Source code for mindspore.ops.composite.base

# This is the Python adaptation and derivative work of Myia (https://github.com/mila-iqia/myia/).
#
# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================

"""Basic composite operations."""
from functools import partial
from types import FunctionType

from mindspore import context
from ..._c_expression import EnvInstance_, GradOperation_, HyperMap_, Map_, MultitypeFuncGraph_, Tail_, \
    TupleAdd_, TupleSlice_, UnpackCall_, ZipOperation_, ListAppend_, TupleGetItemTensor_
from ...common import dtype as mstype
from ...common.api import ms_function, _pynative_exec, _wrap_func
from .. import functional as F
from ...common.tensor import Tensor
from .. import signature as sig

__all__ = [EnvInstance_, TupleAdd_, TupleSlice_, UnpackCall_, TupleGetItemTensor_]


[docs]def add_flags(fn=None, **flags): """ A decorator that adds a flag to the function. Note: Only supports bool value. Args: fn (Function): Function or cell to add flag. Default: None. flags (dict): Flags use kwargs. Default: None. Returns: Function, the function with added flags. Examples: >>> add_flags(net, predit=True) """ def deco(fn): # need set the attr and access on c++ if not hasattr(fn, "_mindspore_flags"): fn._mindspore_flags = {} fn._mindspore_flags.update({**flags}) return fn ret = deco if fn is not None: ret = deco(fn) return ret
[docs]def core(fn=None, **flags): """ A decorator that adds a flag to the function. By default, the function is marked as True, enabling to use this decorator to set flag to a graph. Args: fn (Function): Function to add flag. Default: None. flags (dict): The following flags can be set core, which indicates that this is a core function or other flag. Default: None. """ # need set the attr and access on c++ def deco(fn): fn._mindspore_flags = { 'core': True, **flags, } return fn if fn is not None: ret = deco(fn) else: ret = deco return ret
[docs]class GradOperation(GradOperation_): """ A higher-order function which is used to generate the gradient function for the input function. The gradient function generated by `GradOperation` higher-order function can be customized by construction arguments. Given an input function `net = Net()` that takes `x` and `y` as inputs, and has a parameter `z`, see `Net` in Examples. To generate a gradient function that returns gradients with respect to the first input (see `GradNetWrtX` in Examples). 1. Construct a `GradOperation` higher-order function with default arguments: `grad_op = GradOperation()`. 2. Call it with input function as argument to get the gradient function: `gradient_function = grad_op(net)`. 3. Call the gradient function with input function's inputs to get the gradients with respect to the first input: `grad_op(net)(x, y)`. To generate a gradient function that returns gradients with respect to all inputs (see `GradNetWrtXY` in Examples). 1. Construct a `GradOperation` higher-order function with `get_all=True` which indicates getting gradients with respect to all inputs, they are `x` and `y` in example function `Net()`: `grad_op = GradOperation(get_all=True)`. 2. Call it with input function as argument to get the gradient function: `gradient_function = grad_op(net)`. 3. Call the gradient function with input function's inputs to get the gradients with respect to all inputs: `gradient_function(x, y)`. To generate a gradient function that returns gradients with respect to given parameters (see `GradNetWithWrtParams` in Examples). 1. Construct a `GradOperation` higher-order function with `get_by_list=True`: `grad_op = GradOperation(get_by_list=True)`. 2. Construct a `ParameterTuple` that will be passed to the input function when constructing `GradOperation` higher-order function, it will be used as a parameter filter that determine which gradient to return: `params = ParameterTuple(net.trainable_params())`. 3. Call it with input function and `params` as arguments to get the gradient function: `gradient_function = grad_op(net, params)`. 