1
$\begingroup$

I am developing an MIT open source GAN to generate synthetic structured health data, in the standard FHIR format. I have a running GAN, and it is using 200,000 Synthea Patient FHIR resources to train. I am trying to shape the tensor using the FHIR Metadata. Everything is in the repo here:https://github.com/fhirfly/ganfhir.

I am new at GAN, so keep that in mind.

These are some of the stats from the end of training loop:

Epoch [196/200], Batch complete with [1000 passes, Discriminator Loss: 0.0160, Generator Loss: 4.2750

Epoch [197/200], Batch complete with [1000 passes, Discriminator Loss: 0.0156, Generator Loss: 4.2950

Epoch [198/200], Batch complete with [1000 passes, Discriminator Loss: 0.0158, Generator Loss: 4.2761

Epoch [199/200], Batch complete with [1000 passes, Discriminator Loss: 0.0153, Generator Loss: 4.3070

Epoch [200/200], Batch complete with [1000 passes, Discriminator Loss: 0.0153, Generator Loss: 4.3113

Also after I train, I try to generate a FHIR patient by loading the training in the generator and generating some noise on the network, which brings back a numeric sequence, but I am not sure how to convert that to human readable text. My attempt wrote out 'e' to the terminal.

$\endgroup$

1 Answer 1

0
$\begingroup$

We solved the problem with the following code in the trainer. We also has to squeeze some the dimentions out of tensors without the gradiants there. You can see the full code in the repo above/

Clip discriminator's gradients

    for p in discriminator.parameters():
        p.data.clamp_(-0.01, 0.01)
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.