inference_transpiler.py 10.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import numpy as np
16 17 18
from .. import core
from ..framework import Program
from ..executor import global_scope
19 20 21


class InferenceTranspiler:
L
Luo Tao 已提交
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
    '''
    Convert the fluid program to optimized inference program. 
    
    There are several optimizations, only fuse batch normalization is supported now.

    Examples:
   
    .. code-block:: python

        # As InferenceTranspiler will modify the original program,
        # please clone before use it.
        inference_transpiler_program = program.clone()
        t = fluid.InferenceTranspiler()
        t.transpile(inference_transpiler_program, place)
    '''

L
Luo Tao 已提交
38
    def transpile(self, program, place, scope=None):
39
        '''
L
Luo Tao 已提交
40 41 42 43 44 45
        Run the transpiler.

        Args:
            program (Program): program to transpile
            place (Place): inference place
            scope (Scope|None): inference Scope
L
Luo Tao 已提交
46
        '''
L
Luo Tao 已提交
47 48 49 50 51 52 53 54 55 56
        if not isinstance(program, Program):
            raise TypeError("program should be as Program type")
        if not isinstance(place, core.CPUPlace) and not isinstance(
                place, core.CUDAPlace):
            raise TypeError("place should be as CPUPlace/CUDAPlace type")
        if scope is None:
            scope = global_scope()
        if not isinstance(scope, core.Scope):
            raise TypeError("scope should be as Scope type or None")
        self.fuse_batch_norm(program, place, scope)
L
Luo Tao 已提交
57

L
Luo Tao 已提交
58
    def fuse_batch_norm(self, program, place, scope):
L
Luo Tao 已提交
59 60
        '''
        Transpile the program by fused batch normalization.
61 62 63 64 65
 
        The batch normalization followed the convolution or fully connected layer 
        can be integrated with them. Doing so will give us a forward acceleration, 
        especially in environments like mobile or embedded.
                    
L
Luo Tao 已提交
66 67 68 69 70
        For input :math:`X`:

        - Conv process:        :math:`X = input * W + bias` 
        - Batch norm process:  :math:`X' = (X - mean) / std` 
        - Scale Process:       :math:`Y = a * X' + b`
71 72 73

        After fuse into one operation:

L
Luo Tao 已提交
74 75 76 77
        .. math::

            Y &= (input * W + bias - mean) / std * a + b \\\\
              &= input * a * W / std + ((bias - mean) / std * a + b)
78 79

        The operator transformation is: 
L
Luo Tao 已提交
80

81
        - before:
L
Luo Tao 已提交
82

83 84
          - conv->batch_norm->any_other_op (bias == 0)
          - conv->elementwise_add->batch_norm->any_other_op (bias != 0)
L
Luo Tao 已提交
85
            
86
        - after: 
L
Luo Tao 已提交
87

88 89 90
          - conv->elementwise_add->any_other_op
        
        The transpile stages are:
L
Luo Tao 已提交
91

92
        1. insert elementwise_add op when bias == 0.
93
        2. fuse the batch_norm's parameters to conv and elementwise_add operators.
94 95 96
        3. remove batch_norm ops which are not used in any other ops.
        4. adjust the input of any_other_op to be the output of elementwise_add operator.
        5. remove unused variables.
97

L
Luo Tao 已提交
98 99 100 101 102
        Args:
            program (Program): program to transpile
            place (Place): inference place
            scope (Scope): inference Scope
        
103 104 105
        '''
        self.scope = scope
        self.place = place
106
        self.block = program.block(0)
107 108
        self.input_map = {}  # store the input names should be adjusted 

109
        i = 0
110 111
        while i < len(self.block.ops):
            current_op = self.block.ops[i]
112
            # TODO(luotao1): consider only conv2d now. fc would be delt later.
113
            if current_op.type in ['conv2d']:
L
Luo Tao 已提交
114 115 116
                # TODO(luotao1): consider single chain network now. 
                # For branch network, we counldn't use block.ops[i + 1] as 
                # the judgment condition.
117
                next_op = self.block.ops[i + 1]
118
                # conv2d without bias
119
                if (next_op.type == 'batch_norm'):
120 121 122
                    # insert bias op
                    bias_op = self._insert_bias_op(i + 1, current_op, next_op)
                    # fuse batch_norm
123
                    self._fuse_param(current_op, next_op, bias_op, 0)
124
                    # remove batch_norm_op
125
                    self.block.remove_op(i + 2)
126
                    i = i + 1
127 128 129 130 131 132 133 134 135
                # conv2d with bias, the next_op.type is elementwise_add
                elif (next_op.type == 'elementwise_add'):
                    next_next_op = self.block.ops[i + 2]
                    if (next_next_op.type == 'batch_norm'):
                        # fuse batch_norm
                        self._fuse_param(current_op, next_next_op, next_op, 1)
                        # remove batch_norm_op
                        self.block.remove_op(i + 2)
                        i = i + 1
136 137
            i = i + 1

138
        self._adjust_input()
139
        self._remove_unused_var()
L
Luo Tao 已提交
140 141 142
        # TODO(luotao): use clone() method to flush the program.desc in force, 
        # since some large program.desc will not be flushed immediately. 
        # And a better solution will be considered later.
L
Luo Tao 已提交
143
        program = program.clone()
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161

    # ====================== private transpiler functions =====================
    def _insert_bias_op(self, index, current_op, bn_op):
        '''
        Construct elementwise_add operator for adding bias 
        and insert it into program.
        