4. Call the gradient function with input function's inputs to get the gradients with respect to given parameters: `gradient_function(x, y)`. To generate a gradient function that returns gradients with respect to all inputs and given parameters in the format of ((dx, dy), (dz))(see `GradNetWrtInputsAndParams` in Examples). 1. Construct a `GradOperation` higher-order function with `get_all=True` and `get_by_list=True`: `grad_op = GradOperation(get_all=True, get_by_list=True)`. 2. Construct a `ParameterTuple` that will be passed along input function when constructing `GradOperation` higher-order function: `params = ParameterTuple(net.trainable_params())`. 3. Call it with input function and `params` as arguments to get the gradient function: `gradient_function = grad_op(net, params)`. 4. Call the gradient function with input function's inputs to get the gradients with respect to all inputs and given parameters: `gradient_function(x, y)`. We can configure the sensitivity(gradient with respect to output) by setting `sens_param` as True and passing an extra sensitivity input to the gradient function, the sensitivity input should has the same shape and type with input function's output(see `GradNetWrtXYWithSensParam` in Examples). 1. Construct a `GradOperation` higher-order function with `get_all=True` and `sens_param=True`: `grad_op = GradOperation(get_all=True, sens_param=True)`. 2. Define `grad_wrt_output` as `sens_param` which works as the gradient with respect to output: `grad_wrt_output = Tensor(np.ones([2, 2]).astype(np.float32))`. 3. Call it with input function as argument to get the gradient function: `gradient_function = grad_op(net)`. 4. Call the gradient function with input function's inputs and `sens_param` to get the gradients with respect to all inputs: `gradient_function(x, y, grad_wrt_output)`. Args: get_all (bool): If True, get all the gradients with respect to inputs. Default: False. get_by_list (bool): If True, get all the gradients with respect to Parameter variables. If get_all and get_by_list are both False, get the gradient with respect to first input. If get_all and get_by_list are both True, get the gradients with respect to inputs and Parameter variables at the same time in the form of ((gradients with respect to inputs), (gradients with respect to parameters)). Default: False. sens_param (bool): Whether to append sensitivity (gradient with respect to output) as input. If sens_param is False, a 'ones_like(outputs)' sensitivity will be attached automatically. Default: False. Returns: The higher-order function which takes a function as argument and returns gradient function for it. Examples: >>> class Net(nn.Cell): >>> def __init__(self): >>> super(Net, self).__init__() >>> self.matmul = P.MatMul() >>> self.z = Parameter(Tensor(np.array([1.0], np.float32)), name='z') >>> def construct(self, x, y): >>> x = x * self.z >>> out = self.matmul(x, y) >>> return out >>> >>> class GradNetWrtX(nn.Cell): >>> def __init__(self, net): >>> super(GradNetWrtX, self).__init__() >>> self.net = net >>> self.grad_op = GradOperation() >>> def construct(self, x, y): >>> gradient_function = self.grad_op(self.net) >>> return gradient_function(x, y) >>> >>> x = Tensor([[0.5, 0.6, 0.4], [1.2, 1.3, 1.1]], dtype=mstype.float32) >>> y = Tensor([[0.01, 0.3, 1.1], [0.1, 0.2, 1.3], [2.1, 1.2, 3.3]], dtype=mstype.float32) >>> GradNetWrtX(Net())(x, y) Tensor(shape=[2, 3], dtype=Float32, [[1.4100001 1.5999999 6.6 ] [1.4100001 1.5999999 6.6 ]]) >>> >>> class GradNetWrtXY(nn.Cell): >>> def __init__(self, net): >>> super(GradNetWrtXY, self).