Pytorch Model with multiple inputs about apex

[mcarilli]

Sorry, it's been a busy last couple of weeks and I overlooked this question. It appears to be a limitation of torch.nn.Sequential which is used to define network_to_half . I'll double check with the Pytorch people whether torch.nn.Sequential is intended to support multiple inputs. Seems like a reasonable requirement. Maybe there's something obvious I'm missing.

In the meantime, what network_to_half is doing is really simple: 1. converting input to half if it isn't already, and 2. converting batchnorms to float. Therefore, as long as you manually call .half() on your inputs before starting the forward pass, you don't need 1., which means you don't need the full network_to_half Sequential sequence. Instead of

my_model = network_to_half(my_model)

you can implement only 2. by saying

my_model = BN_convert_float(my_model.half())

and after that my_model should still be able to handle whatever input structure you intended.