search

scipion-em / scipion-em-sphire

Plugin to use Sphire programs within the Scipion framework
GNU General Public License v3.0
0 stars 1 forks source link

Add Cinderella method in its own protocol. #13

Closed dmaluenda closed 4 years ago

dmaluenda commented 5 years ago

Add Cinderella protocol to the plugin would be nice.

It seems no so hard, its installation is very similar to crYOLO (see installation doc). While, for using, it can take .mrcs as input (see tutorial doc).

pconesa commented 5 years ago

We have tested this and is not rendering nice results with relion or xmipp averages...with the current generic model (trainned with isac averages) I don't see it very useful.

delarosatrevin commented 5 years ago

But maybe it can be trained with other 2D classes results?

On Thu, Oct 24, 2019 at 9:46 AM Pablo Conesa notifications@github.com wrote:

We have tested this and is not rendering nice results with relion or xmipp averages...with the current generic model (trainned with isac averages) I don't see it very useful.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/scipion-em/scipion-em-sphire/issues/13?email_source=notifications&email_token=ACAM6WA2RUAASRDMICER7ULQQFHD7A5CNFSM4JBW3JVKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOECEBZYQ#issuecomment-545791202, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACAM6WCSA4X3FO7LV6PSZJTQQFHD7ANCNFSM4JBW3JVA .

pconesa commented 5 years ago

Yep, I think it can. Maybe we can implement the training and keep training it locally (with some users collaboration) and release it in a future version.

JorMaister commented 5 years ago

For the training, I think it should be tested first to check if it works correctly and if the resulting model works fine. In the SPHIRE wiki for Cinderella they say that the current available model was trained with 19 data sets, so I suggest that the first approach should be to get, from another source, a number of data sets close to it and generate a trained model with them. But, it's not possible to know if that number is large enough to get nice results or if it needs to be higher for a generalization until it is tested.

delarosatrevin commented 5 years ago

Yes, I don't know if the final benefits are worth the effort. Because the final outcome is just to select good classes, that is good (if it works well) for automation, but it is not a big thing for normal processing.

JorMaister commented 5 years ago

Once we have the data sets, the initial test would be quick, but if the results aren't as good as expected, the process can be extended in time. Thus, I totally agree that the first thing which should be considered is if the benefits worth the effort.