math_op_patch.py 5.6 KB
Newer Older
Y
Yang Yu 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
# 
# 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.

from ..framework import Variable, unique_name
Y
Fix CI  
Yang Yu 已提交
16
from layer_function_generator import OpProtoHolder
17
from ..initializer import force_init_on_cpu
Y
Yang Yu 已提交
18 19 20 21 22

__all__ = ['monkey_patch_variable']


def monkey_patch_variable():
Y
Yang Yu 已提交
23
    def unique_tmp_name():
Y
Yang Yu 已提交
24 25 26 27 28 29 30 31 32 33 34
        return unique_name("tmp")

    def safe_get_dtype(var):
        try:
            dtype = var.dtype
        except:
            raise ValueError("Cannot get data type from %s", var.name)
        return dtype

    def create_tensor(block, value, dtype, shape):
        value = float(value)
Y
Yang Yu 已提交
35
        tmp_name = unique_tmp_name()
Y
Yang Yu 已提交
36 37 38 39
        var = block.create_var(name=tmp_name, shape=shape, dtype=dtype)
        block.append_op(
            type="fill_constant",
            outputs={'Out': [var]},
40 41 42 43 44 45
            attrs={
                'dtype': var.dtype,
                'shape': shape,
                'value': value,
                'force_cpu': force_init_on_cpu()
            })
Y
Yang Yu 已提交
46 47
        return var

Y
Yang Yu 已提交
48 49 50
    def create_scalar(block, value, dtype):
        return create_tensor(block, value, dtype, shape=[1])

Y
Yang Yu 已提交
51 52 53
    def create_tensor_with_batchsize(ref_var, value, dtype):
        assert isinstance(ref_var, Variable)
        value = float(value)
Y
Yang Yu 已提交
54
        tmp_name = unique_tmp_name()
Y
Yang Yu 已提交
55 56 57 58 59 60 61 62 63 64 65
        var = ref_var.block.create_var(name=tmp_name, dtype=dtype)
        ref_var.block.append_op(
            type='fill_constant_batch_size_like',
            outputs={'Out': [var]},
            inputs={'Input': [ref_var]},
            attrs={'shape': ref_var.shape,
                   'value': value})
        return var

    def astype(self, dtype):
        """
Y
Yang Yu 已提交
66
        Cast a variable to a specified data type.
Y
Yang Yu 已提交
67 68 69 70 71 72 73 74
        NOTE: The variable must be a Tensor
        Args:
            self(Variable): The source variable
            dtype: The target dtype

        Returns:
            Variable with new dtype
        """
Y
Yang Yu 已提交
75
        tmp_name = unique_tmp_name()
Y
Yang Yu 已提交
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
        out = self.block.create_var(name=tmp_name, dtype=dtype)
        self.block.append_op(
            type="cast",
            inputs={"X": [self]},
            outputs={"Out": [out]},
            attrs={"in_dtype": self.dtype,
                   "out_dtype": out.dtype})
        return out

    def _elemwise_method_creator_(method_name, op_type, reverse=False):
        def __impl__(self, other_var):
            lhs_dtype = safe_get_dtype(self)

            if not isinstance(other_var, Variable):
                if reverse:
                    has_batch_size = False
                    for elem in self.shape:
                        if elem < 0:
                            has_batch_size = True
                            break
                    if not has_batch_size:
                        other_var = create_tensor(
                            self.block,
                            other_var,
                            dtype=lhs_dtype,
                            shape=self.shape)
                    else:
                        other_var = create_tensor_with_batchsize(
                            self, other_var, lhs_dtype)
                else:
                    # add fill_op to self.block
                    other_var = create_scalar(
                        self.block, value=other_var, dtype=lhs_dtype)

            rhs_dtype = safe_get_dtype(other_var)
            if lhs_dtype != rhs_dtype:
                other_var = astype(other_var, lhs_dtype)
            if reverse:
                tmp = self
                self = other_var
                other_var = tmp

Y
Yang Yu 已提交
118
            tmp_name = unique_tmp_name()
Y
Yang Yu 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
            out = self.block.create_var(name=tmp_name, dtype=lhs_dtype)
            self.block.append_op(
                type=op_type,
                inputs={'X': [self],
                        'Y': [other_var]},
                outputs={'Out': out})
            return out

        comment = OpProtoHolder.instance().get_op_proto(op_type).comment

        __impl__.__doc__ = """
        {0}
        Args:
            self(Variable): left hand variable
            other_var(Variable|float|int): right hand variable 

        Returns:
            Variable
        """.format(comment)
        __impl__.__name__ = method_name
        return __impl__

    # inject methods
    for method_name, op_type, reverse in (
        ("__add__", "elementwise_add", False),
            # a+b == b+a. Do not need to reverse explicitly
        ("__radd__", "elementwise_add", False),
        ("__sub__", "elementwise_sub", False),
        ("__rsub__", "elementwise_sub", True),
        ("__mul__", "elementwise_mul", False),
            # a*b == b*a. Do not need to reverse explicitly
        ("__rmul__", "elementwise_mul", False),
        ("__div__", "elementwise_div", False),
Q
Qiao Longfei 已提交
152 153
        ("__rdiv__", "elementwise_div", True),
        ("__pow__", "elementwise_pow", False),
154 155 156
        ("__rpow__", "elementwise_pow", True),
            # for logical compare
        ("__eq__", "equal", False),
Q
qiaolongfei 已提交
157
        ("__ne__", "not_equal", False),
Q
qiaolongfei 已提交
158
        ("__lt__", "less_than", False),
Q
qiaolongfei 已提交
159
        ("__le__", "less_equal", False)):
Y
Yang Yu 已提交
160 161 162 163
        setattr(Variable, method_name,
                _elemwise_method_creator_(method_name, op_type, reverse))

    Variable.astype = astype