fix inconsistency in multi-channel network

ShixingWang commented 4 years ago

Pull request recommendations:

[x] Name your pull request: bug-fix/dimension-fix-for-multi-channel.
[x] Link to any relevant issue in the PR description: seems to be relevant to #112
[x] Provide context of changes. I want to train this network over some multi-channel single layer images (shape = [chan, row, col]), so I just copied fnet/nn_modules/fnet_nn_3d_params.py and changed all the 3d functions to 2d. However, some errors showed up. I walked around them with some nasty modifications, and this is the only one that is a must-fix. I do not include other mods because they may not obey your conventions of image shape, and I am not confident with my programming skill and style. Things I am not comfortable with include:
- fnet/data/tiffdataset.py assumes the channel to be one when unsqueezing the images
- fnet/data/bufferedpatchdataset.py assumes the signal and target images have the same number of channels.
[] Provide relevant tests for your feature or bug fix. As the context part said, my own images are 2 dimensional, so I do not have direct tests for this. But after the same fix the network works out fine on 2d.
[ ] Provide or update documentation for any feature added by your pull request.

gregjohnso commented 4 years ago

@JacksonMaxfield is there a way we can run CI tests on this?

gregjohnso commented 4 years ago

@ShixingWang were you able to look at the fnet_nn_2d.py?

evamaxfield commented 4 years ago

@ShixingWang unfortunately I am going to have to close this PR. The branch you are trying to pull request into is specifically locked because that is the version of the repo that was used for the FNET paper. If you want to make a PR in the future, please point it at the master branch.

Apologies for any confusion.

AllenCellModeling / pytorch_fnet

fix inconsistency in multi-channel network #165