tqchen commented on pull request #5756:
URL: https://github.com/apache/incubator-tvm/pull/5756#issuecomment-642241286
Thanks @t-vi I do not disagree, that is why we felt perhaps we should go
with float in the first place, how about also open a dicussion thread about
this topic?
tqchen commented on pull request #5756:
URL: https://github.com/apache/incubator-tvm/pull/5756#issuecomment-642218720
How about we first fix the usage of "float" throughout the codebase, then we
can send a PR to warn about "float" as dtype in the parser.
tqchen commented on pull request #5756:
URL: https://github.com/apache/incubator-tvm/pull/5756#issuecomment-642170343
Aha, that is definitely interesting. It is still worth to have a discussion
about the numpy compatibility vs favors common deep learning land. That could
leads to the
tqchen commented on pull request #5756:
URL: https://github.com/apache/incubator-tvm/pull/5756#issuecomment-642115243
Previously we use this approach. Personally I was in favor of default to
fp32.
However, one reason for us to convert to the current way is to just make
sure that