How's the performance of this implementation

[eglxiang]

No description provided.

[milesial]

The network implementation relies on Pytorch, so its performance is on par with the framework's. There shouldn't be anything special if you use the last version of pytorch. If you use bilinear=True for the upsampling, things are faster and need less memory.

About the data loading and overall pipeline, it's Python so you get what you expect.

[xiaoxingzeng]

This implementation is very different from the origin U-Net, I think its depend on your task.

[milesial]

@xiaoxingzeng Can you tell me what's different from the original architecture ? I think my implementation matches the Fig. 1 from https://arxiv.org/pdf/1505.04597.pdf

[xiaoxingzeng]

@xiaoxingzeng Can you tell me what's different from the original architecture ? I think my implementation matches the Fig. 1 from https://arxiv.org/pdf/1505.04597.pdf

Hi milesial, I check the paper and your code.
Sorry，it's my mistake, I reference the paper not the origin paper you referenced. Your reimplementation is correct.