Static Graph Syntax Support

Overview

In graph mode, Python code is not executed by the Python interpreter. Instead, the code is compiled into a static computation graph, and then the static computation graph is executed.

There are two ways to use the graph mode. The first way is to call the @jit decorator to modify a function or a class member method, and then the decorated function or method will be compiled into a static computation graph. For details about how to use jit, click jit API document. The second way is to set ms.set_context(mode=ms.GRAPH_MODE), then write the code in the construct function of the Cell so that the code in the construct function will be compiled into a static computation graph. For details about the definition of Cell, click Cell API document.

Due to syntax parsing restrictions, the supported data types, syntax, and related operations during graph building are not completely consistent with the Python syntax. As a result, some usage is restricted. Borrowing the traditional JIT compilation idea, considers the unification of static and dynamic graphs from the perspective of graph mode and extends the syntax capabilities of graph patterns. The static graph provides a syntax experience close to that of the dynamic graph, so as to realize the unity of dynamic and static. In order to facilitate users to choose whether to extend the static graph syntax, the JIT syntax support level option ‘jit_syntax_level’ is provided, and its value must be in the range of [STRICT,LAX], and selecting ‘STRICT’ is considered to use the basic syntax and do not extend the static graph syntax. The default value is ‘LAX’, please refer to the Extended Syntaxes (LAX level) section of this article for more information. All backends are supported at all levels.

• STRICT: Only basic syntaxes is supported, and execution performance is optimal.

• LAX: Supporting more complex syntaxes, compatible with all Python syntax as much as possible.

The following describes the data types, syntax, and related operations supported during static graph building. These rules apply only to graph mode.

Basic Syntaxes (STRICT Level)

Constants and Variables Within Static Graphs

In static graphs, constants and variables are an important concept for understanding static graph syntax, and many syntaxes support different methods and degrees in the case of constant input and variable input. Therefore, before introducing the specific syntax supported by static graphs, this section first explains the concepts of constants and variables in static graphs.

In static graph mode, the operation of a program is divided into compilation period and execution period. During compilation, the program is compiled into an intermediate representation graph, and the program does not actually execute, but statically parses the intermediate representation through abstract deduction. This makes it impossible to guarantee that we will get the values of all intermediate representation nodes at compile time. Constants and variables are distinguished by their true values in the compiler.

• Constant: The amount of value that can be obtained during compilation.

• Variable: The amount of value that cannot be obtained during compilation.

In some cases, it is difficult to determine whether a quantity is a constant or a variable, and we can use ‘ops.isconstant’ to determine whether it is a constant or not. For example:

from mindspore import Tensor, jit, ops

a = Tensor([1])

@jit
def foo(a):
    b = Tensor([2])
    m = ops.isconstant(a)
    n = ops.isconstant(b)
    return m, n

In the above code, ‘a’ is the variable, so ‘m’ is ‘false’. ‘b’ is a constant, so ‘n’ is ‘True’.

Constants Generate Scenes

• Scalars, lists, and tuples entered as graph mode are constants (without using the mutable interface). For example:

  from mindspore import Tensor, jit
  
  a = 1
  b = [Tensor([1]), Tensor([2])]
  c = ["a", "b", "c"]
  
  @jit
  def foo(a, b, c):
      return a, b, c
  

  In the above code, enter ‘a’, ‘b’, ‘c’ are constants.

• The result of the constant operation is constant. For example:

  from mindspore import jit, Tensor
  
  @jit
  def foo():
      a = 1
      b = "2"
      c = Tensor([1, 2, 3])
      return a, b, c
  

  In the above code, enter ‘a’, ‘b’, ‘c’ are constants.

• Constant operations obtain a constant result. For example:

  from mindspore import jit, Tensor
  
  @jit
  def foo():
      a = Tensor([1, 2, 3])
      b = Tensor([1, 1, 1])
      c = a + b
      return c
  

  In the above code, ‘a’ and ‘b’ are constants of Tensor generated in the graph mode, so the result of their calculation is also constant. However, if one of them is a variable, its return value will also be a variable.

Variables Generate Scenes

• The return value of all mutable interfaces is a variable (whether mutable is used outside the graph or inside the graph). For example:

  from mindspore import Tensor, jit
  from mindspore.common import mutable
  
  a = mutable([Tensor([1]), Tensor([2])])
  
  @jit
  def foo(a):
      b = mutable(Tensor([3]))
      c = mutable((Tensor([1]), Tensor([2])))
      return a, b, c
  

  In the above code, ‘a’ is generated by calling the mutable interface outside the graph, ‘b’ and ‘c’ are generated by calling the mutable interface inside the graph, and ‘a’, ‘b’, and ‘c’ are variables.

