feat(mge/functional): add param output_padding for deconv ops

GitOrigin-RevId: 8a69608953a69b40db4d23b489435a7ae03c9523

imperative/python/megengine/functional/nn.py

@@ -335,6 +335,7 @@ def conv_transpose2d(
     bias: Optional[Tensor] = None,
     stride: Union[int, Tuple[int, int]] = 1,
     padding: Union[int, Tuple[int, int]] = 0,
+    output_padding: Union[int, Tuple[int, int]] = 0,
     dilation: Union[int, Tuple[int, int]] = 1,
     groups: int = 1,
     conv_mode="cross_correlation",
@@ -352,6 +353,7 @@ def conv_transpose2d(
         stride: stride of the 2D convolution operation. Default: 1
         padding: size of the paddings added to the input on both sides of its
             spatial dimensions. Only zero-padding is supported. Default: 0
+        output_padding: size of paddings appended to output. Default: 0
         dilation: dilation of the 2D convolution operation. Default: 1
         groups: number of groups into which the input and output channels are divided,
             so as to perform a ``grouped convolution``. When ``groups`` is not 1,
@@ -374,6 +376,7 @@ def conv_transpose2d(
 
     stride_h, stride_w = expand_hw(stride)
     pad_h, pad_w = expand_hw(padding)
+    output_pad_h, output_pad_w = expand_hw(output_padding)
     dilate_h, dilate_w = expand_hw(dilation)
 
     compute_mode = _config._get_actual_op_param(compute_mode, _config.__compute_mode)
@@ -389,6 +392,31 @@ def conv_transpose2d(
         compute_mode=compute_mode,
         sparse=sparse_type,
     )
+    if output_pad_h != 0 or output_pad_h != 0:
+        assert (
+            output_pad_h < stride[0]
+        ), "output_padding[0] shoule be less than stride[0]"
+        assert (
+            output_pad_w < stride[1]
+        ), "output_padding[1] shoule be less than stride[1]"
+        Hout = (
+            (inp.shape[2] - 1) * stride[0]
+            - 2 * padding[0]
+            + dilation[0] * (weight.shape[2] - 1)
+            + output_pad_h
+            + 1
+        )
+        Wout = (
+            (inp.shape[3] - 1) * stride[1]
+            - 2 * padding[1]
+            + dilation[1] * (weight.shape[3] - 1)
+            + output_pad_w
+            + 1
+        )
+        output_shape = [inp.shape[0], weight.shape[1], Hout, Wout]
+        output_shape = astensor1d(output_shape)
+        (output,) = apply(op, weight, inp, output_shape)
+    else:
         (output,) = apply(op, weight, inp)
     if bias is not None:
         if amp._enabled:
@@ -528,6 +556,7 @@ def conv_transpose3d(
     bias: Optional[Tensor] = None,
     stride: Union[int, Tuple[int, int, int]] = 1,
     padding: Union[int, Tuple[int, int, int]] = 0,
+    output_padding: Union[int, Tuple[int, int, int]] = 0,
     dilation: Union[int, Tuple[int, int, int]] = 1,
     groups: int = 1,
 ) -> Tensor:
@@ -544,6 +573,7 @@ def conv_transpose3d(
         stride: stride of the 3D convolution operation. Default: 1
         padding: size of the paddings added to the input on all sides of its
             spatial dimensions. Only zero-padding is supported. Default: 0
+        output_padding: size of paddings appended to output. Default: 0
         dilation: dilation of the 3D convolution operation. Default: 1
         groups: number of groups into which the input and output channels are divided,
             so as to perform a ``grouped convolution``. When ``groups`` is not 1,
@@ -558,6 +588,7 @@ def conv_transpose3d(
     pad = expand_dhw(padding)
     stride = expand_dhw(stride)
     dilate = expand_dhw(dilation)
+    output_padding = expand_dhw(output_padding)
 
