shijianjian/se_resnext_3d

SE_ResNeXt_3D

The origin of SENet

Code modified from 2d se-resnext

Note

The first layer was changed to ResNet-like structure due to the high memory cost.

• Before:

  x = conv_layer(x, filter=64, kernel=[3, 3], stride=1, layer_name=scope+'_conv1')
  x = Batch_Normalization(x, training=self.training, scope=scope+'_batch1')

• After:

  x = Conv3D(filters=self.init_filters, kernel_size=(7, 7, 7), strides=(2, 2, 2), padding='same')(x)
  x = MaxPooling3D(pool_size=(3, 3, 3), strides=(2, 2, 2), padding="same")(x)
  x = Batch_Normalization(x, training=self.training, scope=scope+'_batch1')

se_resnext_3d's Issues

关于利用senet进行T的attention！

您好，非常不错的工作。在resnet中可以加入senet。
另外一个想法就是我利用senet的**，将5维张量B×C×T×H×W通过transpose来进行C和T的转换使其变成B×T×C×H×W，然后通过squeeze和excatation的**进行压缩成B×T×1×1×1。然后经过2个FC层变成B×T‘×1×1×1，这样就完成了对时域的attention。请问您知道怎么实现这个操作吗？或者您能提供下代码吗？就是嵌入到resnet中的代码。谢谢！
期待您的回复
@shijianjian