问题描述
我必须复制给定信息为(层,输入形状,过滤器,核,步幅)的模型摘要
LAYER INPUT SHAPE FILTER KERNEL STRIDE
Fc1 [batch,128] - - -
Reshape [batch,6x4x256] - - -
DeConv1 [batch,6,4,256] 128 (5,5,256) (2,2)
DeConv2 [batch,11,8,128] 128 (3,3,128) (2,1)
DeConv3 [batch,22,128] 64 (3,128) (1,1)
....
我设法提出了以下建议:
model.add(Dense(6*4*256,activation="relu",input_dim=128))
model.add(Reshape((6,256)))
model.add(Conv2DTranspose(128,kernel_size=(5,5),strides=(2,2),padding="same"))
model.add(Batchnormalization())
model.add(Activation("relu"))
model.add(Conv2DTranspose(128,kernel_size=(3,3),1),padding="same"))
model.add(Batchnormalization())
model.add(Activation("relu"))
model.add(Conv2DTranspose(64,strides=(1,padding="same"))
model.add(Batchnormalization())
model.add(Activation("relu"))
....
但这会产生:
Layer (type) Output Shape Param #
=================================================================
dense_58 (Dense) (None,6144) 792576
_________________________________________________________________
reshape_44 (Reshape) (None,256) 0
_________________________________________________________________
conv2d_transpose_155 (Conv2D (None,12,128) 819328
_________________________________________________________________
batch_normalization_197 (Bat (None,128) 512
_________________________________________________________________
activation_190 (Activation) (None,128) 0
_________________________________________________________________
conv2d_transpose_156 (Conv2D (None,24,128) 147584
_________________________________________________________________
batch_normalization_198 (Bat (None,128) 512
_________________________________________________________________
activation_191 (Activation) (None,128) 0
_________________________________________________________________
...
因此,我得到的不是(12,128),而是第一个DeConv层的(11,128)。我尝试对here和here给出的输出形状说明进行统计,但仍然无法正确完成。我要去哪里错了,什么会给我(11,128)值?
解决方法
暂无找到可以解决该程序问题的有效方法,小编努力寻找整理中!
如果你已经找到好的解决方法,欢迎将解决方案带上本链接一起发送给小编。
小编邮箱:dio#foxmail.com (将#修改为@)