denoising not accepting flags

[imipolex-K]

Hi,
Great tool, I am excited to see how projects like this, tying together great new ideas with the existing Warp pipeline!

I noticed that the 'easymode denoise' command doesn't accept either the 'method' or 'mode' flags, despite what is detailed in the tutorial and README.
Digging in, I saw that in the 'main.py' file, the mode was hardcoded in line 163: n2n.dispatch(mode='direct',
Is there a reason to not use the 'splits' mode? I changed it in the main.py file to that and it seemed to run normal.

Thanks,
Keith

[mgflast]

Hi Keith, Thanks for the feedback! I had cut out the splits mode because I thought there was no genuine use case for it - the direct mode is faster and allows you to skip splits generation, plus the denoising quality is equivalent as far as I’ve been able to tell. Because it is a general pretrained model the statistical argument for denoising from splits no longer holds. But since the functionality is all in the code I’ll reinstate the mode and method argument when I get the chance - am almost done with a general DeepDeWedge network which would add a second method anyway :) best, Mart

…

On Thu, 28 May 2026 at 05:11, imipolex-K ***@***.***> wrote: *imipolex-K* created an issue (mgflast/easymode#15) <#15> Hi, Great tool, I am excited to see how projects like this, tying together great new ideas with the existing Warp pipeline! I noticed that the 'easymode denoise' command doesn't accept either the 'method' or 'mode' flags, despite what is detailed in the tutorial and README. Digging in, I saw that in the 'main.py' file, the mode was hardcoded in line 163: n2n.dispatch(mode='direct', Is there a reason to not use the 'splits' mode? I changed it in the main.py file to that and it seemed to run normal. Thanks, Keith — Reply to this email directly, view it on GitHub <#15?email_source=notifications&email_token=BSQIFOBY3U7C5ZSUHXGK2HT446U5TA5CNFSL4Z3JMQ5C6L3HNF2C22DVMIXUS43TOVSS6NBVGM3TIMZSGM3TPJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSQIFOE3LDXDQ3V2DBAJLJL446U5TAVCNFSM6AAAAACZQFGD4OVHI2DSMVQWIX3LMV43ASLTON2WKOZUGUZTONBTGIZTONY> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BSQIFOBY6VIGAKZQGDN7453446U5TA5CNFSL4Z3JMQ5C6L3HNF2C22DVMIXUS43TOVSS6NBVGM3TIMZSGM3TPJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/BSQIFOHPTRPC5FOBCNI5WRT446U5TA5CNFSL4Z3JMQ5C6L3HNF2C22DVMIXUS43TOVSS6NBVGM3TIMZSGM3TPJTSMVQXG33OVJZXKYTTMNZGSYTFMSSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[imipolex-K]

Hi Mart,

Thanks for the response, I'm looking forward to the DeepDeWedge implementation!

Do you think easymode segmentation is tolerant of denoising with other noise models and programs?
I am wondering if denoising with the same general noise models that you denoised the tomos you trained the segmentation on is better than custom noise models (like CryoCARE)? Or even combined with missing wedge (like Isonet2)?

I assume that each would affect the noise and signal that the segmentation model expects and might throw it off, but is this naive because of the number of different datasets used to train the model?

Thanks,
Keith

[mgflast]

Hi Keith, I’m sure it would be tolerant to any other noise2noise type denoiser, cryoCARE or Warp’s n2n included. I have seen some particularly nice segmentation results come out of missing wedge filled tomograms (e.g ribosomes, membranes, microtubules - i think it was with IsoNet2) but in some cases the segmentation output also did seem a bit worse; it missed some nuclear pore complexes that in a n2n denoised version of the same tomogram were detected. So it is a little bit unpredictable. That’s also why I finally committed to distilling a general DDW-like net (and IsoNet2 once I free up the GPU time) so I can use that for augmentation during training, so that the segmentation becomes a bit more input-flavour agnostic. In general I would advise: test segmentation on a selected subset of tomograms, maybe 5 or so, see what works and what doesn’t, then go from there. If you do test on some different input flavours I would love to hear about the results. best, Mart

…

On Thu, 28 May 2026 at 15:12, imipolex-K ***@***.***> wrote: *imipolex-K* left a comment (mgflast/easymode#15) <#15 (comment)> Hi Mart, Thanks for the response, I'm looking forward to the DeepDeWedge implementation! Do you think easymode segmentation is tolerant of denoising with other noise models and programs? I am wondering if denoising with the same general noise models that you denoised the tomos you trained the segmentation on is better than custom noise models (like CryoCARE)? Or even combined with missing wedge (like Isonet2)? I assume that each would affect the noise and signal that the segmentation model expects and might throw it off, but is this naive because of the number of different datasets used to train the model? Thanks, Keith — Reply to this email directly, view it on GitHub <#15?email_source=notifications&email_token=BSQIFOEVDMQIWZVQVMFCOS345A3NJA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJWGQZDSMBUHA42M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4564290489>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSQIFOBYUG5C5Y5VJA73OJL45A3NJAVCNFSM6AAAAACZQFGD4OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DKNRUGI4TANBYHE> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BSQIFOENRCYJZWFBAINUOS345A3NJA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJWGQZDSMBUHA42M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/BSQIFOERIT3SLK72CQVIDNT45A3NJA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJWGQZDSMBUHA42M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>