Merge pull request #5718 from chengduoZH/Add_pad_and_stride_for_conv_transpose_op

Add padding for conv_transpose_op

Author: chengduo

paddle/operators/conv_transpose_op.cc

@@ -30,11 +30,6 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
   std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
   std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
 
-  for (size_t i = 0; i < paddings.size(); ++i) {
-    PADDLE_ENFORCE_EQ(paddings[i], 0,
-                      "No Padding allowed in conv transpose op.");
-  }
-
   PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5,
                  "ConvTransposeOp intput should be 4-D or 5-D tensor.");
   PADDLE_ENFORCE_EQ(in_dims.size(), filter_dims.size(),
@@ -52,7 +47,7 @@ void ConvTransposeOp::InferShape(framework::InferShapeContext* ctx) const {
 
   std::vector<int64_t> output_shape({in_dims[0], filter_dims[1]});
   for (size_t i = 0; i < strides.size(); ++i) {
-    output_shape.push_back((in_dims[i + 2] - 1) * strides[i] +
+    output_shape.push_back((in_dims[i + 2] - 1) * strides[i] - 2 * paddings[i] +
                            filter_dims[i + 2]);
   }
   ctx->SetOutputDim("Output", framework::make_ddim(output_shape));

paddle/operators/conv_transpose_op.h

@@ -62,7 +62,6 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
     Tensor* output = context.Output<Tensor>("Output");
 
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
-    // Actually, no paddings and groups allowed in conv transpose.
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
     // TODO(Zhuoyuan): Paddings can be added in future.
     // groups will alway be disabled in conv2dtranspose.
@@ -148,8 +147,8 @@ class GemmConvTransposeKernel : public framework::OpKernel<T> {
       } else if (filter_shape_vec.size() == 3) {
         // col2vol: col_matrix -> dy
         // from (c * k_d * k_h * k_w, d * h * w) to (c, o_d, o_h, o_w)
-        col2vol(context.device_context(), col, dilations, strides,
-                std::vector<int>{0, 0, 0}, &output_batch);
+        col2vol(context.device_context(), col, dilations, strides, paddings,
+                &output_batch);
       }
     }
   }
@@ -173,7 +172,6 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
     if ((!input_grad) && (!filter_grad)) return;
 
     std::vector<int> strides = context.Attr<std::vector<int>>("strides");
-    // Actually, no paddings and groups allowed in conv transpose.
     std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
 
     const int batch_size = static_cast<int>(input->dims()[0]);

paddle/operators/math/im2col.cu

@@ -119,8 +119,8 @@ __global__ void col2im(int n, const T* data_col, int im_height, int im_width,
 
   if (index < n) {
     T val = 0;
-    int w = index % im_width;
-    int h = (index / im_width) % im_height;
+    int w = index % im_width + padding_width;
+    int h = (index / im_width) % im_height + padding_height;
     int c = index / (im_width * im_height);
 
     // compute the start and end of the output

python/paddle/v2/fluid/tests/test_conv2d_op.py

@@ -110,13 +110,30 @@ def init_op_type(self):
         self.op_type = "conv2d"
 
 
+class TestWithPad(TestConv2dOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        assert np.mod(self.input_size[1], self.groups) == 0
+        f_c = self.input_size[1] / self.groups
+        self.filter_size = [6, f_c, 3, 3]
+
+
+class TestWithStride(TestConv2dOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [2, 2]
+        self.input_size = [2, 3, 6, 6]  # NCHW
+        assert np.mod(self.input_size[1], self.groups) == 0
+        f_c = self.input_size[1] / self.groups
+        self.filter_size = [6, f_c, 3, 3]
+
+
 class TestWithGroup(TestConv2dOp):
     def init_group(self):
         self.groups = 3
 
-    def init_op_type(self):
-        self.op_type = "conv2d"
-
 
 class TestWith1x1(TestConv2dOp):
     def init_test_case(self):
@@ -127,15 +144,9 @@ def init_test_case(self):
         f_c = self.input_size[1] / self.groups
         self.filter_size = [6, f_c, 1, 1]
 
-    def init_dilation(self):
-        self.dilations = [1, 1]
-
     def init_group(self):
         self.groups = 3
 
-    def init_op_type(self):
-        self.op_type = "conv2d"
-
 
 class TestWithDilation(TestConv2dOp):
     def init_test_case(self):
@@ -152,14 +163,19 @@ def init_dilation(self):
     def init_group(self):
         self.groups = 3
 
+
+#----------------Conv2dCudnn----------------
+class TestCudnn(TestConv2dOp):
     def init_op_type(self):
-        self.op_type = "conv2d"
+        self.op_type = "conv_cudnn"
 
 
-#----------------Conv2dCudnn----------------
+class TestCudnnWithPad(TestWithPad):
+    def init_op_type(self):
+        self.op_type = "conv_cudnn"
 
 
-class TestCudnn(TestConv2dOp):
+class TestCudnnWithStride(TestWithStride):
     def init_op_type(self):
         self.op_type = "conv_cudnn"
 

python/paddle/v2/fluid/tests/test_conv2d_transpose_op.py

@@ -4,9 +4,7 @@
 
 
 def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param):
-    # [2, 3, 5, 5]
     in_n, in_c, in_h, in_w = input_.shape
-    # [3, 6, 3, 3]
     f_c, out_c, f_h, f_w = filter_.shape
     assert in_c == f_c
 
