Feature/Recv op python with guard #7793
Conversation
typhoonzero
changed the title
|
|
||
| if __name__ == "__main__": | ||
| unittest.main() | ||
| # test = TestRecvOp() |
There was a problem hiding this comment.
Please delete the annotated code.
There was a problem hiding this comment.
Done. Since current CI do not turn on WITH_DISTRIBUTE, I also added a switch to turn off the test of recv_op.
| o = layers.scale(x=x, scale=10.0) | ||
| main.global_block().create_var( | ||
| name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape) | ||
| print main |
There was a problem hiding this comment.
Don't use print in the unit test.
| append_batch_size=False) | ||
| fluid.initializer.Constant(value=1.0)(x, main.global_block()) | ||
| layers.Send("127.0.0.1:6174", [x], [x]) | ||
| print main |
… recv_op_python_with_guard
python/paddle/v2/fluid/layers/io.py
Outdated
| self.outputs = [] | ||
| self.endpoint = endpoint | ||
| self.fan_in = fan_in | ||
| # FIXME(typhoonzero): Add this switch is stupid |
There was a problem hiding this comment.
Maybe write a more precise comment? I could not understand what this means :)
There was a problem hiding this comment.
Oh, thanks. If you are busy I can take #7700 as well. It's not high priority anyways.
There was a problem hiding this comment.
Okay, I'll focus on performance optimization.
| @@ -161,7 +161,6 @@ class RecvOpMaker : public framework::OpProtoAndCheckerMaker { | |||
| public: | |||
| RecvOpMaker(OpProto *proto, OpAttrChecker *op_checker) | |||
| : OpProtoAndCheckerMaker(proto, op_checker) { | |||
| AddInput("RX", "(Tensor) Input tensor to be optimized").AsDuplicable(); | |||
There was a problem hiding this comment.
Why this line is removed, I thought the OP need to explicitly specify its input, otherwise the framework won't know about this dependency.
There was a problem hiding this comment.
Current recv_op should be called listen_and_serv, which is often the only operator in a server program's main block (block 0). Unlike send/recv it does not appear in the calculation blocks.
There was a problem hiding this comment.
Thanks for the explanation, I think I misunderstood the code.
| @@ -161,7 +161,6 @@ class RecvOpMaker : public framework::OpProtoAndCheckerMaker { | |||
| public: | |||
| RecvOpMaker(OpProto *proto, OpAttrChecker *op_checker) | |||
| : OpProtoAndCheckerMaker(proto, op_checker) { | |||
| AddInput("RX", "(Tensor) Input tensor to be optimized").AsDuplicable(); | |||
There was a problem hiding this comment.
Delete the Inputs in RecvOp may cause all the book demo under distributed_book folder can not run.
… recv_op_python_with_guard
This is part of job of #6508