提交 eb070476 编写于 作者: C chengduoZH

add padding

上级 9cf60365
......@@ -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));
......
......@@ -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]);
......
......@@ -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
......
......@@ -110,13 +110,30 @@ class TestConv2dOp(OpTest):
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 @@ class TestWith1x1(TestConv2dOp):
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 @@ class TestWithDilation(TestConv2dOp):
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"
......
......@@ -4,9 +4,7 @@ from op_test import OpTest
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
......@@ -29,6 +27,7 @@ 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
......@@ -36,8 +35,6 @@ 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 @@ class TestConv2dTransposeOp(OpTest):
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 @@ class TestConv2dTransposeOp(OpTest):
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):
......
......@@ -4,9 +4,7 @@ from op_test import OpTest
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
......@@ -14,7 +12,6 @@ def conv3dtranspose_forward_naive(input_, filter_, conv3dtranspose_param):
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):
......@@ -33,6 +30,8 @@ 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
......@@ -40,8 +39,6 @@ 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}
......@@ -49,7 +46,6 @@ class TestConv3dTransposeOp(OpTest):
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 @@ class TestConv3dTransposeOp(OpTest):
self.outputs = {'Output': output}
def test_check_output(self):
print 'check output here'
self.check_output()
def test_check_grad(self):
......@@ -85,7 +80,7 @@ class TestConv3dTransposeOp(OpTest):
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]
......@@ -93,5 +88,25 @@ class TestConv3dTransposeOp(OpTest):
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()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册