Closed BenCretois closed 2 years ago
Hi,
Problem is solved, I was using a Gaussian instead of a Categorical likelihood for my data. The model now can be trained and returns a coherent accuracy (~96% for only 3 epochs of training).
Thank you for letting us know. I’m happy to hear you resolved the issue and find tyxe useful!
On Wed, Feb 2, 2022 at 4:18 AM Benjamin Cretois @.***> wrote:
Hi,
Problem is solved, I was using a Gaussian instead of a Categorical likelihood for my data. The model now can be trained and returns a coherent accuracy (~96% for only 3 epochs of training).
— Reply to this email directly, view it on GitHub https://github.com/TyXe-BDL/TyXe/issues/11#issuecomment-1027881278, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAKWCL4E5ORSB5CHCW6TRP3UZEOHRANCNFSM5NJVQBOA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.
You are receiving this because you are subscribed to this thread.Message ID: @.***>
It's really useful indeed! I first attempted to convert my DL models with Pyro but was unsuccessful. TyXe makes it way easier. I made a notebook about the analysis I was doing - using the classic cat / dogs dataset. I could share it if you think that would be useful for others who would like to use TyXe :)
Glad you managed to work things out. We should make it explicit in the docs what likelihoods correspond to which loss functions, I'm realizing that it's probably not very clear at the moment, thanks for opening the issue.
If you could share a link to your notebook that would be great! I'm sure others would be grateful for seeing more example use cases. I'd be happy to start a collection of contributed examples at the bottom of the readme and just put the link there for now.
Here is the cleaned notebook: https://gitlab.com/nina-data/deepexperiments/-/blob/main/notebooks/bayesianfy.ipynb. I still need to add a bit more text but most of it is already there. It's currently on a Gitlab repo but it's no problem for me to place it on GitHub! And yes would be great to have a section of contributed examples, that would be very helpful!
Error while computing log_prob at site 'obs': Value is not broadcastable with batch_shape+event_shape: torch.Size([32, 3, 128, 128]) vs torch.Size([32, 3, 128]).
I have encountered a similar mistake as you. Do you know how to correct the code? For example, how to modify the size here. By the way, I can't seem to open your notebook
Hi,
First, I want to say that the work you've done is clearly amazing.
I have playing a bit with TyXe and I am trying to convert the FC layer of a pretrained ResNet to a probabilistic layer.
I have been implementing the code shown in this paper, unfortunately I was unsuccessful. When using a homoskedastic Gaussian for the likelihood function I get this error:
My code is very minimalistic:
Thanks for your help already!