Why can the upsampling in the UNet decoder only use interpolation but not transpose convolution?

[hanruiyao163]

The original UNet paper used transpose convolution for upsampling.

[qubvel]

Hi @hanruiyao163, as you might see, this implementation is quite different from the original one. However, it has evolved due to quality and convergence improvements and tested in Kaggle competitions.

The main differences of this implementation

• input/output resolutions are the same
• different number of channels in decoder
• different upsampling strategy
• different encoders (!), original one is build on outdated VGG

[hanruiyao163]

That is to say, is the performance of interpolation better than that of transpose convolution?