     sparse_type = "dense" if groups == 1 else "group"
     op = builtin.Convolution3DBackwardData(
@@ -573,6 +604,41 @@ def conv_transpose3d(
         strategy=get_execution_strategy(),
         sparse=sparse_type,
     )
+    if output_padding[0] != 0 or output_padding[1] != 0 or output_padding[2] != 0:
+        assert (
+            output_padding[0] < stride[0]
+        ), "output_padding[0] shoule be less than stride[0]"
+        assert (
+            output_padding[1] < stride[1]
+        ), "output_padding[1] shoule be less than stride[1]"
+        assert (
+            output_padding[2] < stride[2]
+        ), "output_padding[2] shoule be less than stride[2]"
+        Dout = (
+            (inp.shape[2] - 1) * stride[0]
+            - 2 * padding[0]
+            + dilation[0] * (weight.shape[2] - 1)
+            + output_padding[0]
+            + 1
+        )
+        Hout = (
+            (inp.shape[3] - 1) * stride[1]
+            - 2 * padding[1]
+            + dilation[1] * (weight.shape[3] - 1)
+            + output_padding[1]
+            + 1
+        )
+        Wout = (
+            (inp.shape[4] - 1) * stride[2]
+            - 2 * padding[2]
+            + dilation[2] * (weight.shape[4] - 1)
+            + output_padding[2]
+            + 1
+        )
+        output_shape = [inp.shape[0], weight.shape[1], Dout, Hout, Wout]
+        output_shape = astensor1d(output_shape)
+        (output,) = apply(op, weight, inp, output_shape)
+    else:
         (output,) = apply(op, weight, inp)
     if bias is not None:
         output += bias

imperative/python/megengine/functional/quantized.py

@@ -134,6 +134,7 @@ def conv_transpose2d(
     dtype=None,
     stride: Union[int, Tuple[int, int]] = 1,
     padding: Union[int, Tuple[int, int]] = 0,
+    output_padding: Union[int, Tuple[int, int]] = 0,
     dilation: Union[int, Tuple[int, int]] = 1,
     groups: int = 1,
     conv_mode="cross_correlation",
@@ -156,6 +157,7 @@ def conv_transpose2d(
         )
 
     pad_h, pad_w = _pair(padding)
+    output_pad_h, output_pad_w = _pair(output_padding)
     stride_h, stride_w = _pair_nonzero(stride)
     dilate_h, dilate_w = _pair_nonzero(dilation)
     compute_mode = _config._get_actual_op_param(compute_mode, _config.__compute_mode)
@@ -173,5 +175,30 @@ def conv_transpose2d(
         compute_mode=compute_mode,
         mode=conv_mode,
     )
+    if output_pad_h != 0 or output_pad_h != 0:
+        assert (
+            output_pad_h < stride[0]
+        ), "output_padding[0] shoule be less than stride[0]"
+        assert (
+            output_pad_w < stride[1]
+        ), "output_padding[1] shoule be less than stride[1]"
+        Hout = (
+            (inp.shape[2] - 1) * stride[0]
+            - 2 * padding[0]
+            + dilation[0] * (weight.shape[2] - 1)
+            + output_pad_h
+            + 1
+        )
+        Wout = (
+            (inp.shape[3] - 1) * stride[1]
+            - 2 * padding[1]
+            + dilation[1] * (weight.shape[3] - 1)
+            + output_pad_w
+            + 1
+        )
+        output_shape = [inp.shape[0], weight.shape[1], Hout, Wout]
+        output_shape = Tensor(output_shape)
+        (output,) = apply(op, weight, inp, output_shape)
+    else:
         (output,) = apply(op, weight, inp)
     return output

imperative/python/megengine/module/conv.py

@@ -30,6 +30,7 @@ class _ConvNd(Module):
         kernel_size: Union[int, Tuple[int, int]],
         stride: Union[int, Tuple[int, int]],
         padding: Union[int, Tuple[int, int]],
+        output_padding: Union[int, Tuple[int, int]],
         dilation: Union[int, Tuple[int, int]],
         groups: int,
         bias: bool = True,
@@ -45,6 +46,7 @@ class _ConvNd(Module):
         self.kernel_size = kernel_size
         self.stride = stride
         self.padding = padding
+        self.output_padding = output_padding
         self.dilation = dilation
         self.groups = groups
 
