作者搜集了网上存在的利用Tensorflow计算FLOPs的代码（如下所示），但是几乎都不能准确计算，因此我们尝试通过人工计算。
CNN 模型所需的计算力（flops）和参数（parameters）数量是怎么计算的？
卷积层计算量(FLOPS)和参数量的计算
MACs和FLOPs
def get_flops_params():
sess = tf.Session()
graph = sess.graph
flops = tf.profiler.profile(graph, options=tf.profiler.ProfileOptionBuilder.float_operation())
params = tf.profiler.profile(graph, options=tf.profiler.ProfileOptionBuilder.trainable_variables_parameter())
print('FLOPs: {}; Trainable params: {}'.format(flops.total_float_ops, params.total_parameters))
文章目录
- 1.SCNN
- 1.1.Parameters计算
- 1.2.FLOPs计算
- 2.PET-CGDNN
def SCNN():
# build the CNN model
in_shp = [2, 128]
L = 128 # sample points
xm_input = Input(in_shp)
xm = Reshape([128, 2], input_shape=in_shp)(xm_input)
x1 = Conv1D(128, 16, activation='relu', padding='same', input_shape=[L, 2])(xm)
x2 = BatchNormalization()(x1)
x3 = Dropout(0.5)(x2)
x4 = SeparableConv1D(64, 8, activation='relu', padding='same')(x3)
x5 = BatchNormalization()(x4)
x6 = Dropout(0.5)(x5)
x7 = Flatten()(x6)
x8 = Dense(10)(x7)
predicts = Activation('softmax')(x8)
model = Model(xm_input, predicts)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
return model
输出每一层具体信息:
_________________________________________________________________ Layer (type) Output Shape Param # ================================================================= input_1 (InputLayer) (None, 2, 128) 0 _________________________________________________________________ reshape_1 (Reshape) (None, 128, 2) 0 _________________________________________________________________ conv1d_1 (Conv1D) (None, 128, 128) 4224 _________________________________________________________________ batch_normalization_1 (Batch (None, 128, 128) 512 _________________________________________________________________ dropout_1 (Dropout) (None, 128, 128) 0 _________________________________________________________________ separable_conv1d_1 (Separabl (None, 128, 64) 9280 _________________________________________________________________ batch_normalization_2 (Batch (None, 128, 64) 256 _________________________________________________________________ dropout_2 (Dropout) (None, 128, 64) 0 _________________________________________________________________ flatten_1 (Flatten) (None, 8192) 0 _________________________________________________________________ dense_1 (Dense) (None, 11) 90123 _________________________________________________________________ activation_1 (Activation) (None, 11) 0 ================================================================= Total params: 104,395 Trainable params: 104,011 Non-trainable params: 384 _________________________________________________________________
- 卷积层：
以第一个conv1d_1 (Conv1D)为例，
P C O N V = ( K h ∗ K w ∗ C i n ) ∗ C o u t + C o u t = ( 16 ∗ 2 ) ∗ 128 + 128 = 4224 P_{CONV} = (K_{h} * K_{w} * C_{in}) * C_{out} + C_{out}= (16*2)*128+128 = 4224 PCONV=(Kh∗Kw∗Cin)∗Cout+Cout=(16∗2)∗128+128=4224 C i n = 2 , C o u t = 128 , K h ∗ K w = 16 C_{in}=2, C_{out} = 128, K_{h} * K_{w} = 16 Cin=2,Cout=128,Kh∗Kw=16
- 卷积层：
以第一个conv1d_1 (Conv1D)为例（注意：不考虑bias时有-1，有bias时没有-1，这里的计算考虑了bias)
F C O N V = ( 2 ∗ C i n ∗ K h ∗ K w − 1 ） ∗ H ∗ W ∗ C o u t = ( 2 ∗ 2 ∗ 16 − 1 ) ∗ 128 ∗ 128 = 1 , 032 , 192 F_{CONV} = (2* C_{in} *K_{h} * K_{w}-1） * H*W*C_{out}= (2*2*16-1)*128*128=1,032,192 FCONV=(2∗Cin∗Kh∗Kw−1）∗H∗W∗Cout=(2∗2∗16−1)∗128∗128=1,032,192 - 可分离卷积，参考：该捋清！卷积、可分离卷积的参数和FLOPs计算！
F S e p C O N V = 2 ( K h ∗ K w + C o u t ) ∗ H ∗ W ∗ C i n = 2 ( 8 + 64 ) ∗ 128 ∗ 128 = 2 , 359 , 296 F_{SepCONV} = 2(K_{h} * K_{w}+ C_{out})* H*W*C_{in}= 2(8+64)*128*128=2,359,296 FSepCONV=2(Kh∗Kw+Cout)∗H∗W∗Cin=2(8+64)∗128∗128=2,359,296 - 全连接dense层，参考神经网络层的FLOPs计算,
F D e n s e = 2 ∗ H ∗ W ∗ C i n ∗ C o u t = 2 ∗ 8192 ∗ 11 = 180 , 224 F_{Dense} = 2* H*W*C_{in}*C_{out}= 2*8192*11=180,224 FDense=2∗H∗W∗Cin∗Cout=2∗8192∗11=180,224
F C O N V + F S e p C O N V + F D e n s e = 3 , 571 , 712 F_{CONV}+F_{SepCONV} +F_{Dense}=3,571,712 FCONV+FSepCONV+FDense=3,571,712
def cal1(x):
y = tf.keras.backend.cos(x)
return y
def cal2(x):
y = tf.keras.backend.sin(x)
return y
def MCLDNN(weights=None,
input_shape1=[2, 128],
input_shape2=[128, 1],
classes=11,
**kwargs):
if weights is not None and not (os.path.exists(weights)):
raise ValueError('The `weights` argument should be either '
'`None` (random initialization), '
'or the path to the weights file to be loaded.')
