-
Notifications
You must be signed in to change notification settings - Fork 6.8k
Conversation
I think merging is a good idea. Not sure about submodule though. We have too many submodules already |
I also prefer copying mxnet.cpp directly to using submodule, as it allows direct reference to dependencies like c_api.h and dmlc-core. However, this will diverge the cpp package into two versions, one in dmlc/mxnet.cpp, and the other one in dmlc/mxnet, which is confusing. Maybe we can just move all the development to mxnet? |
Yes. We can mark mxnet.cpp as discontinued. |
We will mark dmlc/mxnet.cpp as deprecated. Future checkins will be made only in dmlc/mxnet. |
Can one of the admins verify this patch? |
any updates on this? |
I'm updating it to remove the submodule, and to include headers from mxnet. |
I have some question about BatchNorm. It recently added two arguments gamma and beta, which don't have default values. But I found that in python example codes, these arguments can be omitted. This makes mxnet.cpp unable to generate correct function signature for BatchNorm. |
not sure what you mean. Are you talking about initializer? what default value? |
For example, the default value for eps in batchnorm is 1e-3 (https://github.com/dmlc/mxnet/blob/master/src/operator/batch_norm-inl.h#L37), and gamma and beta have no default values(https://github.com/dmlc/mxnet/blob/master/src/operator/batch_norm.cc#L81), while some example code calls batchnorm without specifying gamma and beta (https://github.com/dmlc/mxnet/blob/master/tests/python/common/models.py#L16). |
gamma and beta are weights, not parameters. their default values are set by initializer |
ndarray-or-symbol replaced NDArray and Symbol for types. |
This pull request is ready for review now. |
Any idea why windows tests failed? |
Fixed |
As mentioned in dmlc/MXNet.cpp#71, we propose to merge mxnet.cpp into mxnet project.
It will attract more attention to mxnet.cpp, and it will also enable integrated test with mxnet, so that future checkins to mxnet will not break mxnet.cpp. Meanwhile, it is important for industrial environment (e.g, data center), where python and other dependencies are normally not installed.