@@ -178,6 +180,7 @@ class Conv1d(_ConvNd):
             kernel_size,
             stride,
             padding,
+            0,
             dilation,
             groups,
             bias,
@@ -352,6 +355,7 @@ class Conv2d(_ConvNd):
             kernel_size,
             stride,
             padding,
+            0,
             dilation,
             groups,
             bias,
@@ -505,6 +509,7 @@ class Conv3d(_ConvNd):
             kernel_size,
             stride,
             padding,
+            0,
             dilation,
             groups,
             bias,
@@ -572,6 +577,7 @@ class ConvTranspose2d(_ConvNd):
         stride: stride of the 2D convolution operation. Default: 1
         padding: size of the paddings added to the input on both sides of its
             spatial dimensions. Only zero-padding is supported. Default: 0
+        output_padding: size of paddings appended to output. Default: 0
         dilation: dilation of the 2D convolution operation. Default: 1
         groups: number of groups into which the input and output channels are divided,
             so as to perform a ``grouped convolution``. When ``groups`` is not 1,
@@ -591,6 +597,8 @@ class ConvTranspose2d(_ConvNd):
         * ``bias`` usually has shape ``(1, out_channels, *1)``
     """
 
+    output_padding = 0
+
     def __init__(
         self,
         in_channels: int,
@@ -598,6 +606,7 @@ class ConvTranspose2d(_ConvNd):
         kernel_size: Union[int, Tuple[int, int]],
         stride: Union[int, Tuple[int, int]] = 1,
         padding: Union[int, Tuple[int, int]] = 0,
+        output_padding: Union[int, Tuple[int, int]] = 0,
         dilation: Union[int, Tuple[int, int]] = 1,
         groups: int = 1,
         bias: bool = True,
@@ -608,6 +617,7 @@ class ConvTranspose2d(_ConvNd):
         kernel_size = _pair_nonzero(kernel_size)
         stride = _pair_nonzero(stride)
         padding = _pair(padding)
+        output_padding = _pair(output_padding)
         dilation = _pair_nonzero(dilation)
         self.conv_mode = conv_mode
         self.compute_mode = compute_mode
@@ -617,6 +627,7 @@ class ConvTranspose2d(_ConvNd):
             kernel_size,
             stride,
             padding,
+            output_padding,
             dilation,
             groups,
             bias,
@@ -656,6 +667,7 @@ class ConvTranspose2d(_ConvNd):
             bias,
             self.stride,
             self.padding,
+            self.output_padding,
             self.dilation,
             self.groups,
             self.conv_mode,
@@ -817,6 +829,7 @@ class DeformableConv2d(_ConvNd):
             kernel_size,
             stride,
             padding,
+            0,
             dilation,
             groups,
             bias,
@@ -889,6 +902,7 @@ class ConvTranspose3d(_ConvNd):
         stride: stride of the 3D convolution operation. Default: 1
         padding: size of the paddings added to the input on all sides of its
             spatial dimensions. Only zero-padding is supported. Default: 0
+        output_padding: size of paddings appended to output. Default: 0
         dilation: dilation of the 3D convolution operation. Default: 1
         groups: number of groups into which the input and output channels are divided,
             so as to perform a ``grouped convolution``. When ``groups`` is not 1,
@@ -902,6 +916,8 @@ class ConvTranspose3d(_ConvNd):
         * ``bias`` usually has shape ``(1, out_channels, *1)``
     """
 
