U-Net transfer learning

Deep Learning
1

I attempted using transfer learning on a u-net for the Planet Amazon competition.
This was done by attaching a global average pooling layer to the out of the u-net to create the 17D output.
Does this approach make sense?

It achieved ~0.89 on the LB but I was not able to run it on the full sized images because of memory and time constraints. Any inputs much appreciated.

PS: I am in the market for a job and/or looking to work on projects.

2

Personally I’d replace the last convolutional layer instead.

You may also want to think about training each channel alone in 17 networks. Takes a lot of time but the results will probably be better. You can use smaller variants of U-Net to speed it up and train them from scratch.

All that said, wasn’t it a classification competition?