onnx 1.16.2__cp312-cp312-win32.whl → 1.18.0__cp312-cp312-win32.whl

onnx/reference/ops/op_conv.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 
@@ -124,9 +124,9 @@ def _conv_implementation(  # type: ignore
         if B is not None:
             res[:, :, :] += B.reshape((1, -1, 1))  # type: ignore
 
-        for n in range(0, sN):
+        for n in range(sN):
             for nw in range(W.shape[0]):
-                for c in range(0, sC):
+                for c in range(sC):
                     w = W[nw : nw + 1, c : c + 1]
                     for io in range(bh, eh, sth):
                         hr = (io - bh) // sth
@@ -171,9 +171,9 @@ def _conv_implementation(  # type: ignore
         if B is not None:
             res[:, :, :, :] = B.reshape((1, -1, 1, 1))  # type: ignore
 
-        for n in range(0, sN):
+        for n in range(sN):
             for nw in range(W.shape[0]):
-                for c in range(0, sC):
+                for c in range(sC):
                     w = W[nw : nw + 1, c : c + 1]
                     for io in range(bh, eh, sth):
                         hr = (io - bh) // sth
@@ -189,11 +189,13 @@ def _conv_implementation(  # type: ignore
                             iw1, iw2 = max(0, j + ow), min(j + ow + kw, sW)
                             img = X[n : n + 1, c : c + 1, ih1:ih2, iw1:iw2]
                             if img.shape != w.shape:
-                                jh1, jh2 = max(-oh - i, 0), min(
-                                    kh, kh + sH - (i + oh + kh)
+                                jh1, jh2 = (
+                                    max(-oh - i, 0),
+                                    min(kh, kh + sH - (i + oh + kh)),
                                 )
-                                jw1, jw2 = max(-ow - j, 0), min(
-                                    kw, kw + sW - (j + ow + kw)
+                                jw1, jw2 = (
+                                    max(-ow - j, 0),
+                                    min(kw, kw + sW - (j + ow + kw)),
                                 )
                                 w_ = w[:1, :1, jh1:jh2, jw1:jw2]
                                 if img.shape != w_.shape:
@@ -229,9 +231,9 @@ def _conv_implementation(  # type: ignore
         if B is not None:
             res[:, :, :, :, :] = B.reshape((1, -1, 1, 1, 1))  # type: ignore
 
-        for n in range(0, sN):
+        for n in range(sN):
             for nw in range(W.shape[0]):
-                for c in range(0, sC):
+                for c in range(sC):
                     w = W[nw : nw + 1, c : c + 1]
                     for io in range(bh, eh, sth):
                         hr = (io - bh) // sth
@@ -253,14 +255,17 @@ def _conv_implementation(  # type: ignore
                                 iz1, iz2 = max(0, z + oz), min(z + oz + kz, sZ)
                                 img = X[n : n + 1, c : c + 1, ih1:ih2, iw1:iw2, iz1:iz2]
                                 if img.shape != w.shape:
-                                    jh1, jh2 = max(-oh - i, 0), min(
-                                        kh, kh + sH - (i + oh + kh)
+                                    jh1, jh2 = (
+                                        max(-oh - i, 0),
+                                        min(kh, kh + sH - (i + oh + kh)),
                                     )
-                                    jw1, jw2 = max(-ow - j, 0), min(
-                                        kw, kw + sW - (j + ow + kw)
+                                    jw1, jw2 = (
+                                        max(-ow - j, 0),
+                                        min(kw, kw + sW - (j + ow + kw)),
                                     )
-                                    jz1, jz2 = max(-oz - z, 0), min(
-                                        kz, kz + sZ - (z + oz + kz)
+                                    jz1, jz2 = (
+                                        max(-oz - z, 0),
+                                        min(kz, kz + sZ - (z + oz + kz)),
                                     )
                                     w_ = w[:1, :1, jh1:jh2, jw1:jw2, jz1:jz2]
                                     if img.shape != w_.shape:
@@ -271,15 +276,11 @@ def _conv_implementation(  # type: ignore
                                         )
                                     s = np.dot(
                                         img.reshape((1, -1)), w_.reshape((-1, 1))
-                                    )[
-                                        0, 0
-                                    ]  # (img * w_).sum()
+                                    )[0, 0]  # (img * w_).sum()
                                 else:
                                     s = np.dot(
                                         img.reshape((1, -1)), w.reshape((-1, 1))
-                                    )[
-                                        0, 0
-                                    ]  # (img * w).sum()
+                                    )[0, 0]  # (img * w).sum()
                                 res[n, nw, hr, wr, zr] += s  # type: ignore
 
