onnxslim 0.1.72__tar.gz → 0.1.74__tar.gz

{onnxslim-0.1.72/onnxslim.egg-info → onnxslim-0.1.74}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnxslim
-Version: 0.1.72
+Version: 0.1.74
 Summary: OnnxSlim: A Toolkit to Help Optimize Onnx Model
 Home-page: https://github.com/inisis/OnnxSlim
 Author: inisis

onnxslim-0.1.74/VERSION

@@ -0,0 +1 @@
+0.1.74

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim/core/pattern/fusion/__init__.py

@@ -1,6 +1,7 @@
 from .concat_reshape import *
 from .convadd import *
 from .convbn import *
+from .convmul import *
 from .gelu import *
 from .gemm import *
 from .padconv import *

onnxslim-0.1.74/onnxslim/core/pattern/fusion/convmul.py

@@ -0,0 +1,69 @@
+import onnxslim.third_party.onnx_graphsurgeon as gs
+from onnxslim.core.pattern import Pattern, PatternMatcher
+from onnxslim.core.pattern.registry import register_fusion_pattern
+
+
+class ConvMulMatcher(PatternMatcher):
+    def __init__(self, priority):
+        """Initializes the ConvMulMatcher for fusing Conv and Mul layers in an ONNX graph."""
+        pattern = Pattern(
+            """
+            input    input  0  1 conv_0
+            Conv     conv_0 1+ 1 input mul_0
+            Mul      mul_0  2  1 conv_0 ? output
+            output   output 1  0 mul_0
+            """
+        )
+        super().__init__(pattern, priority)
+
+    @property
+    def name(self):
+        """Returns the name of the FusionConvMul pattern."""
+        return "FusionConvMul"
+
+    def rewrite(self, opset=11):
+        match_case = {}
+        conv_node = self.conv_0
+        mul_node = self.mul_0
+        conv_weight = list(conv_node.inputs)[1]
+        if len(conv_node.users) == 1 and conv_node.users[0] == mul_node and isinstance(mul_node.inputs[1], gs.Constant):
+            mul_constant = mul_node.inputs[1].values
+
+            if mul_constant.squeeze().ndim == 1 and mul_constant.squeeze().shape[0] == conv_weight.shape[0]:
+                weight_shape = conv_weight.values.shape
+                reshape_shape = [-1] + [1] * (len(weight_shape) - 1)
+
+                mul_scale_reshaped = mul_constant.squeeze().reshape(reshape_shape)
+                new_weight = conv_weight.values * mul_scale_reshaped
+
+                inputs = []
+                inputs.append(next(iter(conv_node.inputs)))
+
+                weight_name = list(conv_node.inputs)[1].name
+                inputs.append(gs.Constant(weight_name, values=new_weight))
+
+                if len(conv_node.inputs) == 3:
+                    conv_bias = conv_node.inputs[2].values
+                    new_bias = conv_bias * mul_constant.squeeze()
+                    bias_name = list(conv_node.inputs)[2].name
+                    inputs.append(gs.Constant(bias_name, values=new_bias))
+
+                outputs = list(mul_node.outputs)
+
+                conv_node.outputs.clear()
+                mul_node.inputs.clear()
+                mul_node.outputs.clear()
+
+                match_case[conv_node.name] = {
+                    "op": conv_node.op,
+                    "inputs": inputs,
+                    "outputs": outputs,
+                    "name": conv_node.name,
+                    "attrs": conv_node.attrs,
+                    "domain": None,
+                }
+
+        return match_case
+
+
+register_fusion_pattern(ConvMulMatcher(1))

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim/core/pattern/fusion/gemm.py

@@ -289,16 +289,38 @@ class GemmAddPatternMatcher(PatternMatcher):
             )
             and add_bias_variable
             and len(reshape_node.users) == 1
+            and gemm_node.outputs[0].shape
         ):
+
+            def can_broadcast_to(shape_from, shape_to):
+                """Return True if shape_from can broadcast to shape_to per NumPy rules."""
+                if shape_from is None or shape_to is None:
+                    return False
+                try:
+                    np.empty(shape_to, dtype=np.float32) + np.empty(shape_from, dtype=np.float32)
+                    return True
+                except ValueError:
+                    return False
+
             gemm_bias_constant = gemm_node.inputs[2] if len(gemm_node.inputs) == 3 else None
             if gemm_bias_constant:
                 gemm_bias = gemm_bias_constant.values
                 add_bias = add_bias_variable.values
-                gemm_bias_fused = gemm_bias + add_bias
-                gemm_bias_fused_constant = gs.Constant(gemm_bias_constant.name + "_fused", values=gemm_bias_fused)
-                gemm_node.inputs[2] = gemm_bias_fused_constant
+                if (
+                    can_broadcast_to(gemm_bias.shape, gemm_node.outputs[0].shape)
+                    and can_broadcast_to(add_bias.shape, gemm_node.outputs[0].shape)
+                    and add_bias.ndim <= 2
+                ):
+                    gemm_bias_fused = gemm_bias + add_bias
+                    gemm_bias_fused_constant = gs.Constant(gemm_bias_constant.name + "_fused", values=gemm_bias_fused)
+                    gemm_node.inputs[2] = gemm_bias_fused_constant
+                else:
+                    return match_case
             else:
-                gemm_node.inputs[2] = add_bias_variable
+                if can_broadcast_to(add_bias_variable.values.shape, gemm_node.outputs[0].shape):
+                    gemm_node.inputs[2] = add_bias_variable
+                else:
+                    return match_case
 
             add_node.replace_all_uses_with(reshape_node)
 

