[Relay/TOPI][Op] Add erf intrinsic and op #3702
Conversation
|
Would you mind also adding this to the TensorFlow frontend? The op name is Erf. |
|
@soiferj Sure, will add it. |
|
@icemelon9 , is there a timeline to get this PR ready for review? This would be really nice to unblock running BERT-based models! |
|
@soiferj Sorry about the delay. I was working on the dynamic shape function support for Any. I'll start updating this PR this week. |
|
@tqchen Could you help check if the vectorizable intrinsic list is complete? |
| @@ -85,6 +85,18 @@ RELAY_REGISTER_UNARY_OP("exp") | |||
| .set_support_level(1) | |||
| .set_attr<FTVMCompute>("FTVMCompute", RELAY_UNARY_COMPUTE(topi::exp)); | |||
|
|
|||
|
|
|||
| RELAY_REGISTER_UNARY_OP("erf") | |||
| .describe(R"code(Returns the error function value for input array, computed element-wise. | |||
There was a problem hiding this comment.
| .describe(R"code(Returns the error function value for input array, computed element-wise. | |
| .describe(R"code(Returns the Gauss error function value for input array, computed element-wise. |
| @@ -55,10 +55,17 @@ def _mx_fully_connected(inputs, attrs): | |||
| use_flatten = attrs.get_bool("flatten", True) | |||
| if has_flatten and use_flatten: | |||
| inputs[0] = _op.nn.batch_flatten(inputs[0]) | |||
| data_shape = _infer_type(inputs[0]).checked_type.shape | |||
| if len(data_shape) > 2: | |||
| inputs[0] = _op.reverse_reshape(inputs[0], [-1, 0]) | |||
There was a problem hiding this comment.
could you elaborate a bit what are these changes for?
There was a problem hiding this comment.
This is because mxnet dense op allows (d1, d2, ..., dk, in_dim), (out_dim, in_dim) --> (d1, d2, .., dk, out_dim). But this isn't allowed in topi's dense op.
|
Thanks @icemelon9 @yongwww |
* add more ops * stop vectorization for erf * x * cleanup * fix * add whitelist for vectorizable intrin * add tf converter * fix dense * fix * add missing intrin * fix mxnet frontend * fix nvptx
* add more ops * stop vectorization for erf * x * cleanup * fix * add whitelist for vectorizable intrin * add tf converter * fix dense * fix * add missing intrin * fix mxnet frontend * fix nvptx
* add more ops * stop vectorization for erf * x * cleanup * fix * add whitelist for vectorizable intrin * add tf converter * fix dense * fix * add missing intrin * fix mxnet frontend * fix nvptx
@yzhliu @masahi @yongwww @kevinthesun could you help review this pr?