+    output_padding = 0
+
     def __init__(
         self,
         in_channels: int,
@@ -909,6 +925,7 @@ class ConvTranspose3d(_ConvNd):
         kernel_size: Union[int, Tuple[int, int, int]],
         stride: Union[int, Tuple[int, int, int]] = 1,
         padding: Union[int, Tuple[int, int, int]] = 0,
+        output_padding: Union[int, Tuple[int, int, int]] = 0,
         dilation: Union[int, Tuple[int, int, int]] = 1,
         groups: int = 1,
         bias: bool = True,
@@ -923,6 +940,7 @@ class ConvTranspose3d(_ConvNd):
             kernel_size=kernel_size,
             stride=stride,
             padding=padding,
+            output_padding=output_padding,
             dilation=dilation,
             groups=groups,
             bias=bias,
@@ -956,5 +974,11 @@ class ConvTranspose3d(_ConvNd):
 
     def forward(self, inp):
         return conv_transpose3d(
-            inp, self.weight, self.bias, self.stride, self.padding, self.dilation,
+            inp,
+            self.weight,
+            self.bias,
+            self.stride,
+            self.padding,
+            self.output_padding,
+            self.dilation,
         )

imperative/python/megengine/module/qat/conv.py

@@ -74,6 +74,7 @@ class ConvTranspose2d(Float.ConvTranspose2d, QATModule):
             float_module.kernel_size,
             float_module.stride,
             float_module.padding,
+            float_module.output_padding,
             float_module.dilation,
             float_module.groups,
             float_module.bias is not None,

imperative/python/megengine/module/quantized/conv.py

@@ -138,6 +138,8 @@ class ConvTranspose2d(Float.ConvTranspose2d, QuantizedModule):
         dtype: data type of the output, should be qint8.
     """
 
+    output_padding = 0
+
     def __init__(
         self,
         in_channels: int,
@@ -145,6 +147,7 @@ class ConvTranspose2d(Float.ConvTranspose2d, QuantizedModule):
         kernel_size: Union[int, Tuple[int, int]],
         stride: Union[int, Tuple[int, int]] = 1,
         padding: Union[int, Tuple[int, int]] = 0,
+        output_padding: Union[int, Tuple[int, int]] = 0,
         dilation: Union[int, Tuple[int, int]] = 1,
         groups: int = 1,
         bias: bool = True,
@@ -159,6 +162,7 @@ class ConvTranspose2d(Float.ConvTranspose2d, QuantizedModule):
             kernel_size=kernel_size,
             stride=stride,
             padding=padding,
+            output_padding=output_padding,
             dilation=dilation,
             groups=groups,
             bias=bias,
@@ -180,6 +184,7 @@ class ConvTranspose2d(Float.ConvTranspose2d, QuantizedModule):
             qat_module.kernel_size,
             qat_module.stride,
             qat_module.padding,
+            qat_module.output_padding,
             qat_module.dilation,
             qat_module.groups,
             qat_module.bias is not None,
@@ -212,6 +217,7 @@ class ConvTranspose2d(Float.ConvTranspose2d, QuantizedModule):
             dtype=self.output_dtype,
             stride=self.stride,
             padding=self.padding,
+            output_padding=self.output_padding,
             dilation=self.dilation,
             groups=self.groups,
             conv_mode=self.conv_mode,

imperative/python/test/unit/functional/test_functional.py

@@ -18,7 +18,8 @@ from megengine.core._trace_option import use_symbolic_shape
 from megengine.core.autodiff.grad import Grad
 from megengine.core.tensor.utils import make_shape_tuple
 from megengine.device import get_device_count
-from megengine.module import LayerNorm
+from megengine.jit.tracing import trace
+from megengine.module import ConvTranspose2d, ConvTranspose3d, LayerNorm
 
 _assert_allclose = partial(np.testing.assert_allclose, atol=5e-6, rtol=5e-6)
 
