updateOp0 = BatchNormalization('BN_0', out, isConvLayer_=True,
isTraining_=self._isTraining, currentStep_=self._trainingStep)
opticalFlowOut, updateOF = self._buildOpticalFlowNet(pool2)
print("opticalFlowOut.shape = ", opticalFlowOut.shape) # shape = [b*u, 1024]
with tf.name_scope("CNN"):
'''
The input shape = [b, u, g, w, h, c]
after Conv, shape = [b*u*g, w', h', c']
here, decouple the Group dimension, shape = [b*u, g * w' * h' * c']
'''
print("In CNN:")
print("\t darknetOutput.shape = ", out.shape) # shape = [b*u*1, 7, 7, 1024]
out, updateOp1 = ResidualHeadBlock('ResHeadBlock', out, [128, 64, 128], isTraining_=self._isTraining,
trainingStep_=self._trainingStep, activationType_="LEAKY_RELU", isTrainable_=True)
print("\t before Fc, out.shape = ", out.shape)
out = tf.cond(self._isTraining, lambda: tf.nn.dropout(out, self._DROPOUT_PROB), lambda: out)
out = FullyConnectedLayer('Fc1', out, numberOfOutputs_=1024)
out, updateOp2 = BatchNormalization('BN_4', out, isConvLayer_=False,
isTraining_=self._isTraining, currentStep_=self._trainingStep)
self._dictOfInterestedActivations['CNN'] = out
print("\t Fc out.shape = ", out.shape)
with tf.name_scope("Concat"):
out = tf.concat([out, opticalFlowOut], axis=1)
print("In Concat:")
print("\t before Fc, out.shape = ", out.shape)
out = FullyConnectedLayer('Fc', out, numberOfOutputs_=1024)
out, updateOp3 = BatchNormalization('BN_5', out, isConvLayer_=False,
isTraining_=self._isTraining, currentStep_=self._trainingStep)
'''
Note: For tf.nn.rnn_cell.dynamic_rnn(), the input shape of [1:] must be explicit.
i.e., one Can't Reshape the out by:
out = tf.reshape(out, [BATCH_SIZE, UNROLLED_SIZE, -1])
since '-1' is implicit dimension.
'''
featuresShapeInOneBatch = out.shape[1:].as_list()
targetShape = [self._batchSize, self._unrolledSize] + featuresShapeInOneBatch
out = tf.reshape(out, targetShape)
self._dictOfInterestedActivations['Concat'] = out
print("before LSTM, shape = ", out.shape)
out, self._stateTensorOfLSTM_1, self._statePlaceHolderOfLSTM_1 = LSTM( "LSTM_1",
out,
self._NUMBER_OF_NEURONS_IN_LSTM,
isTraining_=self._isTraining,
dropoutProb_=self._DROPOUT_PROB)
self._dictOfInterestedActivations['LSTM'] = out
with tf.name_scope("Fc_Final"):
featuresShapeInOneBatch = out.shape[2:].as_list()
targetShape = [self._batchSize * self._unrolledSize] + featuresShapeInOneBatch
out = tf.reshape(out, targetShape)
out = FullyConnectedLayer('Fc3', out, numberOfOutputs_=dataSettings.NUMBER_OF_CATEGORIES)
self._logits = tf.reshape(out, [self._batchSize, self._unrolledSize, -1])
self._updateOp = tf.group(updateOF, updateOp0, updateOp1, updateOp2, updateOp3)
print()
@property
def logitsOp(self):
return self._logits
@property
def updateOp(self):
return self._updateOp
def GetListOfStatesTensorInLSTMs(self):
'''
You should Not Only sess.run() the net.logits, but also this listOfTensors
to get the States of LSTM. And assign it to PlaceHolder next time.
ex:
>> tupleOfResults = sess.run( [out] + net.GetListOfStatesTensorInLSTMs(), ...)
>> listOfResults = list(tupleOfResults)
>> output = listOfResults.pop(0)
>> listOfStates = listOfResults
See GetFeedDictOfLSTM() method as well
'''
return [self._stateTensorOfLSTM_1]
def GetFeedDictOfLSTM(self, BATCH_SIZE_, listOfPreviousStateValues_=None):
'''
This function will return a dictionary that contained the PlaceHolder-Value map
of the LSTM states.
You can use this function as follows:
>> feed_dict = { netInput : batchOfImages }
>> feedDictOFLSTM = net.GetLSTM_Feed_Dict(BATCH_SIZE, listOfPreviousStateValues)
>> tupleOfOutputs = sess.run( [out] + net.GetListOfStatesTensorInLSTMs(),
feed_dict = feed_dict.update(feedDictOFLSTM) )
>> listOfOutputs = list(tupleOfOutputs)
>> output = listOfOutputs.pop(0)
>> listOfPreviousStateValues = listOfOutputs.pop(0)
'''
if listOfPreviousStateValues_ == None:
'''
For the first time (or, the first of Unrolls), there's no previous state,
return zeros state.
'''
initialCellState = tuple( [np.zeros([BATCH_SIZE_, self._NUMBER_OF_NEURONS_IN_LSTM])] * 2 )
initialCellState = tf.nn.rnn_cell.LSTMStateTuple(initialCellState[0], initialCellState[1])
return {self._statePlaceHolderOfLSTM_1 : initialCellState }
else:
if len(listOfPreviousStateValues_) != 1:
errorMessage = "len(listOfPreviousStateValues_) = " + str( len(listOfPreviousStateValues_) )
errorMessage += "; However, the expected lenght is 1.\n"
errorMessage += "\t Do you change the Network Structure, such as Add New LSTM?\n"
errorMessage += "\t Or, do you add more tensor to session.run()?\n"
return { self._statePlaceHolderOfLSTM_1 : listOfPreviousStateValues_[0] }
def _buildOpticalFlowNet(self, inputTensor_):
with tf.name_scope("OpticalFLow"):
print("In OpticalFlow:")
print("\t before Conv2, out.shape = ", inputTensor_.shape) # shape = [b*u*1, 112, 112, 32]
out = ConvLayer('Conv2', inputTensor_, filterSize_=3, numberOfFilters_=64,
stride_=1, padding_='SAME', isTrainable_=True)
out, updateOp1 = BatchNormalization('BN2', out, isConvLayer_=True, isTraining_=self._isTraining,
currentStep_=self._trainingStep, isTrainable_=True)
out = LeakyRELU('RELU2', out)
out = MaxPoolLayer('Pool2', out, kernelSize_=2, stride_=2, padding_='SAME')
out, updateOp2 = ResidualBlock('ResBlock', out, [128, 128, 64], isTraining_=self._isTraining,
trainingStep_=self._trainingStep, activationType_="LEAKY_RELU", isTrainable_=True)
out = MaxPoolLayer('Pool5', out, kernelSize_=2, stride_=2, padding_='SAME')
print("\t before Fc, out.shape = ", out.shape) # shape = [b*u, 28, 28, 128]
out = tf.cond(self._isTraining, lambda: tf.nn.dropout(out, self._DROPOUT_PROB), lambda: out)
out = FullyConnectedLayer('Fc_of', out, numberOfOutputs_=1024)
out, updateOp3 = BatchNormalization('BN_of', out, isConvLayer_=False,
isTraining_=self._isTraining, currentStep_=self._trainingStep)
self._dictOfInterestedActivations['OpticalFlow'] = out
print("\t Fc final.shape = ", out.shape)
updateOp = tf.group(updateOp1, updateOp2, updateOp3)
return out, updateOp