• Tensors that are inputs to static graphs are variables. For example:

  from mindspore import Tensor, jit
  
  a = Tensor([1])
  b = (Tensor([1]), Tensor([2]))
  
  @jit
  def foo(a, b):
      return a, b
  

  In the above code, ‘a’ is the Tensor input as the graph pattern, so it is a variable. But ‘b’ is a tuple that is input to the graph schema, not a Tensor type, and even if its internal elements are Tensor, ‘b’ is a constant.

• What is calculated by variables is the variable

  If a quantity is the output of an operator, then it is in most cases constant. For example:

  from mindspore import Tensor, jit, ops
  
  a = Tensor([1])
  b = Tensor([2])
  
  @jit
  def foo(a, b):
      c = a + b
      return c
  

  In this case , ‘c’ is the result of calculations of ‘a’ and ‘b’ , and the inputs ‘a’ and ‘b’ used for the calculation are variables , so ‘c’ is also a variable.

Data Types

Built-in Python Data Types

Currently, the following built-in Python data types are supported: Number, String, List, Tuple, and Dictionary.

Number

Supporting int, float, and bool, but does not support complex numbers.

Number can be defined on the network. That is, the syntax y = 1, y = 1.2, and y = True are supported.

When the data is a constant, the value of the data can be achieved at compile time, the forcible conversion to Number is supported in the network. The syntax y = int(x), y = float(x), and y = bool(x) are supported. When the data is a variable, i.e., you can get the value only at runtime. It also supports data type conversion using built-in functions Python Built-in Functions such as int(), float() and bool(). For example:

from mindspore import Tensor, jit

@jit
def foo(x):
  out1 = int(11.1)
  out2 = int(Tensor([10]))
  out3 = int(x.asnumpy())
  return out1, out2, out3

res = foo(Tensor(2))
print("res[0]:", res[0])
print("res[1]:", res[1])
print("res[2]:", res[2])

The result is as follows:

res[0]: 11
res[1]: 10
res[2]: 2

Supporting returning numbers. For example:

import mindspore as ms

@ms.jit
def test_return_scalar(x, y):
    return x + y

res = test_return_scalar(ms.mutable(1), ms.mutable(2))
print(res)

3

String

String can be constructed on the network, i.e., support for using quotes (' or ") to create strings such as x = 'abcd' or y = "efgh". Convert constants to strings by means of str(). Support string concatenation, truncation, and the use of membership operators (in or not in) to determine whether a string contains the specified character. Support for formatting string output by inserting a value into a string with the string format %s. Support for using the format string function str.format() in constant scenarios.

For example:

from mindspore import jit

@jit
def foo():
  var1 = 'Hello!'
  var2 = "MindSpore"
  var3 = str(123)
  var4 = "{} is {}".format("string", var3)
  return var1[0], var2[4:9], var1 + var2, var2 * 2, "H" in var1, "My name is %s!" % var2, var4

res = foo()
print("res:", res)

The result is as follows:

res: ('H', 'Spore', 'Hello!MindSpore', 'MindSporeMindSpore', True, 'My name is MindSpore!', 'string is 123')

List

When ‘JIT_SYNTAX_LEVEL’ is set to ‘LAX’, static graph mode can support the inplace operation of some ‘List’ objects, see Supporting List Inplace Modification Operations.

The basic usage scenarios of ‘List’ are as follows:

• The graph mode supports creating Lists in graph.

  Support creating List objects within graph mode, and the elements of the List objects can contain any of the types supported by the graph mode, as well as multiple levels of nesting. For example:

  import numpy as np
  import mindspore as ms
  
  @ms.jit
  def generate_list():
    a = [1, 2, 3, 4]
    b = ["1", "2", "a"]
    c = [ms.Tensor([1]), ms.Tensor([2])]
    d = [a, b, c, (4, 5)]
    return d
  

  The above sample code, all List objects can be created normally.

