feat(imperative/amp): add dimshuffle before creating nhwc tensor

GitOrigin-RevId: 4461f9a0d3

imperative/python/megengine/amp/convert_format.py

@@ -33,7 +33,7 @@ def convert_tensor_format(x: Tensor, inplace: bool = True):
     if x.format != "nhwc":
         if inplace:
             # hostvalue should still be valid, so no d2h cost.
-            data = x.numpy().transpose(*pattern)
+            data = x.numpy()
             # reset will destroy existed backward grad
             x[...] = Tensor(data, format="nhwc")
         else:

imperative/python/test/unit/core/test_formatted_tensor.py

@@ -38,7 +38,7 @@ def test_basic():
 
 def _compare_nchw_nhwc(data, func, is_symbolic=None):
     x1 = tensor(data)
-    x2 = tensor(data.transpose(0, 2, 3, 1), format="nhwc")
+    x2 = tensor(data, format="nhwc")
     if is_symbolic is not None:
         func = trace(func, symbolic=is_symbolic)
     out1 = func(x1)
@@ -247,8 +247,8 @@ def test_conv2d(is_symbolic):
         if x.format == "nhwc":
             x = F.conv2d(
                 x,
-                weight=mge.tensor(np.ones((3, 1, 1, 2)), format="nhwc"),
-                bias=mge.tensor(np.ones((1, 1, 1, 3)), format="nhwc"),
+                weight=mge.tensor(np.ones((3, 2, 1, 1)), format="nhwc"),
+                bias=mge.tensor(np.ones((1, 3, 1, 1)), format="nhwc"),
             )
             assert x.format == "nhwc"
             return x.numpy()
@@ -265,8 +265,8 @@ def test_group_conv2d(is_symbolic):
         if x.format == "nhwc":
             x = F.conv2d(
                 x,
-                weight=mge.tensor(np.ones((2, 2, 1, 1, 2)), format="nhwc"),
-                bias=mge.tensor(np.ones((1, 1, 1, 4)), format="nhwc"),
+                weight=mge.tensor(np.ones((2, 2, 2, 1, 1)), format="nhwc"),
+                bias=mge.tensor(np.ones((1, 4, 1, 1)), format="nhwc"),
                 groups=2,
             )
             assert x.format == "nhwc"
@@ -286,10 +286,10 @@ def test_bn(is_symbolic):
         if x.format == "nhwc":
             oups = F.batch_norm(
                 x.astype("float32"),
-                running_mean=mge.tensor(np.ones((1, 1, 1, 2)), format="nhwc"),
-                running_var=mge.tensor(np.ones((1, 1, 1, 2)), format="nhwc"),
-                weight=mge.tensor(np.ones((1, 1, 1, 2)), format="nhwc"),
-                bias=mge.tensor(np.ones((1, 1, 1, 2)), format="nhwc"),
+                running_mean=mge.tensor(np.ones((1, 2, 1, 1)), format="nhwc"),
+                running_var=mge.tensor(np.ones((1, 2, 1, 1)), format="nhwc"),
+                weight=mge.tensor(np.ones((1, 2, 1, 1)), format="nhwc"),
+                bias=mge.tensor(np.ones((1, 2, 1, 1)), format="nhwc"),
                 training=True,
                 inplace=False,
             )

imperative/src/impl/transformations/format.cpp

@@ -27,7 +27,11 @@ TypedValueRef<FormattedTensorValue> FormatTransformation::to(
             pattern = {0, 3, 1, 2};
         }
     } else if ((format == FT::NCHW || format == FT::DEFAULT) && target == FT::NHWC) {
-        pattern = {0, 2, 3, 1};
+        if (tensor.value().shape().cast<ShapeValue>().ndim == 5) {
+            pattern = {0, 1, 3, 4, 2};
+        } else {
+            pattern = {0, 2, 3, 1};
+        }
     } else {
         mgb_throw(
                 MegBrainError, "Unsupport format conversion from %s to %s",
@@ -572,8 +576,13 @@ ValueRefList FormatTransformation::apply_transformation(
         }
     } else if (auto* create_tensor = op.as<CreateTensor>()) {
         auto format = create_tensor->format();
-        // TODO: add dimshuffle for nhwc format
-        return {wrap_output(imperative::apply(op, inputs)[0], format)};
+        if (format == FT::NHWC) {
+            auto output = wrap_output(imperative::apply(op, inputs)[0]);
+            output = to(output.cast(m_value_type), FT::NHWC, "");
+            return {output};
+        } else {
+            return {wrap_output(imperative::apply(op, inputs)[0], format)};
+        }
     } else if (auto* get_attr = op.as<GetAttr>()) {
         auto&& input = inputs.item();
         if (!input.is(m_value_type)) {