You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by GitBox <gi...@apache.org> on 2018/09/14 21:45:39 UTC
[GitHub] anirudh2290 closed pull request #12290: [MXNET-798] Fix the dtype
cast from non float32 in Gradient computation
anirudh2290 closed pull request #12290: [MXNET-798] Fix the dtype cast from non float32 in Gradient computation
URL: https://github.com/apache/incubator-mxnet/pull/12290
This is a PR merged from a forked repository.
As GitHub hides the original diff on merge, it is displayed below for
the sake of provenance:
As this is a foreign pull request (from a fork), the diff is supplied
below (as it won't show otherwise due to GitHub magic):
diff --git a/src/executor/graph_executor.cc b/src/executor/graph_executor.cc
index 265554ab391..54a8d224ff4 100644
--- a/src/executor/graph_executor.cc
+++ b/src/executor/graph_executor.cc
@@ -141,8 +141,8 @@ nnvm::NodeEntry AggregateGradient(std::vector<nnvm::NodeEntry>&& v) {
if (v.empty()) {
nnvm::NodePtr ng = nnvm::Node::Create();
- ng->attrs.op = zeros_op;
- ng->attrs.name = "zeros";
+ ng->attrs.op = Op::Get("_zeros_without_dtype");
+ ng->attrs.name = "zeros_without_dtype";
ng->attrs.op->attr_parser(&(ng->attrs));
return nnvm::NodeEntry{ng, 0, 0};
}
diff --git a/src/operator/tensor/init_op.cc b/src/operator/tensor/init_op.cc
index bb23f5d44f6..8554ba85417 100644
--- a/src/operator/tensor/init_op.cc
+++ b/src/operator/tensor/init_op.cc
@@ -30,9 +30,22 @@ namespace op {
DMLC_REGISTER_PARAMETER(InitOpParam);
DMLC_REGISTER_PARAMETER(InitOpWithScalarParam);
+DMLC_REGISTER_PARAMETER(InitOpWithoutDTypeParam);
DMLC_REGISTER_PARAMETER(RangeParam);
DMLC_REGISTER_PARAMETER(EyeParam);
+NNVM_REGISTER_OP(_zeros_without_dtype)
+.describe("fill target with zeros without default dtype")
+.set_num_inputs(0)
+.set_num_outputs(1)
+.set_attr_parser(ParamParser<InitOpWithoutDTypeParam>)
+.set_attr<nnvm::FInferShape>("FInferShape", InitShape<InitOpWithoutDTypeParam>)
+.set_attr<nnvm::FInferType>("FInferType", InitType<InitOpWithoutDTypeParam>)
+.set_attr<FInferStorageType>("FInferStorageType",
+ InitStorageType<InitOpWithoutDTypeParam, true, true>)
+.set_attr<FCompute>("FCompute<cpu>", FillCompute<cpu, 0>)
+.set_attr<FComputeEx>("FComputeEx<cpu>", FillComputeZerosEx<cpu>)
+.add_arguments(InitOpWithoutDTypeParam::__FIELDS__());
NNVM_REGISTER_OP(_zeros)
.describe("fill target with zeros")
diff --git a/src/operator/tensor/init_op.cu b/src/operator/tensor/init_op.cu
index 81d835ee3bd..902b567516b 100644
--- a/src/operator/tensor/init_op.cu
+++ b/src/operator/tensor/init_op.cu
@@ -44,6 +44,9 @@ void FillZerosCsrImpl(mshadow::Stream<mshadow::gpu> *s, const NDArray& dst) {
});
}
+NNVM_REGISTER_OP(_zeros_without_dtype)
+.set_attr<FCompute>("FCompute<gpu>", FillCompute<gpu, 0>)
+.set_attr<FComputeEx>("FComputeEx<gpu>", FillComputeZerosEx<gpu>);
NNVM_REGISTER_OP(_zeros)
.set_attr<FCompute>("FCompute<gpu>", FillCompute<gpu, 0>)
diff --git a/src/operator/tensor/init_op.h b/src/operator/tensor/init_op.h
index 304911a02a7..1a4790acdb2 100644
--- a/src/operator/tensor/init_op.h
+++ b/src/operator/tensor/init_op.h
@@ -61,6 +61,24 @@ struct InitOpParam : public dmlc::Parameter<InitOpParam> {
}
};
+struct InitOpWithoutDTypeParam : public dmlc::Parameter<InitOpWithoutDTypeParam> {
+ TShape shape;
+ std::string ctx;
+ int dtype;
+ DMLC_DECLARE_PARAMETER(InitOpWithoutDTypeParam) {
+ DMLC_DECLARE_FIELD(shape)
+ .set_default(TShape())
+ .describe("The shape of the output");
+ DMLC_DECLARE_FIELD(ctx)
+ .set_default("")
+ .describe("Context of output, in format [cpu|gpu|cpu_pinned](n)."
+ "Only used for imperative calls.");
+ DMLC_DECLARE_FIELD(dtype)
+ .set_default(-1)
+ .describe("Target data type.");
+ }
+};
+
struct EyeParam : public dmlc::Parameter<EyeParam> {
nnvm::dim_t N;
nnvm::dim_t M;
diff --git a/tests/python/unittest/test_infer_type.py b/tests/python/unittest/test_infer_type.py
new file mode 100644
index 00000000000..bad83f3ef01
--- /dev/null
+++ b/tests/python/unittest/test_infer_type.py
@@ -0,0 +1,58 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+# pylint: skip-file
+import mxnet as mx
+import numpy as np
+from common import models, with_seed
+from mxnet import autograd
+from nose.tools import *
+from mxnet.test_utils import assert_almost_equal
+
+@with_seed()
+def test_infer_multiout_op():
+ data = mx.nd.arange(16, dtype=np.float64).reshape((4, 4))
+ data.attach_grad()
+
+ with autograd.record():
+ y = mx.nd.split(data, axis=0, num_outputs=2)
+ y[0].backward()
+ assert data.grad.dtype == np.float64
+
+@with_seed()
+def test_infer_multiout_op2():
+ def test_func(a):
+ q, l = mx.nd.linalg.gelqf(a)
+ return mx.nd.sum(l)
+
+ data32 = mx.nd.random.normal(shape=(2, 3), ctx=mx.cpu(), dtype=np.float32)
+ data32.attach_grad()
+ with autograd.record():
+ test32 = test_func(data32)
+ test32.backward()
+
+ data64 = mx.nd.Cast(data32, dtype=np.float64)
+ data64.attach_grad()
+ with autograd.record():
+ test64 = test_func(data64)
+ test64.backward()
+ assert_almost_equal(data64.grad.asnumpy(), data32.grad.asnumpy(), atol=1e-5, rtol=1e-5)
+
+
+if __name__ == '__main__':
+ import nose
+ nose.runmodule()
\ No newline at end of file
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
With regards,
Apache Git Services