为什么内核在训练开始后很快就死了？

Question

I'm trying to build a generative model for a dataset that I have. I have used a stack of Conv2DTranspose layers, using tensorflow. here is my code:

model = Sequential()

model.add(Conv2DTranspose(filters=2, kernel_size=3, input_shape = (2,2,1)))

for i in range(37):
    
    model.add(Conv2DTranspose(filters = 2*i, kernel_size = 3))
        

model.add(Conv2DTranspose(filters = 3, kernel_size=3))


model.compile(optimizer = 'Adam', loss='mean_squared_error')

model.fit(labels, ims, batch_size=32, epochs = 100) 

the input is of size (2,2,1) and the output is a (80,80,3) array which is supposed to be an image. When I run the code, when it comes to train the model, meaning the line containing:

model.fit()

everything stopes and the running is ceased without showing any error or warning. I have tried running this code using both Jupyter notebook and a simple .py file. The result was same in both case. By the way, the system which I tried to use for running the code is a normal desktop machine exploiting only CPU and RAM (no GPU is included, RAM capacity = 32 GB). How can I fix this?

Answer 1

This is because a mistake in this line:

for i in range(37):  
    model.add(tf.keras.layers.Conv2DTranspose(filters = 2*i, kernel_size = 3))

As you set each layer filters by 2*i , the first layer filters will be 0 because i is 0 . So, your second layer output shape will be something like (None,6,6,0) , which does not make sense, and since tf libraries con not handle this, the kernel dies. So, I think you can change this line to something like this to avoid the mistake:

for i in range(37):  
    model.add(tf.keras.layers.Conv2DTranspose(filters = 2*(i+1), kernel_size = 3))