Merge pull request #8651 from wanghaoshuang/fix_pool

Add ceil_mode option for pool2d and pool3d

paddle/fluid/operators/pool_op.cc

@@ -17,8 +17,15 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-int PoolOutputSize(int input_size, int filter_size, int padding, int stride) {
-  int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
+int PoolOutputSize(int input_size, int filter_size, int padding, int stride,
+                   bool ceil_mode) {
+  int output_size;
+  if (!ceil_mode) {
+    output_size = (input_size - filter_size + 2 * padding) / stride + 1;
+  } else {
+    output_size =
+        (input_size - filter_size + 2 * padding + stride - 1) / stride + 1;
+  }
   PADDLE_ENFORCE(output_size > 0,
                  "Due to the settings of padding(%d), filter_size(%d) and "
                  "stride(%d), the output size is less than 0, please check "
@@ -38,6 +45,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
   std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize");
   std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
   std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+  bool ceil_mode = ctx->Attrs().Get<bool>("ceil_mode");
 
   PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
                  "Pooling intput should be 4-D or 5-D tensor.");
@@ -59,8 +67,8 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const {
 
   std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
   for (size_t i = 0; i < ksize.size(); ++i) {
-    output_shape.push_back(
-        PoolOutputSize(in_x_dims[i + 2], ksize[i], paddings[i], strides[i]));
+    output_shape.push_back(PoolOutputSize(in_x_dims[i + 2], ksize[i],
+                                          paddings[i], strides[i], ceil_mode));
   }
   ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
   ctx->ShareLoD("X", "Out");
@@ -167,6 +175,12 @@ Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       "use_cudnn",
       "(bool, default false) Only used in cudnn kernel, need install cudnn")
       .SetDefault(false);
+  AddAttr<bool>(
+      "ceil_mode",
+      "(bool, default false) Wether to use the ceil function to calculate "
+      "output height and width. False is the default. If it is set to False, "
+      "the floor function will be used.")
+      .SetDefault(false);
   AddAttr<std::string>(
       "data_format",
       "(string, default NCHW) Only used in "
@@ -192,11 +206,16 @@ Example:
        X shape: $(N, C, H_{in}, W_{in})$
   Output:
        Out shape: $(N, C, H_{out}, W_{out})$
-  Where
+  For ceil_mode = false:
        $$
        H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
        W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
        $$
+  For ceil_mode = true:
+       $$
+       H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0] + strides[0] - 1)}{strides[0]} + 1 \\
+       W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1
+       $$
 
 )DOC");
 }
@@ -251,6 +270,12 @@ Pool3dOpMaker::Pool3dOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       "use_cudnn",
       "(bool, default false) Only used in cudnn kernel, need install cudnn")
       .SetDefault(false);
+  AddAttr<bool>(
+      "ceil_mode",
+      "(bool, default false) Wether to use the ceil function to calculate "
+      "output height and width. False is the default. If it is set to False, "
+      "the floor function will be used.")
+      .SetDefault(false);
   AddAttr<std::string>(
       "data_format",
       "(string, default NCHW) Only used in "
@@ -276,12 +301,18 @@ Example:
        X shape: $(N, C, D_{in}, H_{in}, W_{in})$
   Output:
        Out shape: $(N, C, D_{out}, H_{out}, W_{out})$
-  Where
+  For ceil_mode = false:
   $$
        D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
        H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\
        W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1
   $$
+  For ceil_mode = true:
+  $$
+       D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0] + strides[0] -1)}{strides[0]} + 1 \\
+       H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1] + strides[1] -1)}{strides[1]} + 1 \\
+       W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2] + strides[2] -1)}{strides[2]} + 1
+  $$
 
