Keras：替换模型的输入层不会更新模型，摘要等其他层的输出形状

Question

I am using Keras with TensorFlow backend. I want to use a pretrained U-Net model and replace the input layer with another one. I trained the model on images of size (256,256). When I am predicting a bigger scene I want to manipulate the input such that the UNet does what it does, just on another image size such that I don't have to cute the image or anything. Here is my code:

model = load(model_path)
model.layers.pop(0)
new_input = Input(shape = (512,512))
model = Model(new_input,model(new_input_layer))

Now when I am using

print(model.summary())

it outputs

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_1 (InputLayer)         (None, 512, 512, 3)     0
_________________________________________________________________
model_1 (Model)              multiple                  211825
=================================================================

and if I am doing

model.layers[1].summary()

I get

____________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connectedto
=========================================================================
conv2d_1 (Conv2D)             (None, 256, 256, 16)    1216     input_1[0][0]
__________________________________________________________________________
batch_normalization_1       (None, 256, 256, 16)      64      conv2d_1[0][0]

and so on. Typically the output shape of conv_2d_1 should be (None,512,512,16), but it wasn't updated properly (and for other layers as well). Further, when I am using

model.layers[1].layers[0].output_shape

I get the same result as in the summary.

When I am doing a prediction with the adapted model everything works fine with respect to the output. But if the image size is bigger than (512,512), for example like (4096,4096) I'm running into memory problems/ allocation problems with respect to gpu. Therefore I want to calculate the memory that is needed, to predict the image and cut it, if it is too big. But to write a function that does this for me I need the correct informations about the outputs shapes.
Does anyone have any suggestions? Maybe I should replace the input layer in another way?! Maybe I can update the model somehow? Or does some keras function already exist, which calculates the needed memory?(I didn't found any) Thanks for your attention! :)

Answer 1

There is a better alternative to what you are trying to achieve. The CNN layers can handle arbitrary shape if you specify shape=(None,None,3) says some height and width with 3 channels. You can train the original model like that and don't have to adjust tensor / image shapes when you predict.