@@ -1374,3 +1375,37 @@ def test_local_conv2d(stride, padding, dilation, ksize, groups):
     )
     ref = local_conv2d_np(data, weight, stride, padding, dilation)
     np.testing.assert_almost_equal(output.numpy(), ref, 5)
+
+
+def test_conv_transpose2d():
+    m = ConvTranspose2d(
+        16, 33, (3, 5), output_padding=(1, 2), stride=(2, 3), padding=(4, 2)
+    )
+
+    @trace(symbolic=True)
+    def fwd(inp: Tensor):
+        return m(inp)
+
+    input = Tensor(np.random.rand(20, 16, 50, 100))
+    output = fwd(input)
+    output_shape = Tensor(output.shape)
+    np.testing.assert_equal(
+        output_shape.numpy(), np.array([20, 33, 94, 300], dtype=np.int32)
+    )
+
+
+def test_conv_transpose3d():
+    m = ConvTranspose3d(
+        16, 33, (3, 5, 2), output_padding=(2, 1, 1), stride=(3, 2, 2), padding=(0, 4, 2)
+    )
+
+    @trace(symbolic=True)
+    def fwd(inp: Tensor):
+        return m(inp)
+
+    input = Tensor(np.random.rand(20, 16, 10, 50, 100))
+    output = fwd(input)
+    output_shape = Tensor(output.shape)
+    np.testing.assert_equal(
+        output_shape.numpy(), np.array([20, 33, 32, 96, 197], dtype=np.int32)
+    )

imperative/src/impl/ops/convolution.cpp

@@ -5,6 +5,7 @@
 #include "../op_trait.h"
 #include "megbrain/imperative/ops/autogen.h"
 #include "megbrain/opr/internal/megdnn_opr_wrapper.h"
+#include "megbrain/opr/tensor_gen.h"
 
 namespace mgb {
 namespace imperative {
@@ -152,8 +153,11 @@ auto apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
                 inputs[0], inputs[1], conv.param(), conv.policy(), config);
     } else {
         mgb_assert(inputs.size() == 3);
+        auto* src_for_shape =
+                opr::Alloc::make(inputs[2], inputs[0]->dtype(), {}).node();
         return opr::ConvolutionBackwardData::make(
-                inputs[0], inputs[1], inputs[2], conv.param(), conv.policy(), config);
+                inputs[0], inputs[1], src_for_shape, conv.param(), conv.policy(),
+                config);
     }
 }
 
@@ -168,6 +172,14 @@ std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
     if (filter.ndim && diff.ndim) {
         // deduce_layout won't override existing dtype
         dnn_opr.opr().deduce_layout(filter, diff, output_layout);
+        if (inputs.size() == 3) {
+            if (!inputs[2].value.empty()) {
+                cg::copy_tensor_value_to_shape(output_layout, inputs[2].value);
+                output_layout.init_contiguous_stride();
+            } else {
+                output_layout.ndim = 0;
+            }
+        }
     }
     return {{{output_layout, inputs[0].comp_node}}, output_layout.ndim != 0};
 }
@@ -185,8 +197,11 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
             return output_descs[0].layout;
         } else {
             TensorLayout out_layout{inputs[0]->dtype()};
-            dnn_opr.op()->deduce_layout(
-                    inputs[0]->layout(), inputs[1]->layout(), out_layout);
+            if (inputs.size() == 3) {
+                cg::copy_tensor_value_to_shape(
+                        out_layout, inputs[2]->get_value().proxy_to_default_cpu());
+                out_layout.init_contiguous_stride();
+            }
             return out_layout;
         }
     }();
@@ -263,50 +278,74 @@ namespace convolution3d_backward_data {
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def, const SmallVector<LogicalTensorDesc>& inputs) {
     mgb_assert(
-            inputs.size() == 2,
-            "inputs num of conv_transpose3d should be 2 but you give %zu",
+            inputs.size() == 2 || inputs.size() == 3,
+            "inputs num of conv_transpose3d should be 2 or 3 but you give %zu",
             inputs.size());
-    auto&& op_def = def.cast_final_safe<Convolution3DBackwardData>();
-    auto&& weight = inputs[0];
+    auto&& conv3dbwd = def.cast_final_safe<Convolution3DBackwardData>();
+    DnnOprHelper<megdnn::Convolution3DBackwardData> dnn_opr(conv3dbwd.param());
+    auto&& filter = inputs[0];
     auto&& diff = inputs[1];
-    if (!(weight.layout.ndim && diff.layout.ndim)) {
-        return {{{TensorLayout{weight.layout.dtype}, weight.comp_node}}, false};
+
+    if (!(filter.layout.ndim && diff.layout.ndim)) {
+        return {{{TensorLayout{filter.layout.dtype}, filter.comp_node}}, false};
     }
-    DnnOprHelper<megdnn::Convolution3DBackwardData> dnn_opr(op_def.param());
-    auto oup_layout = dnn_opr.deduce_layout(weight.layout, diff.layout);
-    return {{{oup_layout, weight.comp_node}}, true};
+
+    TensorLayout output_layout = dnn_opr.deduce_layout(filter.layout, diff.layout);
+    if (filter.layout.ndim && diff.layout.ndim) {
+        if (inputs.size() == 3) {
+            if (!inputs[2].value.empty()) {
+                cg::copy_tensor_value_to_shape(output_layout, inputs[2].value);
+                output_layout.init_contiguous_stride();
+            } else {
+                output_layout.ndim = 0;
+            }
+        }
+    }
+    return {{{output_layout, inputs[0].comp_node}}, output_layout.ndim != 0};
 }
 
