具有动态输入形状的 CNN

Question

Hello everyone!

Since I am trying to make a Fully Convolutional Neural Network which converts grayscale images to rgb images, I was wondering if I could train and test the model on different sized images (different pixels and ratio). Normally you would just downsample or upsample, what I do not want to do. I heared that it might be possible if I use a Fully Convolutional Neural Network, but I still have no clue what the code would look like. Could you help me out with some code?

Why is this a problem?

Like I said, the input image should not be downsampled, because I do not classify anything. I want to produce a new image with the same size as the input image. So there should not be any loss.

Code for a fixed input shape:

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))
model.add(layers.MaxPooling2D((2, 2)))
...

with images of the size 28*28px

How I thought it might work:

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(None, None, 1)))
model.add(layers.MaxPooling2D((2, 2)))
...

with images of different sizes

• image1 = 256*300
• image2 = 1024*800
• image3 = 500*400

Here is an example of an autoencoder which converts grayscale images to rgb images, but this one has a fixed input shape.

*I am using TF 2.0 Alpha

Answer 1

I figured out, that a convolutional neural network does not care at all about the input shape. What it cares about, are kernel size, stride and padding. For example settingkernel size = 3, stride = 1, padding = 1 does not change the tensors shape. When it comes to pooling, it must be assured, that padding = 1 is added, which is called half/same padding ( http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html ). Thus, it is possible to make a fully convolutional autoencoder which is able to process images of different sizes.