Jonathon Hare
Jonathon Hare
@kkieffer @sarxos I remember now - there was a problem with dyncall in the bridj 0.7 (and 0.7-20140918) builds that meant that it wouldn't pick up libraries loaded from unusual...
In trial.py see method `run` (looping over the epochs): https://github.com/pytorchbearer/torchbearer/blob/9d97c60ec4deb37a0627311ddecb9c6f1429cd82/torchbearer/trial.py#L946 and `_fit_pass` (looping over the batches): https://github.com/pytorchbearer/torchbearer/blob/9d97c60ec4deb37a0627311ddecb9c6f1429cd82/torchbearer/trial.py#L1019 I'd have thought most additions to the training loop can be added via...
really, really (,really!) untested, but something like this should be equivalent based on the above and guessing the correct indentation: ``` loader, optimizer, model, loss_fn = ... swa_model = torch.optim.swa_utils.AveragedModel(model)...
Okay, my bad... I didn't realise that setuptools behaves rather differently in development mode & would mask missing bits from setup.py. Try pulling the latest code now and re-running `python...