deprecation warning/error fixes for new dependency versions
Looks like in the torch-1.0.x, the labels tensor is in CUDA mem by the time it hits reporter.py, leading to problems when scipy forces a conversion of the tensor to numpy. Brought to light in #4; this PR should fix that issue.
Further, now h5py preferred syntax for accessing data values doesn't use .value; using this was emitting a warning for each observation.
Finally, a sum call that should have been a torch.sum call was requiring a further wrapping in torch.tensor, which is a mess and was spitting errors; this has been fixes.
deprecation warning/error fixes for new dependency versions
Looks like in the
torch-1.0.x, thelabelstensor is in CUDA mem by the time it hitsreporter.py, leading to problems whenscipyforces a conversion of the tensor tonumpy. Brought to light in #4; this PR should fix that issue.Further, now
h5pypreferred syntax for accessing data values doesn't use.value; using this was emitting a warning for each observation.Finally, a
sumcall that should have been atorch.sumcall was requiring a further wrapping intorch.tensor, which is a mess and was spitting errors; this has been fixes.