要求
分成应用分类(14类)和流量分类(12类)两个任务,按照运算量,
,
三个档次分别在两个任务上调试3个模型出来
1.基础知识回顾
w_conv1 = weight_variable(name='w_conv1', shape=[1, 3, 1, 32])
b_conv1 = bias_variable([32])
shape:shape为4维tensor,各参数含义:[width, height, channels, kernel_nums]
设input:[batch,length_in,width_in,in_channels_in]
w_conv1中shape是filter的shape:[filter_length,filter_width,channels_in,channels_out]
例如:
input.shape = [5,256,256,32] #5张256*256的图片,输入32
filter.shape = [3,3,32,64] #3*3卷积核,输入32,输出64
output.shape = [5,out_length,out_width,64] #卷积后的结果是5张(卷积后图片尺寸)图片,进入下一个卷积层的通道数为64
2.模型参数设置
通道数没有改变,改变卷积核的尺寸
输出结果会有影响吗?
NIN_1
NIN_14类调参结果
1.14类1500字节结果:32/64/128
改变strides,32/64/128结果:
(1, 1, 2, 128)
(1, 14)
Parsing Inputs...
- FLOPs:
(2953148) Trainable params:
60078)
-
test:96.57% valid:96.39%
2.14类1500字节:16/32/64
(1, 1, 2, 64)
(1, 14)
Parsing Inputs...
- FLOPs:
(819532); Trainable params:
(20030)
-
test:92.66% valid:94.45%
3.12类1500字节结果:32/64/128
改变strides,32/64/128结果:
(1, 1, 2, 128)
(1, 12)
Parsing Inputs...
- FLOPs:
(2952250); Trainable params:
(59948)
- valid_accuracy:97.02%
4.12类1500字节:16/32/64
(1, 1, 2, 128)
(1, 12)
Parsing Inputs...
======================End of Report==========================
- FLOPs:
(818634); Trainable params:
(19900)
网友评论