Outputs are `nan` given a big window shape in iterator

[GlowingHorse]

When I run DASP on a convolutional neural network (mnist dataset), I found something weird.
If I use a ImagePlayerIterator with a big window_shape (like 1/4 of the image size), then y1 and y2 will produce nan.

y1, y2 = self.dasp_model.predict(inputs, batch_size=batch_size)
y1 = y1.reshape(len(ks), x.shape[0], -1, 2)
y2 = y2.reshape(len(ks), x.shape[0], -1, 2)

I tried to check if something like delta (0.0001) is missing in denominator or divisor, but it seems like no problem in codes of all lpdn or dasp layers.
Have you any idea about this? Thanks in advance.

[GlowingHorse]

Seems like some zero padding makes the output conv_count of the conv operation at zero.
-> dasp/layers/convolution.py
conv_count = self._conv2d(ghost, tf.ones_like(self.kernel))

I just add this
conv_count = tf.maximum(conv_count, tf.reduce_sum(tf.ones_like(self.kernel)))
after it. Is it same with your original thought?

[marcoancona]

Hi! It is still not clear to me how the result conv_count can be zero because both the input (ghost) and the kernel are made of ones.

EDIT: I did not see your mention to zero-padding. Your solution seems to make sense indeed.

[sebp]

I encountered the same issue. It also occurs when using an ImagePlayerIterator with the spatial dimensions of window_shape set to the kernel size of the first convolutional layer. In that case, conv_count will yield zero for the patch currently masked out, because the shape of the masked elements exactly matches that of the kernel size.

I'm not sure whether the fix proposed by @GlowingHorse would work. When taking the maximum, wouldn't conv_count always be the kernel size, disregarding the actual mask?