Make bilinear_interp_op support attrs from input. (#11041)

wanghaoshuang · web-flow · commit 85c203b11779 · 2018-06-01T13:14:23.000+08:00
* Make bilinear_interp_op support attrs from input.

* Fix python api.
diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc
@@ -34,9 +34,22 @@ class BilinearInterpOp : public framework::OperatorWithKernel {
     int out_w = ctx->Attrs().Get<int>("out_w");
     PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4");
 
+    if (ctx->HasInput("OutSize")) {
+      auto out_size_dim = ctx->GetInputDim("OutSize");
+      PADDLE_ENFORCE_EQ(out_size_dim.size(), 1,
+                        "OutSize's dimension size must be 1");
+      PADDLE_ENFORCE_EQ(out_size_dim[0], 2, "OutSize's dim[0] must be 2");
+    }
     std::vector<int64_t> dim_out({dim_x[0], dim_x[1], out_h, out_w});
     ctx->SetOutputDim("Out", framework::make_ddim(dim_out));
   }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace());
+  }
 };
 
 class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -45,6 +58,10 @@ class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X",
              "(Tensor) The input tensor of bilinear interpolation, "
              "This is a 4-D tensor with shape of (N x C x h x w)");
+    AddInput("OutSize",
+             "(Tensor) This is a 1-D tensor with two number. "
+             "The first number is height and the second number is width.")
+        .AsDispensable();
     AddOutput("Out",
               "(Tensor) The dimension of output is (N x C x out_h x out_w]");
 
@@ -78,6 +95,12 @@ class BilinearInterpOpGrad : public framework::OperatorWithKernel {
       ctx->SetOutputDim(framework::GradVarName("X"), dim_x);
     }
   }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("X")->type()), ctx.GetPlace());
+  }
 };
 
 }  // namespace operators
diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu
@@ -102,10 +102,21 @@ class BilinearInterpOpCUDAKernel : public framework::OpKernel<T> {
     auto* input_t = ctx.Input<Tensor>("X");      // float tensor
     auto* output_t = ctx.Output<Tensor>("Out");  // float tensor
     auto* input = input_t->data<T>();
-    auto* output = output_t->mutable_data<T>(ctx.GetPlace());
 
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+    auto out_dims = output_t->dims();
+    auto out_size_t = ctx.Input<Tensor>("OutSize");
+    if (out_size_t != nullptr) {
+      Tensor sizes;
+      framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes);
+      auto size_data = sizes.data<int>();
+      out_h = size_data[0];
+      out_w = size_data[1];
+    }
+    auto* output = output_t->mutable_data<T>(
+        {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace());
+
     int batch_size = input_t->dims()[0];
     int channels = input_t->dims()[1];
     int in_h = input_t->dims()[2];
@@ -139,8 +150,8 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* d_input_t = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto* d_output_t = ctx.Input<Tensor>(framework::GradVarName("Out"));
-    auto* d_input = d_input_t->mutable_data<T>(ctx.GetPlace());
     auto* d_output = d_output_t->data<T>();
+    auto* d_input = d_input_t->mutable_data<T>(ctx.GetPlace());
 
     auto& device_ctx =
         ctx.template device_context<platform::CUDADeviceContext>();
@@ -149,6 +160,16 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel<T> {
 
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+
+    auto out_size_t = ctx.Input<Tensor>("OutSize");
+    if (out_size_t != nullptr) {
+      Tensor sizes;
+      framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes);
+      auto size_data = sizes.data<int>();
+      out_h = size_data[0];
+      out_w = size_data[1];
+    }
+
     int batch_size = d_input_t->dims()[0];
     int channels = d_input_t->dims()[1];
     int in_h = d_input_t->dims()[2];
diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h
@@ -24,11 +24,18 @@ class BilinearInterpKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* input_t = ctx.Input<Tensor>("X");      // float tensor
     auto* output_t = ctx.Output<Tensor>("Out");  // float tensor
+    auto out_dims = output_t->dims();
     auto* input = input_t->data<T>();
-    auto* output = output_t->mutable_data<T>(ctx.GetPlace());
-
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+    auto out_size_t = ctx.Input<Tensor>("OutSize");
+    if (out_size_t != nullptr) {
+      auto out_size_data = out_size_t->data<int>();
+      out_h = out_size_data[0];
+      out_w = out_size_data[1];
+    }
+    auto* output = output_t->mutable_data<T>(
+        {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace());
     int batch_size = input_t->dims()[0];
     int channels = input_t->dims()[1];
     int in_h = input_t->dims()[2];
@@ -83,16 +90,23 @@ class BilinearInterpGradKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* d_input_t = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto* d_output_t = ctx.Input<Tensor>(framework::GradVarName("Out"));
-    auto* d_input = d_input_t->mutable_data<T>(ctx.GetPlace());
     auto* d_output = d_output_t->data<T>();
-
+    auto* d_input = d_input_t->mutable_data<T>(ctx.GetPlace());
     auto& device_ctx =
         ctx.template device_context<platform::CPUDeviceContext>();
     math::SetConstant<platform::CPUDeviceContext, T> zero;
     zero(device_ctx, d_input_t, static_cast<T>(0.0));
 