@@ -29,15 +27,14 @@ def conv2dtranspose_forward_naive(input_, filter_, conv2dtranspose_param):
                     j1, j2 = j * stride[0], j * stride[0] + f_w
                     out[n, k, i1:i2, j1:j2] += tmp_out
 
+    out = out[:, :, pad[0]:out_h - pad[0], pad[1]:out_w - pad[1]]
     return out
 
 
 class TestConv2dTransposeOp(OpTest):
     def setUp(self):
         # init as conv transpose
         self.init_op_type()
-
-        # [2, 3, 5, 5] -> kernel [3, 6, 3, 3] -> output [2, 6, 7, 7]
         self.init_test_case()
 
         conv2dtranspose_param = {'stride': self.stride, 'pad': self.pad}
@@ -55,7 +52,6 @@ def setUp(self):
         self.outputs = {'Output': output}
 
     def test_check_output(self):
-        print 'check output here for', self.op_type
         self.check_output()
 
     def test_check_grad_no_input(self):
@@ -88,6 +84,26 @@ def init_op_type(self):
         self.op_type = "conv2d_transpose"
 
 
+class TestWithPad(TestConv2dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [1, 1]
+        self.dilations = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3]
+
+
+class TestWithStride(TestConv2dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1]
+        self.stride = [2, 2]
+        self.dilations = [1, 1]
+        self.input_size = [2, 3, 5, 5]  # NCHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3]
+
+
 # ------------ test_cudnn ------------
 class TestCudnn(TestConv2dTransposeOp):
     def init_op_type(self):

python/paddle/v2/fluid/tests/test_conv3d_transpose_op.py

@@ -4,17 +4,14 @@
 
 
 def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
-    # [2, 3, 5, 5, 5]
     in_n, in_c, in_d, in_h, in_w = input_.shape
-    # [3, 6, 3, 3, 3]
     f_c, out_c, f_d, f_h, f_w = filter_.shape
     assert in_c == f_c
 
     stride, pad = conv3dtranspose_param['stride'], conv3dtranspose_param['pad']
     out_d = (in_d - 1) * stride[0] + f_d
     out_h = (in_h - 1) * stride[1] + f_h
     out_w = (in_w - 1) * stride[2] + f_w
-
     out = np.zeros((in_n, out_c, out_d, out_h, out_w))
 
     for n in range(in_n):
@@ -33,23 +30,22 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
                         j1, j2 = j * stride[2], j * stride[2] + f_w
                         out[n, k, d1:d2, i1:i2, j1:j2] += tmp_out
 
+    out = out[:, :, pad[0]:out_d - pad[0], pad[1]:out_h - pad[1], pad[2]:out_w -
+              pad[2]]
     return out
 
 
 class TestConv3dTransposeOp(OpTest):
     def setUp(self):
         # init as conv transpose
         self.init_op_type()
-
-        # [2, 3, 5, 5, 5] -> kernel [3, 6, 3, 3, 3] -> output [2, 6, 7, 7, 7]
         self.init_test_case()
 
         conv3dtranspose_param = {'stride': self.stride, 'pad': self.pad}
         input_ = np.random.random(self.input_size).astype("float32")
         filter_ = np.random.random(self.filter_size).astype("float32")
         output = conv3dtranspose_forward_naive(
             input_, filter_, conv3dtranspose_param).astype("float32")
-        # print 'deconv output py', output, output.shape
 
         self.inputs = {'Input': input_, 'Filter': filter_}
         self.attrs = {
@@ -60,7 +56,6 @@ def setUp(self):
         self.outputs = {'Output': output}
 
     def test_check_output(self):
-        print 'check output here'
         self.check_output()
 
     def test_check_grad(self):
@@ -85,13 +80,33 @@ def init_test_case(self):
         self.pad = [0, 0, 0]
         self.stride = [1, 1, 1]
         self.dilations = [1, 1, 1]
-        self.input_size = [2, 3, 5, 5, 5]  # NCHW
+        self.input_size = [2, 3, 5, 5, 5]  # NCDHW
         f_c = self.input_size[1]
         self.filter_size = [f_c, 6, 3, 3, 3]
 
     def init_op_type(self):
         self.op_type = "conv3d_transpose"
 
 
+class TestWithPad(TestConv3dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1, 1]
+        self.stride = [1, 1, 1]
+        self.dilations = [1, 1, 1]
+        self.input_size = [2, 3, 5, 5, 5]  # NCDHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3, 3]
+
+
+class TestWithStride(TestConv3dTransposeOp):
+    def init_test_case(self):
+        self.pad = [1, 1, 1]
+        self.stride = [2, 2, 2]
+        self.dilations = [1, 1, 1]
+        self.input_size = [2, 3, 5, 5, 5]  # NCDHW
+        f_c = self.input_size[1]
+        self.filter_size = [f_c, 6, 3, 3, 3]
+
+
 if __name__ == '__main__':
     unittest.main()