Keras层（函数）对应于tf.nn.conv2d_transpose

Question

In Keras , what are the layers(functions) corresponding to tf.nn.conv2d_transpose in Tensorflow ? I once saw the comment that we can Just use combinations of UpSampling2D and Convolution2D as appropriate . Is that right?

In the following two examples, they all use this kind of combination.

1) In Building Autoencoders in Keras , author builds decoder as follows.

2) In an u-uet implementation , author builds deconvolution as follows

up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4], mode='concat', concat_axis=1)
conv6 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(up6)
conv6 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(conv6)

Answer 1

The corresponding layers in Keras are Deconvolution2D layers.

It's worth to mention that you should be really careful with them because they sometimes might behave in unexpected way. I strongly advise you to read this Stack Overflow question (and its answer) before you start to use this layer.

UPDATE:

1. Deconvolution is a layer which was add relatively recently - and maybe this is the reason why people advise you to use Convolution2D * UpSampling2D .
2. Because it's relatively new - it may not work correctly in some cases. It also need some experience to use them properly.
3. In fact - from a mathematical point of view - every Deconvolution might be presented as a composition of Convolution2D and UpSampling2D - so maybe this is the reason why it was mentioned in texts you provided.

UPDATE 2:

Ok. I think I found an easy explaination why Deconvolution2D might be presented in a form of a composition of Convolution2D and UpSampling2D . We would use a definition that Deconvolution2D is a gradient of some convolution layer. Let's consider three most common cases:

1. The easiest one is a Convolutional2D without any pooling. In this case - as it's the linear operation - its gradient is a function itself - so Convolution2D .
2. The more tricky one is a gradient of Convolution2D with AveragePooling . So: (AveragePooling2D * Convolution2D)' = AveragePooling2D' * Convolution2D' . But a gradient of AveragePooling2D = UpSample2D * constant - so it's also in this case when the preposition is true.
3. The most tricky one is one with MaxPooling2D . In this case still (MaxPooling2D * Convolution2D)' = MaxPooling2D' * Convolution2D' But MaxPooling2D' != UpSample2D . But in this case one can easily find an easy Convolution2D which makes MaxPooling2D' = Convolution2D * UpSample2D (intuitively - a gradient of MaxPooling2D is a zero matrix with only one 1 on its diagonal. As Convolution2D might express a matrix operation - it may also represent the injection from a identity matrix to a MaxPooling2D gradient). So: (MaxPooling2D * Convolution2D)' = UpSampling2D * Convolution2D * Convolution2D = UpSampling2D * Convolution2D' .

The final remark is that all parts of the proof have shown that Deconvolution2D is a composition of UpSampling2D and Convolution2D instead of opposite. One can easily proof that every function of a form a composition of UpSampling2D and Convolution2D might be easily presented in a form of a composition of UpSampling2D and Convolution2D . So basically - the proof is done :)