From 324876bbbfb0dd84f2172f951a2a4880bee32df4 Mon Sep 17 00:00:00 2001
From: Abhinav Arora <abhinavarora28@gmail.com>
Date: Tue, 3 Oct 2017 17:26:02 -0700
Subject: [PATCH] Changing learning rate from type Input(float) to
 Input(tensor) (#4578)

---
 paddle/operators/sgd_op.cc                      | 3 +++
 paddle/operators/sgd_op.h                       | 2 +-
 python/paddle/v2/framework/tests/test_sgd_op.py | 2 +-
 3 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc
index 8f9eae4186..1a4d3fb8c5 100644
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -32,6 +32,9 @@ class SGDOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput("param_out"),
                    "Output(param_out) of SGDOp should not be null.");
 
+    auto lr_dims = ctx->GetInputDim("learning_rate");
+    PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
+                      "Learning rate should have 1 element");
     auto param_dim = ctx->GetInputDim("param");
     PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("grad"),
                       "Two input of SGD Op's dimension must be same.");
diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h
index 977d201ced..e2ae65beb0 100644
--- a/paddle/operators/sgd_op.h
+++ b/paddle/operators/sgd_op.h
@@ -31,7 +31,7 @@ class SGDOpKernel : public framework::OpKernel<T> {
     auto param = ctx.Input<Tensor>("param");
     auto grad = ctx.Input<Tensor>("grad");
     auto param_out = ctx.Output<Tensor>("param_out");
-    float lr = *ctx.Input<float>("learning_rate");
+    float lr = ctx.Input<Tensor>("learning_rate")->data<float>()[0];
 
     param_out->mutable_data<T>(ctx.GetPlace());
 
diff --git a/python/paddle/v2/framework/tests/test_sgd_op.py b/python/paddle/v2/framework/tests/test_sgd_op.py
index f1125f4edb..c05364490f 100644
--- a/python/paddle/v2/framework/tests/test_sgd_op.py
+++ b/python/paddle/v2/framework/tests/test_sgd_op.py
@@ -8,7 +8,7 @@ class TestSGDOp(OpTest):
         self.op_type = "sgd"
         w = np.random.random((102, 105)).astype("float32")
         g = np.random.random((102, 105)).astype("float32")
-        lr = 0.1
+        lr = np.array([0.1]).astype("float32")
 
         self.inputs = {'param': w, 'grad': g, 'learning_rate': lr}
         self.outputs = {'param_out': w - lr * g}
-- 
GitLab