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim/core/pattern/fusion/padconv.py

@@ -24,6 +24,7 @@ class PadConvMatcher(PatternMatcher):
     def parameter_check(self) -> bool:
         """Validates if the padding parameter for a convolutional node is a constant."""
         pad_node = self.pad_0
+
         return isinstance(pad_node.inputs[1], gs.Constant)
 
     def rewrite(self, opset=11):
@@ -42,7 +43,7 @@ class PadConvMatcher(PatternMatcher):
         ):
             if (
                 isinstance(pad_node.inputs[1], gs.Constant)
-                and pad_node.attrs["mode"] == "constant"
+                and pad_node.attrs.get("mode", "constant") == "constant"
                 and conv_node.inputs[1].shape
             ):
                 conv_weight_dim = len(conv_node.inputs[1].shape)
@@ -67,9 +68,10 @@ class PadConvMatcher(PatternMatcher):
                         pad_node.inputs.clear()
                         pad_node.outputs.clear()
 
-                    conv_pads = attrs["pads"]
                     pads = pad_value[2:conv_weight_dim] + pad_value[conv_weight_dim + 2 :]
-                    pads = [pad + conv_pad for pad, conv_pad in zip(pads, conv_pads)]
+                    if hasattr(attrs, "pads"):
+                        conv_pads = attrs["pads"]
+                        pads = [pad + conv_pad for pad, conv_pad in zip(pads, conv_pads)]
 
                     attrs["pads"] = pads
                     match_case[conv_node.name] = {

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim/misc/tabulate.py

@@ -1606,7 +1606,7 @@ def tabulate(
         given header. Possible values are: "global" (no override), "same"
         (follow column alignment), "right", "center", "left".
 
-    Note on intended behaviour: If there is no `tabular_data`, any column
+    Note on intended behavior: If there is no `tabular_data`, any column
         alignment argument is ignored. Hence, in this case, header
         alignment cannot be inferred from column alignment.
 

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim/third_party/_sympy/functions.py

@@ -52,7 +52,7 @@ def simple_floordiv_gcd(p: sympy.Basic, q: sympy.Basic) -> sympy.Basic:
 
     We try to rewrite p and q in the form n*e*p1 + n*e*p2 and n*e*q0, where n is the greatest common integer factor and
     e is the largest syntactic common factor (i.e., common sub-expression) in p and q. Then the gcd returned is n*e,
-    cancelling which we would be left with p1 + p2 and q0.
+    canceling which we would be left with p1 + p2 and q0.
 
     Note that further factoring of p1 + p2 and q0 might be possible with sympy.factor (which uses domain-specific
     theories). E.g., we are unable to find that x*y + x + y + 1 is divisible by x + 1. More generally, when q is of the

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim/third_party/symbolic_shape_infer.py

@@ -532,6 +532,7 @@ class SymbolicShapeInference:
             initializers = []
             if (get_opset(self.out_mp_) >= 9) and (
                 node.op_type == "Unsqueeze" or node.op_type == "ReduceMax" or node.op_type == "ReduceMean"
+                or node.op_type == "DFT" or node.op_type == "ReduceL2" or node.op_type == "ReduceMin"
             ):
                 initializers = [
                     self.initializers_[name]

onnxslim-0.1.74/onnxslim/version.py

@@ -0,0 +1 @@
+__version__ = "0.1.74"

{onnxslim-0.1.72 → onnxslim-0.1.74/onnxslim.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnxslim
-Version: 0.1.72
+Version: 0.1.74
 Summary: OnnxSlim: A Toolkit to Help Optimize Onnx Model
 Home-page: https://github.com/inisis/OnnxSlim
 Author: inisis

{onnxslim-0.1.72 → onnxslim-0.1.74}/onnxslim.egg-info/SOURCES.txt

@@ -35,6 +35,7 @@ onnxslim/core/pattern/fusion/__init__.py
 onnxslim/core/pattern/fusion/concat_reshape.py
 onnxslim/core/pattern/fusion/convadd.py
 onnxslim/core/pattern/fusion/convbn.py
+onnxslim/core/pattern/fusion/convmul.py
 onnxslim/core/pattern/fusion/gelu.py
 onnxslim/core/pattern/fusion/gemm.py
 onnxslim/core/pattern/fusion/padconv.py

onnxslim-0.1.72/VERSION

@@ -1 +0,0 @@
-0.1.72

onnxslim-0.1.72/onnxslim/version.py

@@ -1 +0,0 @@
-__version__ = "0.1.72"