 )DOC");
 }

python/paddle/fluid/layers/nn.py

@@ -1438,6 +1438,7 @@ def pool2d(input,
            pool_padding=0,
            global_pooling=False,
            use_cudnn=True,
+           ceil_mode=False,
            name=None):
     """
     This function adds the operator for pooling in 2 dimensions, using the
@@ -1474,7 +1475,8 @@ def pool2d(input,
             "global_pooling": global_pooling,
             "strides": pool_stride,
             "paddings": pool_padding,
-            "use_cudnn": use_cudnn
+            "use_cudnn": use_cudnn,
+            "ceil_mode": ceil_mode
         })
 
     return pool_out

python/paddle/fluid/tests/unittests/test_pool2d_op.py

@@ -19,12 +19,21 @@ import paddle.fluid.core as core
 from op_test import OpTest
 
 
-def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0):
+def max_pool2D_forward_naive(x,
+                             ksize,
+                             strides,
+                             paddings,
+                             global_pool=0,
+                             ceil_mode=False):
     N, C, H, W = x.shape
     if global_pool == 1:
         ksize = [H, W]
-    H_out = (H - ksize[0] + 2 * paddings[0]) / strides[0] + 1
-    W_out = (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1
+    H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1
+             ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 *
+                                                   paddings[0]) / strides[0] + 1
+    W_out = (W - ksize[1] + 2 * paddings[1] + strides[1] - 1
+             ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 *
+                                                   paddings[1]) / strides[1] + 1
     out = np.zeros((N, C, H_out, W_out))
     for i in xrange(H_out):
         for j in xrange(W_out):
@@ -38,12 +47,21 @@ def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0):
     return out
 
 
-def avg_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=0):
+def avg_pool2D_forward_naive(x,
+                             ksize,
+                             strides,
+                             paddings,
+                             global_pool=0,
+                             ceil_mode=False):
     N, C, H, W = x.shape
     if global_pool == 1:
         ksize = [H, W]
-    H_out = (H - ksize[0] + 2 * paddings[0]) / strides[0] + 1
-    W_out = (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1
+    H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1
+             ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 *
+                                                   paddings[0]) / strides[0] + 1
+    W_out = (W - ksize[1] + 2 * paddings[1] + strides[1] - 1
+             ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 *
+                                                   paddings[1]) / strides[1] + 1
     out = np.zeros((N, C, H_out, W_out))
     for i in xrange(H_out):
         for j in xrange(W_out):
@@ -65,12 +83,13 @@ class TestPool2d_Op(OpTest):
         self.init_global_pool()
         self.init_op_type()
         self.init_pool_type()
+        self.init_ceil_mode()
         if self.global_pool:
             self.paddings = [0 for _ in range(len(self.paddings))]
         input = np.random.random(self.shape).astype("float32")
         output = self.pool2D_forward_naive(input, self.ksize, self.strides,
-                                           self.paddings,
-                                           self.global_pool).astype("float32")
+                                           self.paddings, self.global_pool,
+                                           self.ceil_mode).astype("float32")
         self.inputs = {'X': input}
 
         self.attrs = {
@@ -80,6 +99,7 @@ class TestPool2d_Op(OpTest):
             'pooling_type': self.pool_type,
             'global_pooling': self.global_pool,
             'use_cudnn': self.use_cudnn,
+            'ceil_mode': self.ceil_mode,
             'data_format': 'AnyLayout'  # TODO(dzhwinter) : should be fix latter
         }
 
@@ -116,6 +136,9 @@ class TestPool2d_Op(OpTest):
     def init_global_pool(self):
         self.global_pool = True
 
+    def init_ceil_mode(self):
+        self.ceil_mode = False
+
 
 class TestCase1(TestPool2d_Op):
     def init_test_case(self):
@@ -217,5 +240,25 @@ class TestCUDNNCase6(TestCase5):
         self.op_type = "pool2d"
 
 
+class TestCeilModeCase1(TestCUDNNCase1):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
+class TestCeilModeCase2(TestCUDNNCase2):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
+class TestCeilModeCase3(TestCase1):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
+class TestCeilModeCase4(TestCase2):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_pool3d_op.py