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
+
+    auto out_size_t = ctx.Input<Tensor>("OutSize");
+    if (out_size_t != nullptr) {
+      auto out_size_data = out_size_t->data<int>();
+      out_h = out_size_data[0];
+      out_w = out_size_data[1];
+    }
+
     int batch_size = d_input_t->dims()[0];
     int channels = d_input_t->dims()[1];
     int in_h = d_input_t->dims()[2];
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
@@ -3944,7 +3944,7 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None):
         input (Variable): The input tensor of bilinear interpolation,
                           This is a 4-D tensor of the shape
                           (num_batches, channels, in_h, in_w).
-        out_shape(list|tuple|None): Output shape of bilinear interpolation
+        out_shape(list|tuple|Variable|None): Output shape of bilinear interpolation
                                     layer, the shape is (out_h, out_w).
                                     Default: None
         scale(int|None): The multiplier for the input height or width.
@@ -3971,21 +3971,28 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None):
     def _is_list_or_turple_(data):
         return (isinstance(data, list) or isinstance(data, tuple))
 
+    out_h = 0
+    out_w = 0
+    inputs = {"X": input}
     if out_shape is not None:
-        if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2):
+        if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2) and (
+                out_shape is not Variable):
             raise ValueError('out_shape should be a list or tuple ',
                              'with length 2, (out_h, out_w).')
-        out_shape = list(map(int, out_shape))
-        out_h = out_shape[0]
-        out_w = out_shape[1]
+        if _is_list_or_turple_(out_shape):
+            out_shape = list(map(int, out_shape))
+            out_h = out_shape[0]
+            out_w = out_shape[1]
+        else:
+            inputs['OutSize'] = out_shape
     else:
         out_h = int(input.shape[2] * scale)
         out_w = int(input.shape[3] * scale)
 
     out = helper.create_tmp_variable(dtype)
     helper.append_op(
         type="bilinear_interp",
-        inputs={"X": input},
+        inputs=inputs,
         outputs={"Out": out},
         attrs={"out_h": out_h,
                "out_w": out_w})
diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py
@@ -17,7 +17,10 @@
 from op_test import OpTest
 
 
-def bilinear_interp_np(input, out_h, out_w):
+def bilinear_interp_np(input, out_h, out_w, out_size):
+    if out_size is not None:
+        out_h = out_size[0]
+        out_w = out_size[1]
     batch_size, channel, in_h, in_w = input.shape
     if out_h > 1:
         ratio_h = (in_h - 1.0) / (out_h - 1.0)
@@ -49,12 +52,15 @@ def bilinear_interp_np(input, out_h, out_w):
 
 class TestBilinearInterpOp(OpTest):
     def setUp(self):
+        self.out_size = None
         self.init_test_case()
         self.op_type = "bilinear_interp"
         input_np = np.random.random(self.input_shape).astype("float32")
-        output_np = bilinear_interp_np(input_np, self.out_h, self.out_w)
-
+        output_np = bilinear_interp_np(input_np, self.out_h, self.out_w,
+                                       self.out_size)
         self.inputs = {'X': input_np}
+        if self.out_size is not None:
+            self.inputs['OutSize'] = self.out_size
         self.attrs = {'out_h': self.out_h, 'out_w': self.out_w}
         self.outputs = {'Out': output_np}
 
@@ -68,6 +74,7 @@ def init_test_case(self):
         self.input_shape = [2, 3, 4, 4]
         self.out_h = 2
         self.out_w = 2
+        self.out_size = np.array([3, 3]).astype("int32")
 
 
 class TestCase1(TestBilinearInterpOp):
@@ -91,5 +98,29 @@ def init_test_case(self):
         self.out_w = 128
 
 
+class TestCase4(TestBilinearInterpOp):
+    def init_test_case(self):
+        self.input_shape = [4, 1, 7, 8]
+        self.out_h = 1
+        self.out_w = 1
+        self.out_size = np.array([2, 2]).astype("int32")
+
+
+class TestCase5(TestBilinearInterpOp):
+    def init_test_case(self):
+        self.input_shape = [3, 3, 9, 6]
+        self.out_h = 12
+        self.out_w = 12
+        self.out_size = np.array([11, 11]).astype("int32")
+
+
+class TestCase6(TestBilinearInterpOp):
+    def init_test_case(self):
+        self.input_shape = [1, 1, 128, 64]
+        self.out_h = 64
+        self.out_w = 128
+        self.out_size = np.array([65, 129]).astype("int32")
+
+
 if __name__ == "__main__":
     unittest.main()
