Merge pull request #201 from Idein/support-convTranspose

Support convTranspose

runtime/README.md

@@ -61,6 +61,10 @@ This runtime supports only below operators.
 - ConcatFromSequence
 - Constant
 - ConstantOfShape
+- ConvTranspose
+  - support 2d only
+  - `group` should be 1
+  - `auto_pad` should be `"NOTSET"` (default value)
 - Cos
 - Cosh
 - DepthToSpace

runtime/onnion_runtime/__init__.py

@@ -19,6 +19,7 @@
 from .concatfromsequence import ConcatFromSequence  # noqa: F401
 from .constant import Constant  # noqa: F401
 from .constantofshape import ConstantOfShape  # noqa: F401
+from .convtranspose import ConvTranspose  # noqa: F401
 from .cos import Cos  # noqa: F401
 from .cosh import Cosh  # noqa: F401
 from .depthtospace import DepthToSpace  # noqa: F401

runtime/onnion_runtime/convtranspose.py

@@ -0,0 +1,106 @@
+from typing import Any, List, Optional
+
+import numpy as np
+
+from .error import RunError
+
+
+# https://github.com/onnx/onnx/blob/main/docs/Operators.md#ConvTranspose
+class ConvTranspose:
+    auto_pad: str
+    group: int
+    dilations: Optional[List[int]]
+    strides: Optional[List[int]]
+    kernel_shape: Optional[List[int]]
+    output_shape: Optional[List[int]]
+    output_padding: Optional[List[int]]
+    pads: Optional[List[int]]
+
+    def __init__(self, opset_version: int, **kwargs: Any):
+        self.version = opset_version
+        self.auto_pad = kwargs.get("auto_pad", "NOTSET")
+        self.dilations = kwargs.get("dilations", None)
+        self.group = kwargs.get("group", 1)
+        self.kernel_shape = kwargs.get("kernel_shape", None)
+        self.output_padding = kwargs.get("output_padding", None)
+        self.output_shape = kwargs.get("output_shape", None)
+        self.pads = kwargs.get("pads", None)
+        self.strides = kwargs.get("strides", None)
+
+    def run(self, x: np.ndarray, W: np.ndarray, b: Optional[np.ndarray] = None) -> List[np.ndarray]:
+        """
+        2D Convolution Transpose
+        input shapes:
+            x: [batch, in_ch, in_h, in_w]
+            W: [in_ch, out_ch/group, kernel_h, kernel_w]
+            b: [out_ch]
+        output shape:
+            [batch, out_ch, out_h, out_w]
+        """
+
+        # define parameters
+        dim = len(x.shape) - 2
+        group = self.group
+        batch = x.shape[0]
+        in_ch = x.shape[1]
+        out_ch = W.shape[1]
+        dilations = self.dilations or [1] * dim
+        strides = self.strides or [1] * dim
+        output_padding = self.output_padding or [0] * dim
+        kernel_shape = self.kernel_shape or W.shape[2:]
+        input_shape = x.shape[2:]
+        pads = self.pads or [0] * (dim * 2)
+
+        if dim != 2:
+            raise RunError("ConvTranspose", self.version, "support 2d only")
+
+        if group != 1:
+            raise RunError("ConvTranspose", self.version, "support group=1 only")
+
+        if self.auto_pad != "NOTSET":
+            raise RunError("ConvTranspose", self.version, "support auto_pad=NOTSET only")
+
+        # calculate pads and output_shape
+        if self.output_shape is not None:
+            output_shape = self.output_shape
+            total_padding = [
+                strides[i] * (input_shape[i] - 1)
+                + output_padding[i]
+                + ((kernel_shape[i] - 1) * dilations[i] + 1)
+                - output_shape[i]
+                for i in range(len(input_shape))
+            ]
+            for i in range(len(input_shape)):
+                pads[i] = total_padding[i] - (total_padding[i] // 2)
+                pads[i + dim] = total_padding[i] // 2
+        else:
+            output_shape = [
+                strides[i] * (input_shape[i] - 1)
+                + output_padding[i]
+                + ((kernel_shape[i] - 1) * dilations[i] + 1)
+                - pads[i]
+                - pads[i + dim]
+                for i in range(dim)
+            ]
+
+        # calculate output
+        result = np.zeros([batch, out_ch, *output_shape], dtype=x.dtype)
+
+        for n in range(batch):
+            for och in range(out_ch):
+                if b is not None:
+                    result[n, och, :, :] += b[och]
+                for ih in range(input_shape[0]):
+                    for iw in range(input_shape[1]):
+                        for kh in range(kernel_shape[0]):
+                            for kw in range(kernel_shape[1]):
+                                oh = strides[0] * ih + kh * dilations[0] - pads[0]
+                                ow = strides[1] * iw + kw * dilations[1] - pads[1]
+                                if oh < 0 or ow < 0 or oh >= output_shape[0] or ow >= output_shape[1]:
+                                    continue
+                                v = np.float32(0)
+                                for ich in range(in_ch):
+                                    v += x[n, ich, ih, iw] * W[ich, och, kh, kw]
+                                result[n, och, oh, ow] += v
+
+        return [result]

runtime/onnion_runtime/error.py

@@ -1,4 +1,8 @@
+from typing import Optional
+
+
 class RunError(Exception):
-    def __init__(self, op, version):
+    def __init__(self, op: str, version: int, reason: Optional[str] = None):
         self.op = op
         self.version = version
+        self.reason = reason

