Improve unet_simple

[Nelson-Gon]

Description

I would like to improve the initialization of unet_simple especially regarding making it similar to what the original paper proposed.

Similar Features

It is currently implemented in unet_simple

Feature Details

• Improve upsampling

• Review Downsampling

Proposed Implementation

None yet

Thank you

[nirolada]

What did the original paper propose?

[Nelson-Gon]

Hi,

In the original paper, the authors used 23 convolutional layers.

I think these may be too many for some datasets and it would be better to provide a way to customize the number of layers in the model to avoid overfitting.

The problem with unet_simple is that it does not accurately follow the downsample-upsample workflow.

The original architecture from the paper:

[nirolada]

I think you might want to add an argument to the function for the conv depth.
def unet-simple(..., depth=3)
'depth' should range from 1 to 3 (representing the number of conv layers at each stage).
What do you think?

[Nelson-Gon]

Yes, this is in line what I had in mind. I was also planning to rewrite the whole unet method to move away from manually creating layers. In that case, unet_simple would not be needed at all. Instead, one would just define the target depth and this would be automatically built.

[nirolada]

I will try to look into it soon. If I manage to build something, I will let you know. No need to assign me for now.

[nirolada]

Assign it to me, I'll be finished soon.