|
| 1 | +# Copyright (c) Microsoft Corporation. |
| 2 | +# Licensed under the MIT License. |
| 3 | +"""Fuses BatchNormalization nodes into preceding nodes. Supported fusion patterns: |
| 4 | +- BatchNormalization + Conv -> Conv |
| 5 | +- BatchNormalization + ConvTranpose -> ConvTranpose |
| 6 | +- BatchNormalization + Gemm -> Gemm |
| 7 | +
|
| 8 | +Approach: |
| 9 | + Given an inbound operation output: Y = W * X + B |
| 10 | + And a BatchNormalization outputs: Y_BN = (gamma * (Y - μ) / std) + β, where std = sqrt(var + eps) |
| 11 | +
|
| 12 | + The fusion updates the inbound weights as follows: |
| 13 | + - W_fused = W * (gamma / std) |
| 14 | + - B_fused = (B - μ) * (gamma / std) + β |
| 15 | +""" |
| 16 | + |
| 17 | +from abc import ABC, abstractmethod |
| 18 | + |
| 19 | +import numpy as np |
| 20 | + |
| 21 | +from onnxscript import ir |
| 22 | +from onnxscript.rewriter import pattern as orp |
| 23 | + |
| 24 | + |
| 25 | +class FuseBatchNormBase(orp.RewriteRuleClassBase, ABC): |
| 26 | + """Interface for BatchNormalization nodes fusion.""" |
| 27 | + |
| 28 | + def __init__( |
| 29 | + self, |
| 30 | + op_type: str, |
| 31 | + name: str | None = None, |
| 32 | + remove_nodes: bool = True, |
| 33 | + as_function: bool = False, |
| 34 | + ) -> None: |
| 35 | + super().__init__(name=name, remove_nodes=remove_nodes, as_function=as_function) |
| 36 | + self.op_type = op_type |
| 37 | + |
| 38 | + @abstractmethod |
| 39 | + def get_filters_axis(self, attributes) -> int: |
| 40 | + """Return the axis along which BatchNorm scale should be broadcasted.""" |
| 41 | + |
| 42 | + def _reshape_for_broadcast(self, x: np.ndarray, rank: int, axis: int = 1) -> np.ndarray: |
| 43 | + # Convert axis to positive |
| 44 | + if axis < 0: |
| 45 | + axis += rank |
| 46 | + |
| 47 | + # Build shape: 1s everywhere except -1 at the target axis |
| 48 | + broadcast_shape = [1 if axis != i else -1 for i in range(rank)] |
| 49 | + return np.reshape(x, broadcast_shape) |
| 50 | + |
| 51 | + def rewrite(self, op, x: ir.Value, inbound_out: ir.Value, batchnorm_out: ir.Value): |
| 52 | + batchnorm_node = batchnorm_out.producer() |
| 53 | + # Get BatchNorm parameters |
| 54 | + gamma, beta, input_mean, input_var = [ |
| 55 | + inp.const_value.numpy() for inp in batchnorm_node.inputs[1:] |
| 56 | + ] |
| 57 | + |
| 58 | + # 1e-5 is the default value for epsilon according to |
| 59 | + # https://onnx.ai/onnx/operators/onnx__BatchNormalization.html#attributes |
| 60 | + default_eps = ir.Attr("epsilon", ir.AttributeType.FLOAT, 1e-5) |
| 61 | + eps = batchnorm_node.attributes.get("epsilon", default_eps).as_float() |
| 62 | + |
| 63 | + # Compute the scale_factor to update the inbound weights and bias |
| 64 | + scale_factor = gamma / np.sqrt(input_var + eps) |
| 65 | + |
| 66 | + # Update inbound weights |
| 67 | + inbound_node = inbound_out.producer() |
| 68 | + weights = inbound_node.inputs[1].const_value.numpy() |
| 69 | + |
| 70 | + # Reshape scale factor so it is broadcastable |
| 71 | + axis = self.get_filters_axis(inbound_node.attributes) |
| 72 | + fused_weights = ir.tensor( |
| 73 | + weights * self._reshape_for_broadcast(scale_factor, weights.ndim, axis=axis) |
| 74 | + ) |
| 75 | + |
| 76 | + # Update bias |
| 77 | + if len(inbound_node.inputs) > 2: |
| 78 | + original_bias = inbound_node.inputs[2].const_value.numpy() |
| 79 | + bias_name = inbound_node.inputs[2].name |
| 80 | + else: |
| 81 | + original_bias = np.zeros_like(input_mean) |
| 82 | + bias_name = x.name + "_bias" |
| 83 | + fused_bias = ir.tensor((original_bias - input_mean) * scale_factor + beta) |
| 84 | + |
| 85 | + return op.op( |
| 86 | + self.op_type, |
| 87 | + inputs=[ |
| 88 | + x, |
