Merge pull request #7371 from emailweixu/assign_value_op

assign_value operator

paddle/framework/tensor_util.h

@@ -116,8 +116,8 @@ inline void Copy(const Tensor& src, const platform::Place& dst_place,
  * @param[in] src        The external tensor.
  * @param[in] ctx        The device context contains device resources.
  *
- * * @note    CopyFromVector assumes that the tensor has been resized
- *            before invoking.
+ * * @note    CopyFromVector will resize dst to an 1D tensor with the same
+ *            size as src.
  */
 template <typename T>
 inline void CopyFromVector(const std::vector<T>& src,

paddle/operators/assign_value_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed 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. */
+
+#include "paddle/operators/assign_value_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AssignValueOp : public framework::OperatorWithKernel {
+ public:
+  AssignValueOp(const std::string &type,
+                const framework::VariableNameMap &inputs,
+                const framework::VariableNameMap &outputs,
+                const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of AssignValueOp should not be null.");
+    auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    ctx->SetOutputDim("Out", framework::make_ddim(shape));
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    return framework::OpKernelType(
+        framework::proto::DataType(ctx.Attr<int>("dtype")), ctx.GetPlace());
+  }
+};
+
+class AssignValueOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AssignValueOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddOutput("Out", "(Tensor) Output tensor of assign_value operator.");
+    AddAttr<std::vector<int>>("shape",
+                              "(vector<int>) "
+                              "Shape of values.");
+    AddAttr<int>("dtype", "data type of values")
+        .InEnum({framework::proto::DataType::INT32,
+                 framework::proto::DataType::FP32});
+    AddAttr<std::vector<float>>("fp32_values", "store the float values")
+        .SetDefault({});
+    AddAttr<std::vector<int>>("int32_values", "store the int values")
+        .SetDefault({});
+    AddComment(R"DOC(
+AssignValue operator
+
+$$Out = values$$
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(assign_value, ops::AssignValueOp, ops::AssignValueOpMaker);
+REGISTER_OP_CPU_KERNEL(assign_value, ops::AssignValueKernel<int>,
+                       ops::AssignValueKernel<float>);

paddle/operators/assign_value_op.cu.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+Indicesou 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. */
+
+#include "paddle/operators/assign_value_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(assign_value, ops::AssignValueKernel<int>,
+                        ops::AssignValueKernel<float>);

paddle/operators/assign_value_op.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed 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. */
+
+#pragma once
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class AssignValueKernel : public framework::OpKernel<T> {
+ public:
+  virtual void Compute(const framework::ExecutionContext& ctx) const {
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    int dtype = ctx.Attr<int>("dtype");
+    const char* value_name = nullptr;
+    switch (dtype) {
+      case framework::proto::DataType::INT32:
+        value_name = "int32_values";
+        break;
+      case framework::proto::DataType::FP32:
+        value_name = "fp32_values";
+        break;
+      default:
+        PADDLE_THROW("Unsupported dtype for assign_value_op: %d", dtype);
+        break;
+    }
+    auto values = ctx.Attr<std::vector<T>>(value_name);
+    framework::CopyFromVector(values, ctx.device_context(), out);
+    out->Resize(framework::make_ddim(shape));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/v2/fluid/layers/tensor.py

 from ..layer_helper import LayerHelper
 from ..param_attr import ParamAttr
+from ..framework import convert_np_dtype_to_dtype_
+from ..framework import Variable
+from ..core import DataType
+import numpy
 
 __all__ = [
     'create_tensor', 'create_parameter', 'cast', 'concat', 'sums', 'assign',
@@ -121,7 +125,7 @@ def assign(input, output):
     This function copies the *input* Variable to the *output* Variable.
 
     Args:
-        input(Variable): The source variable
+        input(Variable|numpy.ndarray): The source variable
         output(Variable): The destination variable
 
     Returns:
@@ -134,11 +138,37 @@ def assign(input, output):
           fluid.layers.assign(hidden, out)
     """
     helper = LayerHelper('assign', **locals())
-    helper.append_op(
-        type='scale',
-        inputs={'X': [input]},
-        outputs={'Out': [output]},
-        attrs={'scale': 1.0})
+    if isinstance(input, Variable):
+        helper.append_op(
+            type='scale',
+            inputs={'X': [input]},
+            outputs={'Out': [output]},
+            attrs={'scale': 1.0})
+    elif isinstance(input, numpy.ndarray):
+        dtype = convert_np_dtype_to_dtype_(input.dtype)
+        if dtype == DataType.FP32:
+            value_name = "fp32_values"
+            values = [float(v) for v in input.flat]
+        elif dtype == DataType.INT32:
+            value_name = "int32_values"
+            values = [int(v) for v in input.flat]
+        else:
+            raise ValueError("Unsupported dtype %s", input.dtype)
+        if input.size > 1024 * 1024:
+            raise ValueError("The size of input is too big. Please consider "
+                             "saving it to file and 'load_op' to load it")
+
+        helper.append_op(
+            type='assign_value',
+            outputs={'Out': [output]},
+            attrs={
+                'dtype': dtype,
+                'shape': list(input.shape),
+                value_name: values
+            })
+    else:
+        raise ValueError("Wrong type for assign input: %s" % type(input))
+
     return output
 
 

python/paddle/v2/fluid/tests/test_assign_value_op.py

+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.layers as layers
+import op_test
+import numpy
+import unittest
+import paddle.v2.fluid.framework as framework
+
+
+class TestAssignValueOp(op_test.OpTest):
+    def setUp(self):
+        self.op_type = "assign_value"
+        x = numpy.random.random(size=(2, 5)).astype(numpy.float32)
+        self.inputs = {}
+        self.outputs = {'Out': x}
+        self.attrs = {
+            'shape': x.shape,
+            'dtype': framework.convert_np_dtype_to_dtype_(x.dtype),
+            'fp32_values': [float(v) for v in x.flat]
+        }
+
+    def test_forward(self):
+        self.check_output()
+
+    def test_assign(self):
+        val = (
+            -100 + 200 * numpy.random.random(size=(2, 5))).astype(numpy.int32)
+        x = layers.create_tensor(dtype="float32")
+        layers.assign(input=val, output=x)
+        exe = fluid.Executor(fluid.CPUPlace())
+        fetched_x = exe.run(fluid.default_main_program(),
+                            feed={},
+                            fetch_list=[x])[0]
+        self.assertTrue(
+            numpy.array_equal(fetched_x, val),
+            "fetch_x=%s val=%s" % (fetched_x, val))
+        self.assertEqual(fetched_x.dtype, val.dtype)
+
+
+if __name__ == '__main__':
+    unittest.main()