        :param index: insert location of bias_op
        :type index: Int
        :param current_op: current operator (conv or fc)
        :type current_op: Operator
        :param bn_op: batch norm operator
        :type bn_op: Operator
        :return: bias_op
        :rtype: Operator
        '''
        # The input of bias_op is current_op's output and Bias of bn_op
        # The output of bias_op is bn_op's output
162 163 164 165 166 167 168 169 170 171 172
        x_var = self.block.var(current_op.output("Output")[0])
        y_var = self.block.var(bn_op.input("Bias")[0])
        out_var = self.block.var(bn_op.output("Y")[0])

        bias_op = self.block.insert_op(
            index,
            type="elementwise_add",
            inputs={"X": x_var,
                    "Y": y_var},
            outputs={"Out": out_var},
            attrs={"axis": 1})  # dim_start=1
173 174
        return bias_op

175
    def _fuse_param(self, current_op, bn_op, bias_op, with_bias):
176 177 178 179 180 181 182 183 184
        '''
        fuse the batch_norm_op' parameters to current_op (conv or fc)
        
        :param current_op: current operator (conv or fc)
        :type current_op: Operator
        :param bn_op: batch norm operator
        :type bn_op: Operator
        :param bias_op: elementwise_add operator for adding bias
        :type bias_op: Operator
185 186
        :param with_bias: If current operator has bias, with_bias = 1; otherwise 0. 
        :type with_bias: Int
187 188
        '''

L
Luo Tao 已提交
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
        def _update_param(op, old_param_name, new_param):
            # For the sake of remaining the original variables the same as before,
            # create new variables in scope to store the new parameters.
            old_param_name = old_param_name[0]
            old_var = self.block.vars[old_param_name]
            new_param_name = old_param_name + '_fuse_bn'
            new_var = self.block.create_parameter(
                name=new_param_name.encode('ascii'),
                type=old_var.type,
                dtype=old_var.dtype,
                shape=old_var.shape)
            op.rename_input(old_param_name, new_param_name)
            self.scope.var(new_param_name)

            tensor = self.scope.find_var(new_param_name).get_tensor()
            tensor.set(np.array(new_param), self.place)
205 206

        def _load_param(param_name):
L
Luo Tao 已提交
207
            return np.array(self.scope.find_var(param_name[0]).get_tensor())
208 209 210 211 212 213 214 215 216 217 218 219

        bias_bn = _load_param(bn_op.input("Bias"))  #Bias
        scale_bn = _load_param(bn_op.input("Scale"))  #Scale
        mean_bn = _load_param(bn_op.input("Mean"))  #Mean
        var_bn = _load_param(bn_op.input("Variance"))  #Variance

        # TODO(luotao1): consider only conv2d now. fc would be delt later.
        current_param = _load_param(current_op.input("Filter"))
        std_bn = np.float32(np.sqrt(np.add(var_bn, 1e-5)))
        tmp = np.float32(np.divide(scale_bn, std_bn))

        # add bias of batch_norm_op to conv2d
220 221 222 223
        if with_bias:
            bias = _load_param(bias_op.input("Y"))
        else:
            bias = np.zeros(bias_bn.shape)
224 225 226 227 228 229 230 231 232
        bias = np.float32(
            np.add(np.multiply(np.subtract(bias, mean_bn), tmp), bias_bn))

        # re-compute weight of conv2d
        tmp = tmp.reshape(tmp.shape[0], -1)
        dst_param = current_param.reshape((tmp.shape[0], -1))
        dst_param = np.float32(np.multiply(dst_param, tmp))
        dst_param = dst_param.reshape(current_param.shape)

L
Luo Tao 已提交
233 234 235
        # update parameters
        _update_param(current_op, current_op.input("Filter"), dst_param)
        _update_param(bias_op, bias_op.input("Y"), bias)
236

237 238 239
        # collect the renamed input
        self.input_map[bn_op.output("Y")[0]] = bias_op.output("Out")[0]

240
    def _adjust_input(self):
241 242 243 244 245 246 247
        for i in range(len(self.block.ops)):
            current_op = self.block.ops[i]
            for input_arg in current_op.input_arg_names:
                if input_arg in self.input_map:
                    current_op.rename_input(input_arg,
                                            self.input_map[input_arg])

248 249
    def _remove_unused_var(self):
        '''
250
        remove unused varibles in program
251 252
        '''
        args = []
253 254 255 256
        for i in range(len(self.block.ops)):
            current_op = self.block.ops[i]
            args += current_op.input_arg_names
            args += current_op.output_arg_names
257 258
        args = list(set(args))  # unique the input and output arguments

259 260 261
        for var in self.block.vars.keys():
            if var not in args:
                self.block.remove_var(var)