@@ -19,13 +19,24 @@ import paddle.fluid.core as core
 from op_test import OpTest
 
 
-def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0):
+def max_pool3D_forward_naive(x,
+                             ksize,
+                             strides,
+                             paddings,
+                             global_pool=0,
+                             ceil_mode=False):
     N, C, D, H, W = x.shape
     if global_pool == 1:
         ksize = [D, H, W]
-    D_out = (D - ksize[0] + 2 * paddings[0]) / strides[0] + 1
-    H_out = (H - ksize[1] + 2 * paddings[1]) / strides[1] + 1
-    W_out = (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1
+    D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1
+             ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 *
+                                                   paddings[0]) / strides[0] + 1
+    H_out = (H - ksize[1] + 2 * paddings[1] + strides[1] - 1
+             ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 *
+                                                   paddings[1]) / strides[1] + 1
+    W_out = (W - ksize[2] + 2 * paddings[2] + strides[2] - 1
+             ) / strides[2] + 1 if ceil_mode else (W - ksize[2] + 2 *
+                                                   paddings[2]) / strides[2] + 1
     out = np.zeros((N, C, D_out, H_out, W_out))
     for k in xrange(D_out):
         d_start = np.max((k * strides[0] - paddings[0], 0))
@@ -42,13 +53,24 @@ def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0):
     return out
 
 
-def avg_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=0):
+def avg_pool3D_forward_naive(x,
+                             ksize,
+                             strides,
+                             paddings,
+                             global_pool=0,
+                             ceil_mode=False):
     N, C, D, H, W = x.shape
     if global_pool == 1:
         ksize = [D, H, W]
-    D_out = (D - ksize[0] + 2 * paddings[0]) / strides[0] + 1
-    H_out = (H - ksize[1] + 2 * paddings[1]) / strides[1] + 1
-    W_out = (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1
+    D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1
+             ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 *
+                                                   paddings[0]) / strides[0] + 1
+    H_out = (H - ksize[1] + 2 * paddings[1] + strides[1] - 1
+             ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 *
+                                                   paddings[1]) / strides[1] + 1
+    W_out = (W - ksize[2] + 2 * paddings[2] + strides[2] - 1
+             ) / strides[2] + 1 if ceil_mode else (W - ksize[2] + 2 *
+                                                   paddings[2]) / strides[2] + 1
     out = np.zeros((N, C, D_out, H_out, W_out))
     for k in xrange(D_out):
         d_start = np.max((k * strides[0] - paddings[0], 0))
@@ -73,13 +95,14 @@ class TestPool3d_Op(OpTest):
         self.init_global_pool()
         self.init_op_type()
         self.init_pool_type()
+        self.init_ceil_mode()
 
         if self.global_pool:
             self.paddings = [0 for _ in range(len(self.paddings))]
         input = np.random.random(self.shape).astype("float32")
         output = self.pool3D_forward_naive(input, self.ksize, self.strides,
-                                           self.paddings,
-                                           self.global_pool).astype("float32")
+                                           self.paddings, self.global_pool,
+                                           self.ceil_mode).astype("float32")
         self.inputs = {'X': input}
 
         self.attrs = {
@@ -89,6 +112,7 @@ class TestPool3d_Op(OpTest):
             'pooling_type': self.pool_type,
             'global_pooling': self.global_pool,
             'use_cudnn': self.use_cudnn,
+            'ceil_mode': self.ceil_mode,
             'data_format': 'AnyLayout'  # TODO(dzhwinter) : should be fix latter
         }
 
@@ -125,6 +149,9 @@ class TestPool3d_Op(OpTest):
     def init_global_pool(self):
         self.global_pool = True
 
+    def init_ceil_mode(self):
+        self.ceil_mode = False
+
 
 class TestCase1(TestPool3d_Op):
     def init_test_case(self):
@@ -227,5 +254,25 @@ class TestCUDNNCase6(TestCase5):
         self.op_type = "pool3d"
 
 
+class TestCeilModeCase1(TestCUDNNCase1):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
+class TestCeilModeCase2(TestCUDNNCase2):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
+class TestCeilModeCase3(TestCase1):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
+class TestCeilModeCase4(TestCase2):
+    def init_ceil_mode(self):
+        self.ceil_mode = True
+
+
 if __name__ == '__main__':
     unittest.main()