dr = 0.5 # dropout rate (%)
input = Input(input_shape1 + [1], name='input1')
input1 = Input(input_shape2, name='input2')
input2 = Input(input_shape2, name='input3')
x1 = Flatten()(input)
x1 = Dense(1, name='fc2')(x1)
x1 = Activation('linear')(x1)
cos1 = Lambda(cal1)(x1)
sin1 = Lambda(cal2)(x1)
x11 = Multiply()([input1, cos1])
x12 = Multiply()([input2, sin1])
x21 = Multiply()([input2, cos1])
x22 = Multiply()([input1, sin1])
y1 = Add()([x11, x12])
y2 = Subtract()([x21, x22])
y1 = Reshape(target_shape=(128, 1), name='reshape1')(y1)
y2 = Reshape(target_shape=(128, 1), name='reshape2')(y2)
x11 = concatenate([y1, y2])
x3 = Reshape(target_shape=((128, 2, 1)), name='reshape3')(x11)
# spatial feature
x3 = Conv2D(75, (8, 2), padding='valid', activation="relu", name="conv1_1", kernel_initializer='glorot_uniform')(
x3)
x3 = Conv2D(25, (5, 1), padding='valid', activation="relu", name="conv1_2", kernel_initializer='glorot_uniform')(
x3)
# temporal feature
x4 = Reshape(target_shape=((117, 25)), name='reshape4')(x3)
x4 = keras.layers.GRU(units=128)(x4)
# x4 = Flatten()(x4)
#
x = Dense(classes, activation='softmax', name='softmax')(x4)
model = Model(inputs=[input, input1, input2], outputs=x)
输出每一层具体信息:
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input1 (InputLayer) (None, 2, 128, 1) 0
__________________________________________________________________________________________________
flatten_1 (Flatten) (None, 256) 0 input1[0][0]
__________________________________________________________________________________________________
fc2 (Dense) (None, 1) 257 flatten_1[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 1) 0 fc2[0][0]
__________________________________________________________________________________________________
input2 (InputLayer) (None, 128, 1) 0
__________________________________________________________________________________________________
lambda_1 (Lambda) (None, 1) 0 activation_1[0][0]
__________________________________________________________________________________________________
input3 (InputLayer) (None, 128, 1) 0
__________________________________________________________________________________________________
lambda_2 (Lambda) (None, 1) 0 activation_1[0][0]
__________________________________________________________________________________________________
multiply_1 (Multiply) (None, 128, 1) 0 input2[0][0]
lambda_1[0][0]
__________________________________________________________________________________________________
multiply_2 (Multiply) (None, 128, 1) 0 input3[0][0]
lambda_2[0][0]
__________________________________________________________________________________________________
multiply_3 (Multiply) (None, 128, 1) 0 input3[0][0]
lambda_1[0][0]
__________________________________________________________________________________________________
multiply_4 (Multiply) (None, 128, 1) 0 input2[0][0]
lambda_2[0][0]
__________________________________________________________________________________________________
add_1 (Add) (None, 128, 1) 0 multiply_1[0][0]
multiply_2[0][0]
__________________________________________________________________________________________________
subtract_1 (Subtract) (None, 128, 1) 0 multiply_3[0][0]
multiply_4[0][0]
