No missing wedge compensation after prediction

[Yunjie Chang]

Hi all,

I have a weird issue and can not figure out the reasons.

I am using Isonet v0.2 and it works well with the tutorial data.

When I try my own data, with tomogram reconstructed by Warp (Linux version) or tomo3D, all Isonet steps run smoothly without any errors or warnings. However, after prediction, the tomograms are only denoised, but no missing wedge compensation.

I also tried to skip the deconvolution step, but nothing helps.

Some other informations about my own data:

1. Collected by Krios with Falcon4i, raw data in eer format.

2. Tilt angle in -51˚ to 51˚.

3. Tomograms are reconstructed by WBP and always saved in 32-bit real type.

4. Tried tomograms reconstructed with pixel size 14.4Å and 18Å, no big differences.

5. I have also tested tomograms reconstructed by IMOD SIRT, but nothing helps.

6. I alway use default parameters for the training step, only tune the deconvolution/mask parameters when necessary.

7. I checked the *json files in the training results, the "losses" are not NAN, the values are around 0.15 for the 25th iterations.

Any idea about how to fix the issue?

Thanks!

Yunjie

[YUNTAO LIU]

Hi Yunjie,

We are also puzzled by this, IsoNet does not perform well for some tomograms. I have a feeling of why this happened, but I need more tests. 

I think these following might help: 1: try WBP reconstructed with IMOD. 2. try to use tomograms from -60 to 60 

--
You received this message because you are subscribed to the Google Groups "IsoNet" group.
To unsubscribe from this group and stop receiving emails from it, send an email to isonet+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/isonet/6c0f7f89-e091-44ea-9184-09af207d0909n%40googlegroups.com.

--

Best Regards,

Yuntao Liu,  Postdoc.

California NanoSystem Institute

University of California Los Angeles

[Yunjie Chang]

Hi Yuntao,

Thanks for the reply.

We just tested IMOD WBP reconstructions recently.

Another issue happened is that, if I use only 1 IMOD WBP tomogram for training and prediction, then Isonet works and it compensates (at least part of) the missing wedge. However, if I use 5 IMOD WBP tomograms (they are from the same data collection session) for training and prediction, then Isonet does not work and there is no missing wedge compensation.

We guess maybe it's related to the training part. Somehow the training might fail when we feed multiple tomograms. Do you think it might help if we tune the "learning rate" or "epoch number" parameters? Or use a more precise or more broad mask? If this might help, how should we check what parameters we should choose? By checking the loss and val_loss curves? Or something else?

Thanks

Yunjie

--

Best

Yunjie Chang