__init__() >>> self.net = net >>> self.grad_op = GradOperation(get_all=True) >>> def construct(self, x, y): >>> gradient_function = self.grad_op(self.net) >>> return gradient_function(x, y) >>> >>> x = Tensor([[0.8, 0.6, 0.2], [1.8, 1.3, 1.1]], dtype=mstype.float32) >>> y = Tensor([[0.11, 3.3, 1.1], [1.1, 0.2, 1.4], [1.1, 2.2, 0.3]], dtype=mstype.float32) >>> GradNetWrtXY(Net())(x, y) (Tensor(shape=[2, 3], dtype=Float32, [[4.5099998 2.7 3.6000001] [4.5099998 2.7 3.6000001]]), Tensor(shape=[3, 3], dtype=Float32, [[2.6 2.6 2.6 ] [1.9 1.9 1.9 ] [1.3000001 1.3000001 1.3000001]])) >>> >>> class GradNetWrtXYWithSensParam(nn.Cell): >>> def __init__(self, net): >>> super(GradNetWrtXYWithSensParam, self).__init__() >>> self.net = net >>> self.grad_op = GradOperation(get_all=True, sens_param=True) >>> self.grad_wrt_output = Tensor([[0.1, 0.6, 0.2], [0.8, 1.3, 1.1]], dtype=mstype.float32) >>> def construct(self, x, y): >>> gradient_function = self.grad_op(self.net) >>> return gradient_function(x, y, self.grad_wrt_output) >>> >>> x = Tensor([[0.8, 0.6, 0.2], [1.8, 1.3, 1.1]], dtype=mstype.float32) >>> y = Tensor([[0.11, 3.3, 1.1], [1.1, 0.2, 1.4], [1.1, 2.2, 0.3]], dtype=mstype.float32) >>> GradNetWrtXYWithSensParam(Net())(x, y) (Tensor(shape=[2, 3], dtype=Float32, [[2.211 0.51 1.4900001] [5.588 2.68 4.07 ]]), Tensor(shape=[3, 3], dtype=Float32, [[1.52 2.82 2.14 ] [1.1 2.05 1.55 ] [0.90000004 1.55 1.25 ]])) >>> >>> class GradNetWithWrtParams(nn.Cell): >>> def __init__(self, net): >>> super(GradNetWithWrtParams, self).__init__() >>> self.net = net >>> self.params = ParameterTuple(net.trainable_params()) >>> self.grad_op = GradOperation(get_by_list=True) >>> def construct(self, x, y): >>> gradient_function = self.grad_op(self.net, self.params) >>> return gradient_function(x, y) >>> >>> x = Tensor([[0.8, 0.6, 0.2], [1.8, 1.3, 1.1]], dtype=mstype.float32) >>> y = Tensor([[0.11, 3.3, 1.1], [1.1, 0.2, 1.4], [1.1, 2.2, 0.3]], dtype=mstype.float32) >>> GradNetWithWrtParams(Net())(x, y) (Tensor(shape=[1], dtype=Float32, [21.536]),) >>> >>> class GradNetWrtInputsAndParams(nn.Cell): >>> def __init__(self, net): >>> super(GradNetWrtInputsAndParams, self).__init__() >>> self.net = net >>> self.params = ParameterTuple(net.trainable_params()) >>> self.grad_op = GradOperation(get_all=True, get_by_list=True) >>> def construct(self, x, y): >>> gradient_function = self.grad_op(self.net, self.params) >>> return gradient_function(x, y) >>> >>> x = Tensor([[0.1, 0.6, 1.2], [0.5, 1.3, 0.1]], dtype=mstype.float32) >>> y = Tensor([[0.12, 2.3, 1.1], [1.3, 0.2, 2.4], [0.1, 2.2, 0.3]], dtype=mstype.float32) >>> GradNetWrtInputsAndParams(Net())(x, y) ((Tensor(shape=[2, 3], dtype=Float32, [[3.52 3.9 2.6 ] [3.52 3.9 2.6 ]]), Tensor(shape=[3, 3], dtype=Float32, [[0.6 0.6 0.6 ] [1.9 1.9 1.9 ] [1.3000001 1.3000001 1.3000001]])), (Tensor(shape=[1], dtype=Float32, [12.902]),)) """ def __init__(self, get_all=False, get_by_list=False, sens_param=False): if not isinstance(get_all, bool): raise TypeError(f'get_all should be bool, but got {type(get_all)}') if not isinstance(get_by_list, bool): raise TypeError(f'get_by_list should be bool, but got {type(get_by_list)}') if not isinstance(sens_param, bool): raise TypeError(f'sens_param should be bool, but got {type(sens_param)}') self.get_all = get_all self.get_by_list = get_by_list self.sens_param = sens_param GradOperation_.__init__(self, 'grad', get_all, get_by_list, sens_param) self.grad_fn = None self.fn = None self.need_forward = False def _pynative_forward_run(self, args, kwargs, fn): """ Pynative forward run to build grad graph. """ if self.sens_param: args = args[:-1] for arg in args: if not isinstance(arg, Tensor): raise TypeError("grad inputs should be tensor in pynative mode") if isinstance(fn, FunctionType): _pynative_exec.set_grad_flag(True) _pynative_exec.new_graph(fn, *args, **kwargs) output = fn(*args, **kwargs) _pynative_exec.end_graph(fn, output, *args, **kwargs) else: if fn.already_run and not fn.requires_grad: raise ValueError("obj must set_grad.") if not fn.already_run: self.need_forward = True if self.need_forward: fn.set_grad() fn(*args, **kwargs) fn.already_run = False def __call__(self, fn, weights=None): grad_ = GradOperation(self.get_all, self.get_by_list, self.sens_param) if self.grad_fn is None or self.fn != fn: if context.get_context("mode") == context.GRAPH_MODE: if self.get_by_list: @ms_function(obj=fn) def after_grad(*args): return grad_(fn, weights)(*args) else: @ms_function(obj=fn) def after_grad(*args): return grad_(fn)(*args) else: @_wrap_func def after_grad(*args, **kwargs): self._pynative_forward_run(args, kwargs, fn) _pynative_exec.grad(grad_, fn, weights, *args, **kwargs) out = _pynative_exec(*args, **kwargs) _pynative_exec.clear() return out self.grad_fn = after_grad self.fn = fn return self.grad_fn