| 89 | + op.initializer(fused_weights, name=inbound_node.inputs[1].name), |
| 90 | + op.initializer(fused_bias, name=bias_name), |
| 91 | + ], |
| 92 | + attributes=inbound_node.attributes, |
| 93 | + ) |
| 94 | + |
| 95 | + def check(self, context, x, inbound_out, batchnorm_out) -> orp.MatchResult: |
| 96 | + del context # Unused |
| 97 | + check_result = orp.MatchResult() |
| 98 | + |
| 99 | + inbound_node = inbound_out.producer() |
| 100 | + batchnorm_node = batchnorm_out.producer() |
| 101 | + |
| 102 | + # Check that inbound weights + (inbound bias) + batchnorm params are initializers |
| 103 | + initializers = [inbound_node.inputs[1], *batchnorm_node.inputs[1:]] |
| 104 | + if len(inbound_node.inputs) > 2: |
| 105 | + initializers.append(inbound_node.inputs[2]) |
| 106 | + |
| 107 | + for initializer in initializers: |
| 108 | + if not initializer.is_initializer() or initializer.const_value is None: |
| 109 | + return check_result.fail(f"{initializer.name} is not a constant initializer") |
| 110 | + |
| 111 | + return check_result |
| 112 | + |
| 113 | + |
| 114 | +class FuseBatchNormIntoConv(FuseBatchNormBase): |
| 115 | + """Replaces ``BatchNormalization(Conv(x))`` with ``Conv(x)``.""" |
| 116 | + |
| 117 | + def __init__(self): |
| 118 | + super().__init__("Conv") |
| 119 | + |
| 120 | + def get_filters_axis(self, attributes) -> int: |
| 121 | + return 0 |
| 122 | + |
| 123 | + def pattern(self, op, x): |
| 124 | + return op.BatchNormalization( |
| 125 | + op.Conv(x, _allow_other_inputs=True, _outputs=["inbound_out"]), |
| 126 | + _allow_other_inputs=True, |
| 127 | + _outputs=["batchnorm_out"], |
| 128 | + ) |
| 129 | + |
| 130 | + |
| 131 | +class FuseBatchNormIntoConvTranspose(FuseBatchNormBase): |
| 132 | + """Replaces ``BatchNormalization(ConvTranspose(x))`` with ``ConvTranspose(x)``.""" |
| 133 | + |
| 134 | + def __init__(self): |
| 135 | + super().__init__("ConvTranspose") |
| 136 | + |
| 137 | + def get_filters_axis(self, attributes) -> int: |
| 138 | + return 1 |
| 139 | + |
| 140 | + def pattern(self, op, x): |
| 141 | + return op.BatchNormalization( |
| 142 | + op.ConvTranspose(x, _allow_other_inputs=True, _outputs=["inbound_out"]), |
| 143 | + _allow_other_inputs=True, |
| 144 | + _outputs=["batchnorm_out"], |
| 145 | + ) |
| 146 | + |
| 147 | + |
| 148 | +class FuseBatchNormIntoGemm(FuseBatchNormBase): |
| 149 | + """Replaces ``BatchNormalization(Gemm(x))`` with ``Gemm(x)``.""" |
| 150 | + |
| 151 | + def __init__(self): |
| 152 | + super().__init__("Gemm") |
| 153 | + |
| 154 | + def get_filters_axis(self, attributes) -> int: |
| 155 | + return 0 if attributes.get("transB") is not None and attributes["transB"].value else 1 |
| 156 | + |
| 157 | + def pattern(self, op, x): |
| 158 | + return op.BatchNormalization( |
| 159 | + op.Gemm(x, _allow_other_inputs=True, _outputs=["inbound_out"]), |
| 160 | + _allow_other_inputs=True, |
| 161 | + _outputs=["batchnorm_out"], |
| 162 | + ) |
| 163 | + |
| 164 | + |
| 165 | +fuse_batchnorm_into_conv_rule = FuseBatchNormIntoConv().rule() |
| 166 | +fuse_batchnorm_into_convtranspose_rule = FuseBatchNormIntoConvTranspose().rule() |
| 167 | +fuse_batchnorm_into_gemm_rule = FuseBatchNormIntoGemm().rule() |
| 168 | + |
| 169 | + |
| 170 | +def fuse_batchnorm_rule_set() -> orp.RewriteRuleSet: |
| 171 | + """Returns a set of rewrite rules that fuse BatchNormalization nodes |
| 172 | + into preceding nodes such as Conv, ConvTranspose, and Gemm. |
| 173 | +
|
| 174 | + Returns: |
| 175 | + RewriteRuleSet |
| 176 | + """ |
| 177 | + return orp.RewriteRuleSet( |
| 178 | + [ |
| 179 | + fuse_batchnorm_into_conv_rule, |
| 180 | + fuse_batchnorm_into_convtranspose_rule, |
| 181 | + fuse_batchnorm_into_gemm_rule, |
| 182 | + ] |
| 183 | + ) |
0 commit comments