Image size issues

[qubvel]

Image size issues

[psinger]

Might have figured it out myself, dimension should be multiples of 32.

[julienguegan]

@psinger I have also noticed that when training Unet ... Do you know why ? Is it specific to Unet or is it the same with other architecture ?

[JulienMaille]

@qubvel this question comes over an over, any idea if a better error message could be thrown? Or could we pad them to avoid this?

@julienguegan images needs to be a multiple of 2^depth because they will go through subsampling depth times (with rounding), before being up-sampled and concatenated in the skip connections.
In the example above :

 525 ---------------> 544
   \                  /
   263 -----------> 272
     \              /
     132 -------> 136
       \          /
       66 -----> 68
         \      /
         33 -> 34 !
           \  /
            17

You can see the concatenation will fail Sizes of tensors must match except in dimension 1. Got 33 and 34 in dimension 3

[qubvel]

It is possible to add abstract _check_input method for SegmentationModel and implement it for different models

[qubvel]

Or change unet upsampling, instead of factor specify shape for interpolation

[qubvel]

Such kind of upsamplig have been chosen to simplify conversion between different frameworks, not sure that all of them support shape upsampling