__________________________________________________________________________________________________
reshape1 (Reshape) (None, 128, 1) 0 add_1[0][0]
__________________________________________________________________________________________________
reshape2 (Reshape) (None, 128, 1) 0 subtract_1[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 128, 2) 0 reshape1[0][0]
reshape2[0][0]
__________________________________________________________________________________________________
reshape3 (Reshape) (None, 128, 2, 1) 0 concatenate_1[0][0]
__________________________________________________________________________________________________
conv1_1 (Conv2D) (None, 121, 1, 75) 1275 reshape3[0][0]
__________________________________________________________________________________________________
conv1_2 (Conv2D) (None, 117, 1, 25) 9400 conv1_1[0][0]
__________________________________________________________________________________________________
reshape4 (Reshape) (None, 117, 25) 0 conv1_2[0][0]
__________________________________________________________________________________________________
gru_1 (GRU) (None, 128) 59136 reshape4[0][0]
__________________________________________________________________________________________________
softmax (Dense) (None, 11) 1419 gru_1[0][0]
==================================================================================================
Total params: 71,487
Trainable params: 71,487
Non-trainable params: 0
F
f
c
2
=
2
∗
H
∗
W
∗
C
i
n
∗
C
o
u
t
=
2
∗
256
∗
1
=
512
F_{fc2} = 2* H*W*C_{in}*C_{out}= 2*256*1=512
Ffc2=2∗H∗W∗Cin∗Cout=2∗256∗1=512
F
c
o
n
v
1
1
=
(
2
∗
C
i
n
∗
K
h
∗
K
w
−
1
）
∗
H
∗
W
∗
C
o
u
t
=
(
2
∗
1
∗
8
∗
2
−
1
)
∗
121
∗
1
∗
75
=
281
,
325
F_{conv1_1} = (2* C_{in} *K_{h} * K_{w}-1） * H*W*C_{out}= (2*1*8*2-1)*121*1*75=281,325
Fconv11=(2∗Cin∗Kh∗Kw−1）∗H∗W∗Cout=(2∗1∗8∗2−1)∗121∗1∗75=281,325
F
c
o
n
v
1
2
=
(
2
∗
C
i
n
∗
K
h
∗
K
w
−
1
）
∗
H
∗
W
∗
C
o
u
t
=
(
2
∗
75
∗
5
∗
1
−
1
)
∗
117
∗
1
∗
25
=
2
,
190
,
825
F_{conv1_2} = (2* C_{in} *K_{h} * K_{w}-1） * H*W*C_{out}= (2*75*5*1-1)*117*1*25=2,190,825
Fconv12=(2∗Cin∗Kh∗Kw−1）∗H∗W∗Cout=(2∗75∗5∗1−1)∗117∗1∗25=2,190,825
注：
1.我们如果用tensorflow作为backend时候，系统会默认是channel_last。
2.如何计算GRU卷积神经网络CNN中的参数量（parameters）和计算量（FLOPs ）这篇文章里面的结论直接用：FLOPs = paras * H * W * 2
我们以上面的计算结果为例
F
f
c
2
=
p
a
r
a
s
∗
H
∗
W
∗
2
=
257
∗
1
∗
2
=
514
F_{fc2} =paras * H * W* 2=257*1*2=514
Ffc2=paras∗H∗W∗2=257∗1∗2=514
514和512差别不大。
F
c
o
n
v
1
1
=
p
a
r
a
s
∗
H
∗
W
∗
2
=
1275
∗
121
∗
2
=
308
,
550
F_{conv1_1} =paras * H * W* 2 = 1275 * 121*2 = 308,550
Fconv11=paras∗H∗W∗2=1275∗121∗2=308,550
308,550和281,325差别不大。
F
c
o
n
v
1
2
=
p
a
r
a
s
∗
H
∗
W
∗
2
=
9400
∗
117
∗
2
=
2
,
199
,
600
F_{conv1_2} =paras * H * W* 2 = 9400 * 117*2 = 2,199,600
Fconv12=paras∗H∗W∗2=9400∗117∗2=2,199,600
2,199,600和2,190,825差别不大。
因此虽然没有找到GRU的FLOPs计算公式，我们也尝试用这个公式计算
F
g
r
u
1
=
p
a
r
a
s
∗
H
∗
W
∗
2
=
59136
∗
2
=
118
,
272
F_{gru_1} = paras * H * W* 2 = 59136 * 2 = 118,272
Fgru1=paras∗H∗W∗2=59136∗2=118,272
于是得到总的
F
f
c
2
+
F
c
o
n
v
1
1
+
F
c
o
n
v
1
2
+
F
g
r
u
1
=
514
+
308
,
550
+
2
,
199
,
600
+
118
,
272
=
2
,
626
,
936
F_{fc2} + F_{conv1_1}+F_{conv1_2}+F_{gru_1} = 514+ 308,550+2,199,600+118,272=2,626,936
Ffc2+Fconv11+Fconv12+Fgru1=514+308,550+2,199,600+118,272=2,626,936