         return res

onnx/reference/ops/op_conv_integer.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_conv_transpose.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 
@@ -24,8 +24,6 @@ class ConvTranspose(OpRun):
         pads=None,
         strides=None,
     ):
-        if group != 1:
-            raise RuntimeError(f"group={group} != 1 is not implemented yet.")
         if dilations is None:
             dilations = [1 for s in X.shape[2:]]
         if kernel_shape is None:
@@ -103,8 +101,39 @@ class ConvTranspose(OpRun):
                         res += B[c]
                     final[image_id, c, ...] = res[...]
         else:
-            raise NotImplementedError(
-                f"Implementation for group={group} > 1 is not available yet."
-            )
+            final = np.zeros((X.shape[0], num_output_channels, *output_shape))
+            output_array = []
+
+            for group_id in range(group):
+                group_X = X[:, group_id * C // group : (group_id + 1) * C // group, ...]
+                group_W = W[
+                    group_id * num_output_channels // group : (group_id + 1)
+                    * num_output_channels
+                    // group,
+                    ...,
+                ]
+
+                group_output = self._run(
+                    group_X,
+                    group_W,
+                    B=B,
+                    auto_pad=auto_pad,
+                    dilations=dilations,
+                    group=1,
+                    kernel_shape=kernel_shape,
+                    output_padding=output_padding,
+                    output_shape=output_shape,
+                    pads=pads,
+                    strides=strides,
+                )
+                group_output = np.array(group_output[0])
+                output_array.append(group_output)
+
+            for image_id in range(X.shape[0]):
+                for group_id in range(group):
+                    group_output = output_array[group_id]
+                    final[image_id, group_id : (group_id + 1), ...] = group_output[
+                        image_id, ...
+                    ]
 
         return (final.astype(X.dtype),)  # type: ignore[union-attr]

onnx/reference/ops/op_cos.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_cosh.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_cum_sum.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 
@@ -9,12 +9,7 @@ from onnx.reference.op_run import OpRun
 
 
 class CumSum(OpRun):
-    def _run(self, x, *axis, exclusive=None, reverse=None):  # type: ignore
-        axis = None if not axis else axis[0]  # type: ignore
-        if axis is None:  # type: ignore
-            if reverse or exclusive:
-                raise NotImplementedError("reverse=1 or exclusive=1 not implemented")
-            return (np.cumsum(x),)
+    def _run(self, x, axis, exclusive=None, reverse=None):  # type: ignore
         if not isinstance(axis, (np.int32, np.int64)):
             if len(axis.shape) > 1 or (len(axis.shape) > 0 and axis.shape[0] != 1):  # type: ignore
                 raise RuntimeError(
@@ -34,7 +29,7 @@ class CumSum(OpRun):
             res = np.zeros(x.shape, dtype=x.dtype)
             np.cumsum(x[tuple(indices_c)], axis=axis, out=res[tuple(indices_d)])  # type: ignore
         else:
-            res = np.cumsum(x, axis=axis)  # type: ignore
+            res = np.cumsum(x, axis=axis, dtype=x.dtype)  # type: ignore
         if reverse:
             res = res[tuple(rev_indices)]
         return (res,)

onnx/reference/ops/op_deform_conv.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 #
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_depth_to_space.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_dequantize_linear.py

@@ -1,16 +1,13 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
-
-from typing import Optional
+from __future__ import annotations
 
 import numpy as np
 
 from onnx import TensorProto
-from onnx.helper import np_dtype_to_tensor_dtype
-from onnx.numpy_helper import float8e4m3_to_float32, float8e5m2_to_float32
-from onnx.reference.custom_element_types import (
+from onnx._custom_element_types import (
+    float4e2m1,
     float8e4m3fn,
     float8e4m3fnuz,
     float8e5m2,
@@ -18,6 +15,12 @@ from onnx.reference.custom_element_types import (
     int4,
     uint4,
 )
+from onnx.helper import np_dtype_to_tensor_dtype, tensor_dtype_to_np_dtype
+from onnx.numpy_helper import (
+    float8e4m3_to_float32,
+    float8e5m2_to_float32,
+    unpacked_float4e2m1_to_float32,
+)
 from onnx.reference.op_run import OpRun
 from onnx.reference.ops.op_quantize_linear import reshape_input
 
