Possible bug introduced in 1.0.49

fastai fastai dev
1

I’m doing segmentation using a unet model with standard configuration and a dice metric. Training correctly outputs the train_loss, valid_loss, and dice when running on 1.0.46. After upgrading to 1.0.49 the valid_loss is always #nan# and dice values are not output.

Machine configuration:

=== Software ===
python        : 3.7.2
fastai        : 1.0.49
fastprogress  : 0.1.20
torch         : 1.0.0
torch cuda    : 9.0 / is available
torch cudnn   : 7005 / is enabled

=== Hardware ===
torch devices : 2
  - gpu0      : GeForce GTX 1080 Ti
  - gpu1      : GeForce GTX 1080 Ti

=== Environment ===
platform      : Windows-10-10.0.16299-SP0
conda env     : fastai_v1
python        : C:\Users\lproc\AppData\Local\Continuum\anaconda3\envs\fastai_v1\python.exe
sys.path      : C:\Windows\system32
C:\Users\lproc\AppData\Local\Continuum\anaconda3\envs\fastai_v1\python37.zip
C:\Users\lproc\AppData\Local\Continuum\anaconda3\envs\fastai_v1\DLLs
C:\Users\lproc\AppData\Local\Continuum\anaconda3\envs\fastai_v1\lib
C:\Users\lproc\AppData\Local\Continuum\anaconda3\envs\fastai_v1
C:\Users\lproc\AppData\Local\Continuum\anaconda3\envs\fastai_v1\lib\site-packages
no nvidia-smi is found
2

There is nothing we can do without a reproducible example.