[docs]class MultitypeFuncGraph(MultitypeFuncGraph_): """ Generate overloaded functions. MultitypeFuncGraph is a class used to generate overloaded functions, considering different types as inputs. Initialize an `MultitypeFuncGraph` object with name, and use `register` with input types as the decorator for the function to be registed. And the object can be called with different types of inputs, and work with `HyperMap` and `Map`. Args: name (str): Operator name. read_value (bool): If the registered function not need to set value on Parameter, and all inputs will pass by value, set `read_value` to True. Default: False. Raises: ValueError: If failed to find find a matching function for the given arguments. Examples: >>> # `add` is a metagraph object which will add two objects according to >>> # input type using ".register" decorator. >>> from mindspore import Tensor >>> from mindspore.ops import Primitive, operations as P >>> from mindspore import dtype as mstype >>> >>> scala_add = Primitive('scala_add') >>> tensor_add = P.TensorAdd() >>> >>> add = MultitypeFuncGraph('add') >>> @add.register("Number", "Number") ... def add_scala(x, y): ... return scala_add(x, y) >>> @add.register("Tensor", "Tensor") ... def add_tensor(x, y): ... return tensor_add(x, y) >>> add(1, 2) 3 >>> add(Tensor(1, mstype.float32), Tensor(2, mstype.float32)) Tensor(shape=[], dtype=Float32, 3) """ def __init__(self, name, read_value=False): MultitypeFuncGraph_.__init__(self, name) self.entries = list() if read_value: self.set_signatures(( sig.make_sig('args', sig.sig_rw.RW_READ, sig.sig_kind.KIND_VAR_POSITIONAL),)) def __call__(self, *args): types = tuple(map(mstype.get_py_obj_dtype, args)) for sigs, fn in self.entries: if len(sigs) != len(types): continue if any(not mstype.issubclass_(type_, sig) for sig, type_ in zip(sigs, types)): continue output = fn(*args) return output raise ValueError("Cannot find fn match given args.")
[docs] def register(self, *type_names): """ Register a function for the given type string. Args: type_names (Union[str, :class:`mindspore.dtype`]): Inputs type names or types list. Return: decorator, a decorator to register the function to run, when called under the types described in `type_names`. """ def deco(fn): def convert_type(type_input): if isinstance(type_input, str): return mstype.typing.str_to_type(type_input) if not isinstance(type_input, mstype.Type): raise TypeError(f"MultitypeFuncGraph register only support str or {mstype.Type}") return type_input types = tuple(map(convert_type, type_names)) self.register_fn(type_names, fn) self.entries.append((types, fn)) return fn return deco
[docs]class HyperMap(HyperMap_): """ Hypermap will apply the set operation to input sequences. Apply the operations to every elements of the sequence or nested sequence. Different from `Map`, the `HyperMap` supports to apply on nested structure. Args: ops (Union[MultitypeFuncGraph, None]): `ops` is the operation to apply. If `ops` is `None`, the operations should be put in the first input of the instance. Inputs: - **args** (Tuple[sequence]) - If `ops` is `None`, all the inputs should be sequences with the same length. And each row of the sequences will be the inputs of the operation. If `ops` is not `None`, the first input is the operation, and the others are inputs. Outputs: Sequence or nested sequence, the sequence of output after applying the function. e.g. `operation(args[0][i], args[1][i])`. Examples: >>> from mindspore import dtype as mstype >>> nest_tensor_list = ((Tensor(1, mstype.float32), Tensor(2, mstype.float32)), ... (Tensor(3, mstype.float32), Tensor(4, mstype.float32))) >>> # square all the tensor in the nested list >>> >>> square = MultitypeFuncGraph('square') >>> @square.register("Tensor") ... def square_tensor(x): ... return F.square(x) >>> >>> common_map = HyperMap() >>> common_map(square, nest_tensor_list) ((Tensor(shape=[], dtype=Float32, 1), Tensor(shape=[], dtype=Float32, 4)), (Tensor(shape=[], dtype=Float32, 9), Tensor(shape=[], dtype=Float32, 16)) >>> square_map = HyperMap(square) >>> square_map(nest_tensor_list) ((Tensor(shape=[], dtype=Float32, 1), Tensor(shape=[], dtype=Float32, 4)), (Tensor(shape=[], dtype=Float32, 9), Tensor(shape=[], dtype=Float32, 16)) """ def __init__(self, ops=None): self.ops = ops if ops: HyperMap_.__init__(self, ops) else: HyperMap_.__init__(self) def __call__(self, *args): func = self.ops args_list = args hypermap = self if self.ops is None: func = args[0] args_list = args[1:] hypermap = partial(self, func) # is leaf if not isinstance(args_list[0], (tuple, list)): return func(*args_list) return tuple(map(hypermap, *args_list))
class Map(Map_): """ Map will apply the set operation on input sequences. Apply the operations to every elements of the sequence. Args: ops (Union[MultitypeFuncGraph, None]): `ops` is the operation to apply. If `ops` is `None`, the operations should be put in the first input of the instance. Default: None Inputs: - **args** (Tuple[sequence]) - If `ops` is not `None`, all the inputs should be the same length sequences, and each row of the sequences. e.g. If args length is 2, and for `i` in length of each sequence `(args[0][i], args[1][i])` will be the input of the operation. If `ops` is not `None`, the first input is the operation, and the other is inputs. Outputs: Sequence, the sequence of output after applying the function. e.g. `operation(args[0][i], args[1][i])`. Examples: >>> from mindspore import dtype as mstype >>> tensor_list = (Tensor(1, mstype.float32), Tensor(2, mstype.float32), Tensor(3, mstype.float32)) >>> # square all the tensor in the list >>> >>> square = MultitypeFuncGraph('square') >>> @square.register("Tensor") >>> def square_tensor(x): ... return F.square(x) >>> >>> common_map = Map() >>> common_map(square, tensor_list) (Tensor(shape=[], dtype=Float32, 1), Tensor(shape=[], dtype=Float32, 4), Tensor(shape=[], dtype=Float32, 9)) >>> square_map = Map(square) >>> square_map(tensor_list) (Tensor(shape=[], dtype=Float32, 1), Tensor(shape=[], dtype=Float32, 4), Tensor(shape=[], dtype=Float32, 9)) """ def __init__(self, ops=None): self.ops = ops if ops: Map_.__init__(self, ops) else: Map_.__init__(self) def __call__(self, *args): func = self.ops args_list = args if self.ops is None: func = args[0] args_list = args[1:] return tuple(map(func, *args_list)) class _ListAppend(ListAppend_): """ A metafuncgraph class that append one element to list. Args: name (str): The name of the metafuncgraph object. """ def __init__(self, name): ListAppend_.__init__(self, name) def __call__(self, *args): pass _append = _ListAppend("append") class _Tail(Tail_): """ A metafuncgraph class that generates tail elements of the tuple. Args: name (str): The name of the metafuncgraph object. """ def __init__(self, name): Tail_.__init__(self, name) def __call__(self, *args): pass tail = _Tail('tail') class _ZipOperation(ZipOperation_): """Generates a tuple of zip iterations for inputs.""" def __init__(self, name): ZipOperation_.__init__(self, name) def __call__(self, *args): pass zip_operation = _ZipOperation('zip_operation') """`zip_operation` will generate a tuple of zip iterations of inputs.""" env_get = MultitypeFuncGraph("env_get") @env_get.register("EnvType", "Tensor") def _tensor_env_get(env, parameter): """Used to get env.""" return F.env_getitem(env, F.ref_to_embed(parameter), F.zeros_like(parameter))