@@ -37,6 +40,8 @@ class _CommonDequantizeLinear(OpRun):
             tensor_dtype = TensorProto.UINT4
         elif x.dtype == int4 and x.dtype.descr[0][0] == "int4":
             tensor_dtype = TensorProto.INT4
+        elif x.dtype == float4e2m1 and x.dtype.descr[0][0] == "float4e2m1":
+            tensor_dtype = TensorProto.FLOAT4E2M1
         else:
             tensor_dtype = np_dtype_to_tensor_dtype(x.dtype)
         return tensor_dtype
@@ -45,9 +50,10 @@ class _CommonDequantizeLinear(OpRun):
         self,
         x: np.ndarray,
         x_scale: np.ndarray,
-        x_zero_point: Optional[np.ndarray] = None,
-        axis: Optional[int] = None,
-        block_size: Optional[int] = None,
+        x_zero_point: np.ndarray | None = None,
+        axis: int | None = None,
+        block_size: int | None = None,
+        output_dtype: int | None = None,
     ):  # type: ignore
         x_type = self.get_x_type(x)
         fp8_type = x_type in {
@@ -56,7 +62,11 @@ class _CommonDequantizeLinear(OpRun):
             TensorProto.FLOAT8E5M2,
             TensorProto.FLOAT8E5M2FNUZ,
         }
-        if x_zero_point is not None and not fp8_type:
+        if (
+            x_zero_point is not None
+            and not fp8_type
+            and x_type != TensorProto.FLOAT4E2M1
+        ):
             zero_type = self.get_x_type(x_zero_point)
             if x_type != zero_type:
                 raise ValueError(
@@ -83,10 +93,18 @@ class _CommonDequantizeLinear(OpRun):
                 dx = float8e5m2_to_float32(x)
             elif x_type == TensorProto.FLOAT8E5M2FNUZ:
                 dx = float8e5m2_to_float32(x, fn=True, uz=True)
+            elif x_type == TensorProto.FLOAT4E2M1:
+                dx = unpacked_float4e2m1_to_float32(x)
             else:
                 dx = x.astype(np.float32)
         y = dx * reshape_input(x_scale, x.shape, axis, block_size)
-        return (y.astype(x_scale.dtype),)
+        return (
+            y.astype(
+                tensor_dtype_to_np_dtype(output_dtype)
+                if output_dtype
+                else x_scale.dtype
+            ),
+        )
 
 
 class DequantizeLinear_19(_CommonDequantizeLinear):
@@ -100,3 +118,11 @@ class DequantizeLinear_21(_CommonDequantizeLinear):
     def _run(self, *args, axis=None, block_size=None):  # type: ignore
         # args: x, y_scale, zero_point
         return super()._run(*args, axis=axis, block_size=block_size)  # type: ignore
+
+
+class DequantizeLinear_23(_CommonDequantizeLinear):
+    def _run(self, *args, axis=None, block_size=None, output_dtype=None):  # type: ignore
+        # args: x, y_scale, zero_point
+        return super()._run(
+            *args, axis=axis, block_size=block_size, output_dtype=output_dtype
+        )  # type: ignore

onnx/reference/ops/op_det.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_dft.py

@@ -82,7 +82,14 @@ def _cifft(
 
 
 class DFT_17(OpRun):
-    def _run(self, x: np.ndarray, dft_length: int | None = None, axis: int = 1, inverse: bool = False, onesided: bool = False) -> tuple[np.ndarray]:  # type: ignore
+    def _run(
+        self,
+        x: np.ndarray,
+        dft_length: int | None = None,
+        axis: int = 1,
+        inverse: bool = False,
+        onesided: bool = False,
+    ) -> tuple[np.ndarray]:  # type: ignore
         # Convert to positive axis
         axis = axis % len(x.shape)
         if dft_length is None:
@@ -95,7 +102,14 @@ class DFT_17(OpRun):
 
 
 class DFT_20(OpRun):
-    def _run(self, x: np.ndarray, dft_length: int | None = None, axis: int = -2, inverse: bool = False, onesided: bool = False) -> tuple[np.ndarray]:  # type: ignore
+    def _run(
+        self,
+        x: np.ndarray,
+        dft_length: int | None = None,
+        axis: int = -2,
+        inverse: bool = False,
+        onesided: bool = False,
+    ) -> tuple[np.ndarray]:  # type: ignore
         # Convert to positive axis
         axis = axis % len(x.shape)
         if dft_length is None:

onnx/reference/ops/op_div.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_dropout.py