 SmallVector<TensorPtr> apply_on_physical_tensor(
         const OpDef& def, const SmallVector<TensorPtr>& inputs,
         SmallVector<LogicalTensorDesc>& output_descs, const bool& validated) {
-    auto&& conv = def.cast_final_safe<Convolution3DBackwardData>();
-    auto cn = inputs[0]->comp_node();
-
-    auto&& wlayout = inputs[0]->layout();
-    auto&& dlayout = inputs[1]->layout();
-
-    DnnOprCaller<megdnn::Convolution3DBackwardData> dnn_op(
-            cn, conv.param(), conv.policy());
-
-    auto oup_layout = [&] {
+    auto&& conv3dbwd = def.cast_final_safe<Convolution3DBackwardData>();
+    CompNode cn = inputs[0]->comp_node();
+    DnnOprCaller<megdnn::Convolution3DBackwardData> dnn_opr(
+            cn, conv3dbwd.param(), conv3dbwd.policy());
+    auto out_layout = [&] {
         if (validated) {
             return output_descs[0].layout;
         } else {
-            return dnn_op.deduce_layout(wlayout, dlayout);
+            TensorLayout out_layout{inputs[0]->dtype()};
+            dnn_opr.op()->deduce_layout(
+                    inputs[0]->layout(), inputs[1]->layout(), out_layout);
+            if (inputs.size() == 3) {
+                cg::copy_tensor_value_to_shape(
+                        out_layout, inputs[2]->get_value().proxy_to_default_cpu());
+                out_layout.init_contiguous_stride();
+            }
+            return out_layout;
         }
     }();
-    auto oup = Tensor::make(oup_layout, cn);
-    dnn_op.exec_fastrun(inputs[0], inputs[1], oup);
-    return {oup};
+    auto out = Tensor::make(out_layout, cn);
+    dnn_opr.exec_fastrun(inputs[0], inputs[1], out);
+    return {out};
 }
 
 auto apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
     auto&& conv = static_cast<const Convolution3DBackwardData&>(def);
     OperatorNodeConfig config{conv.make_name()};
-    mgb_assert(inputs.size() == 2);
+    if (inputs.size() == 2) {
         return opr::Convolution3DBackwardData::make(
                 inputs[0], inputs[1], conv.param(), conv.policy(), config);
+    } else {
+        mgb_assert(inputs.size() == 3);
+        // The output shape is calculated in advance and given as input
+        auto* src_for_shape =
+                opr::Alloc::make(inputs[2], inputs[0]->dtype(), {}).node();
+        return opr::Convolution3DBackwardData::make(
+                inputs[0], inputs[1], src_for_shape, conv.param(), conv.policy(),
+                config);
+    }
 }
 
 OP_TRAIT_REG(Convolution3DBackwardData, Convolution3DBackwardData)