• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2024 Huawei Technologies Co., Ltd
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ============================================================================
15
16import time
17import stat
18import os
19from mindspore.nn import Cell
20from mindspore.ops.composite import GradOperation
21from mindspore.common import ParameterTuple
22
23
24class _Grad(Cell):
25    def __init__(self, grad, network, wrt_params=False, real_inputs_count=None):
26        super().__init__()
27        self.network = network
28        self.grad = grad
29        self.sens_param = self.grad.sens_param
30        self.wrt_params = wrt_params
31        self.real_inputs_count = real_inputs_count
32        if self.wrt_params:
33            self.params = ParameterTuple(self.network.trainable_params())
34
35    def __call__(self, *inputs):
36        if self.sens_param and self._dynamic_shape_inputs is not None:
37            # not support dynamic shape sens
38            if self.real_inputs_count is None:
39                dyn_inputs = self._dynamic_shape_inputs[:-1]
40                real_sens = inputs[-1:]
41            else:
42                idx = self.real_inputs_count
43                dyn_inputs = self._dynamic_shape_inputs[:idx]
44                real_sens = inputs[idx:]
45            static_sens = list(dyn_inputs) + list(real_sens)
46            super().set_inputs(*static_sens)
47
48        a = time.perf_counter()
49        out = super().__call__(*inputs)
50        b = time.perf_counter()
51        if os.environ.get("perf") == '1':
52            phase = os.environ.get("PHASE")
53            flags = os.O_WRONLY | os.O_CREAT
54            modes = stat.S_IWUSR | stat.S_IRUSR
55            with os.fdopen(os.open(phase, flags, modes), 'w') as f:
56                f.write(str(b - a))
57        return out
58
59    def construct(self, *inputs):
60        if self.wrt_params:
61            if self.real_inputs_count is None or self.sens_param is False:
62                return self.grad(self.network, self.params)(*inputs)
63            real_inputs = inputs[:self.real_inputs_count]
64            sense_param_inputs = inputs[self.real_inputs_count:]
65            return self.grad(self.network, self.params)(*real_inputs, sense_param_inputs)
66        if self.real_inputs_count is None or self.sens_param is False:
67            return self.grad(self.network)(*inputs)
68        real_inputs = inputs[:self.real_inputs_count]
69        sense_param_inputs = inputs[self.real_inputs_count:]
70        return self.grad(self.network)(*real_inputs, sense_param_inputs)
71
72
73class GradOfAllInputsAndParams(_Grad):
74    """
75    get grads of all inputs and params
76    """
77    def __init__(self, network, sens_param=True, real_inputs_count=None):
78        super().__init__(grad=GradOperation(get_all=True, get_by_list=True,
79                                            sens_param=sens_param),
80                         network=network, wrt_params=True, real_inputs_count=real_inputs_count)
81
82class GradOfFirstInput(_Grad):
83    """
84    get grad of first input
85    """
86
87    def __init__(self, network, sens_param=True, real_inputs_count=None):
88        super().__init__(grad=GradOperation(sens_param=sens_param),
89                         network=network, real_inputs_count=real_inputs_count)
90