@@ -1,9 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
-
-from typing import Optional, Tuple
+from __future__ import annotations
 
 import numpy as np
 from numpy.random import RandomState  # type: ignore
@@ -14,10 +12,10 @@ from onnx.reference.op_run import OpRun
 def _dropout(
     X: np.ndarray,
     drop_probability: float = 0.5,
-    seed: Optional[int] = None,
+    seed: int | None = None,
     training_mode: bool = False,
     return_mask: bool = False,
-) -> Tuple[np.ndarray]:
+) -> tuple[np.ndarray]:
     if drop_probability == 0 or not training_mode:
         if return_mask:
             return X, np.ones(X.shape, dtype=bool)  # type: ignore
@@ -37,10 +35,10 @@ class DropoutBase(OpRun):
     def _private_run(
         self,
         X: np.ndarray,
-        seed: Optional[int] = None,
+        seed: int | None = None,
         ratio: float = 0.5,
         training_mode: bool = False,
-    ) -> Tuple[np.ndarray]:
+    ) -> tuple[np.ndarray]:
         return _dropout(
             X,
             ratio,
@@ -61,5 +59,8 @@ class Dropout_12(DropoutBase):
         ratio = 0.5 if len(inputs) <= 1 else inputs[1]
         training_mode = False if len(inputs) <= 2 else inputs[2]
         return self._private_run(
-            X, seed=seed, ratio=ratio, training_mode=training_mode  # type: ignore
+            X,
+            seed=seed,
+            ratio=ratio,
+            training_mode=training_mode,  # type: ignore
         )

onnx/reference/ops/op_dynamic_quantize_linear.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_einsum.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_elu.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_equal.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_erf.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 from math import erf
 

onnx/reference/ops/op_exp.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_expand.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_eyelike.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 
@@ -11,7 +11,7 @@ from onnx.reference.op_run import OpRun
 
 
 class EyeLike(OpRun):
-    def _run(self, data, *args, dtype=None, k=None):
+    def _run(self, data, *args, dtype=None, k=None):  # noqa: ARG002
         if dtype is None:
             if data is None:
                 _dtype = np.float32

onnx/reference/ops/op_flatten.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_floor.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_gather.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_gather_elements.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 
@@ -28,10 +28,10 @@ def gather_numpy(self: np.ndarray, dim: int, index: np.ndarray) -> np.ndarray:
 
     try:
         gathered = np.choose(index_swaped, data_swaped, mode="wrap")
-    except ValueError as e:
+    except ValueError:
         if len(index_swaped.shape) == 2 and len(data_swaped.shape) == 2:
             return gather_numpy_2(self, index)
-        raise e  # pragma: no cover
+        raise  # pragma: no cover
 
     return np.swapaxes(gathered, 0, dim)
 

onnx/reference/ops/op_gathernd.py

@@ -1,9 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
-
-from typing import Tuple
+from __future__ import annotations
 
 import numpy as np
 
@@ -12,7 +10,7 @@ from onnx.reference.op_run import OpRun
 
 def _gather_nd_impl(
     data: np.ndarray, indices: np.ndarray, batch_dims: int
-) -> Tuple[np.ndarray]:
+) -> tuple[np.ndarray]:
     # Note the data rank - will be reused multiple times later
     data_rank = len(data.shape)
 

onnx/reference/ops/op_gemm.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 
@@ -37,7 +37,17 @@ def _gemm11(a, b, c, alpha, beta):  # type: ignore
 
 
 class Gemm_6(OpRun):
-    def _run(self, a, b, c=None, alpha=None, beta=None, transA=None, transB=None, broadcast=None):  # type: ignore
+    def _run(
+        self,
+        a,
+        b,
+        c=None,
+        alpha=None,
+        beta=None,
+        transA=None,
+        transB=None,
+        broadcast=None,
+    ):  # type: ignore
         if broadcast == 0:
             if transA:
                 _meth = _gemm11 if transB else _gemm10

onnx/reference/ops/op_global_average_pool.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_global_max_pool.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np
 

onnx/reference/ops/op_greater.py

@@ -1,7 +1,7 @@
 # Copyright (c) ONNX Project Contributors
 
 # SPDX-License-Identifier: Apache-2.0
-
+from __future__ import annotations
 
 import numpy as np