runtime/tests/test_convtranspose.py

@@ -0,0 +1,143 @@
+import numpy as np
+from onnion_runtime import ConvTranspose
+
+from .utils import check
+
+
+def test_convtranspose_00() -> None:
+    opset_version = 13
+    attrs = dict()
+    x = np.array([[[[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]]]]).astype(np.float32)  # (1, 1, 3, 3)
+
+    W = np.array(
+        [
+            [
+                [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],  # (1, 2, 3, 3)
+                [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
+            ]
+        ]
+    ).astype(np.float32)
+
+    inputs = [x, W]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+def test_convtranspose_01() -> None:
+    opset_version = 13
+    attrs = {"strides": [3, 2], "output_padding": [1, 1]}
+
+    x = np.array([[[[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]]]]).astype(np.float32)  # (1, 1, 3, 3)
+
+    W = np.array(
+        [
+            [
+                [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],  # (1, 2, 3, 3)
+                [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
+            ]
+        ]
+    ).astype(np.float32)
+
+    inputs = [x, W]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# test dillation
+def test_convtranspose_02() -> None:
+    opset_version = 13
+    attrs = {"dilations": [2, 2]}
+    x = np.random.randn(1, 1, 3, 3).astype(np.float32)
+    W = np.random.randn(1, 1, 2, 2).astype(np.float32)
+
+    inputs = [x, W]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# test pads
+def test_convtranspose_03() -> None:
+    opset_version = 13
+    attrs = {"strides": [3, 2], "pads": [1, 2, 1, 2]}
+    x = np.random.randn(1, 1, 3, 3).astype(np.float32)
+    W = np.random.randn(1, 2, 3, 3).astype(np.float32)
+    b = np.random.randn(2).astype(np.float32)
+
+    inputs = [x, W, b]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# specify output shape
+def test_convtranspose_04() -> None:
+    opset_version = 13
+    attrs = {"strides": [3, 2], "output_shape": [10, 8]}
+    x = np.random.randn(1, 1, 3, 3).astype(np.float32)
+    W = np.random.randn(1, 2, 3, 3).astype(np.float32)
+
+    inputs = [x, W]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# specify output shape and output padding
+def test_convtranspose_05() -> None:
+    opset_version = 13
+    attrs = {"strides": [3, 2], "output_shape": [10, 8], "kernel_shape": [3, 3], "output_padding": [1, 1]}
+    x = np.random.randn(1, 1, 3, 3).astype(np.float32)
+    W = np.random.randn(1, 2, 3, 3).astype(np.float32)
+    b = np.random.randn(2).astype(np.float32)
+
+    inputs = [x, W, b]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# larger channel number
+def test_convtranspose_06() -> None:
+    opset_version = 13
+    attrs = {"strides": [2, 2], "kernel_shape": [2, 2], "pads": [0, 0, 0, 0]}
+    x = np.random.randn(2, 24, 12, 12).astype(np.float32)
+    W = np.random.randn(24, 24, 2, 2).astype(np.float32)
+
+    inputs = [x, W]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# larger channel number (with bias)
+def test_convtranspose_07() -> None:
+    opset_version = 13
+    attrs = {"strides": [2, 2], "kernel_shape": [2, 2], "pads": [0, 0, 0, 0]}
+    x = np.random.randn(2, 24, 12, 12).astype(np.float32)
+    W = np.random.randn(24, 24, 2, 2).astype(np.float32)
+    b = np.random.randn(24).astype(np.float32)
+
+    inputs = [x, W, b]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# opset 1
+def test_convtranspose_08() -> None:
+    opset_version = 1
+    attrs = {"strides": [3, 2], "output_shape": [10, 8], "kernel_shape": [3, 3], "output_padding": [1, 1]}
+    x = np.random.randn(1, 1, 3, 3).astype(np.float32)
+    W = np.random.randn(1, 2, 3, 3).astype(np.float32)
+
+    inputs = [x, W]
+
+    check(ConvTranspose, opset_version, attrs, inputs)
+
+
+# opset 1 (larager channel number)
+def test_convtranspose_09() -> None:
+    opset_version = 1
+    attrs = {"strides": [2, 2], "kernel_shape": [2, 2], "pads": [0, 0, 0, 0]}
+    x = np.random.randn(2, 24, 12, 12).astype(np.float32)
+    W = np.random.randn(24, 24, 2, 2).astype(np.float32)
+    b = np.random.randn(24).astype(np.float32)
+
+    inputs = [x, W, b]
+
+    check(ConvTranspose, opset_version, attrs, inputs)

runtime/tests/utils.py

@@ -10,7 +10,7 @@
 try:
     import onnx
     import onnxruntime
-    from onnx import checker, helper, mapping, numpy_helper
+    from onnx import checker, helper, numpy_helper
 
     WITHOUT_ONNXRUNTIME = False
 except Exception:
@@ -24,7 +24,7 @@ def on_arm32():
         result = bool(int(os.environ["ONNION_TEST_ON_ARM32"]))
     except Exception:
         arch = platform.machine()
-        if arch == "x86_64":
+        if arch == "x86_64" or arch == "arm64":
             result = False
         elif arch == "armv7l":
             result = True
@@ -81,7 +81,7 @@ def check_by_data(expected, result, max_error=1e-4):
 
 def _convert_type(dtype):
     assert not WITHOUT_ONNXRUNTIME
-    return mapping.NP_TYPE_TO_TENSOR_TYPE[dtype]
+    return helper.np_dtype_to_tensor_dtype(dtype)
 
 
 def _run_onnx(model, inputs, output_names):