search

jiazhihao / TASO

The Tensor Algebra SuperOptimizer for Deep Learning
Apache License 2.0
687 stars 90 forks source link

When weight_initial is NULL then new_weight will fail #8

Open OuHangKresnik opened 4 years ago

OuHangKresnik commented 4 years ago

In ops.cc method new_weight, current logic is when weight_initial != NULL, then allocate_memory, else weight_ptr would be null ptr.

I don't know whether this logic is what we wanted, but there is an exception about it.

When look into ops_cudnn.cu:allocate_memory, there is a null condition inside actually.

So I think even we don't check weight_initial in ops.cc:new_weight, the logic still holds and no exception.

OuHangKresnik commented 4 years ago

Actually in this scenario, this tensor has a dimension of 0. I do believe this unusual situation needs more than a few line of code here

jiazhihao commented 4 years ago

Thanks for pointing out the issue. The bug should have been fixed in 6b884.

The logic for new_weight is that TASO will always allocate a device memory for the new weight. If weight_initial is provided, the allocated memory will be initialized using the given value.

The logic should work for 0-d tensors as well, though we haven't explicitly test TASO with 0-d tensors. Let me know if you observe any other issues with 0-d tensors.

OuHangKresnik commented 4 years ago

Thank you for your quick fixes.

noop.cc: create_weight, there is an assertion as well.

jiazhihao commented 4 years ago

I have replaced the assertion in noop.cc to a warning message in commit 49e804e8.

Presumably, when TASO encounters an uninitialized weight tensor, it is likely due to a bug in previous layers, and we should report a warning message to users.