• The graph mode supports returning List

  Before MindSpore version 2.0, List is converted to Tuple when the graph mode returns a List object. In MindSpore version 2.0, List objects can be returned. For example:

  import mindspore as ms
  
  @ms.jit
  def list_func():
      a = [1, 2, 3, 4]
      return a
  
  output = list_func()  # output: [1, 2, 3, 4]
  

  In the same way that a List is created within a graph mode, the graph mode returns a List object that can include any of the types supported by the graph mode, as well as multiple levels of nesting.

• The graph mode supports obtaining List objects from global variables

  import mindspore as ms
  
  global_list = [1, 2, 3, 4]
  
  @ms.jit
  def list_func():
      global_list.reverse()
      return global_list
  
  output = list_func()  # output: [4, 3, 2, 1]
  

  It should be noted that the list returned in the following pattern in the basic scenario is not the same object as the list of global variables, and when ‘JIT_SYNTAX_LEVEL’ is set to ‘LAX’, the returned object and the global object are unified objects.

• Graph mode supports List as input

  The graph mode supports List as input to static graphs. The elements of the List object used as input must be of an input type supported by the graph mode, which also supports multiple levels of nesting.

  import mindspore as ms
  
  list_input = [1, 2, 3, 4]
  
  @ms.jit
  def list_func(x):
      return x
  
  output = list_func()  # output: [1, 2, 3, 4]
  

  It should be noted that when ‘List’ is input as a static graph, it is always treated as a constant, regardless of the type of element inside it.

• Graph mode supports built-in methods for List

  The ‘List’ built-in method is described in detail below:

  • List Index Value

    Basic syntax: element = list_object[index].

    Basic semantics: Extract the element in the ‘List’ object in the ‘index’ bit (‘index’ starts at 0). Supporting multi-level index values.

    Index value ‘index’ supported types include ‘int’, ‘Tensor’, and ‘slice’. Among them, inputs of type ‘int’ and ‘Tensor’ can support constants and variables, and ‘slice’ internal data must be constants that can be determined at compile time.

    Examples are as follows:

    import mindspore as ms
    @ms.jit()
    def list_getitem_func():
        x = [[1, 2], 3, 4]
        a = x[0]
        b = x[0][ms.Tensor([1])]
        c = x[1:3:1]
        return a, b, c
    
    a, b, c = list_getitem_func()
    print('a:{}'.format(a))
    print('b:{}'.format(b))
    print('c:{}'.format(c))
    

    The results are as follows:

    a:[1, 2]
    b:2
    c:[3, 4]
    

  • List index assignment

    Basic syntax: list_object[index] = target_element.

    Basic semantics: Assign the element in the ‘List’ object at bit ‘index’ to ‘target_element’ (‘index’ starts at 0). Support for multi-tier index assignment.

    Index value ‘index’ supported types include ‘int’, ‘Tensor’, and ‘slice’. Among them, inputs of type ‘int’ and ‘Tensor’ can support constants and variables, and the internal data of ‘slice’ must be constant that can be determined at compile time.

    The index assignment object ‘target_element’ supports all data types supported by graph modes.

    Currently, the ‘List’ index assignment does not support the inplace operation, and a new object will be generated after the index is assigned. This operation will support the inplace operation in the future.

    Examples are as follows:

    import mindspore as ms
    
    @ms.jit()
    def test_setitem_func():
        x = [[0, 1], 2, 3, 4]
        x[1] = 10
        x[2] = "ok"
        x[3] = (1, 2, 3)
        x[0][1] = 88
        return x
    
    output = test_setitem_func()
    print('output:{}'.format(output))
    

    The results are as follows:

    output:[[0, 88], 10, 'ok', (1, 2, 3)]
    

  • List.append

    Basic syntax: list_object.append(target_element).

    Basic semantics: Append the element ‘target_element’ to the last list_object’ of the ‘List’ object.

    Currently, ‘List.append’ does not support the inplace operation, and a new object will be generated after index assignment. This operation will support the inplace operation in the future.

    Examples are as follows:

    import mindspore as ms
    
    @ms.jit()
    def test_list():
        x = [1, 2, 3]
        x.append(4)
        return x
    
    x = test_list()
    print('x:{}'.format(x))
    

    The results are as follows:

    x:[1, 2, 3, 4]
    

  • List.clear

    Basic syntax: list_object.clear().

    Base semantics: Empty the elements contained in the ‘List’ object ‘list_object’.

    Currently, ‘List.clear’ does not support inplace, and a new object will be generated after index assignment. This operation will support inplace in the future.

    Examples are as follows:

    import mindspore as ms
    
    @ms.jit()
    def test_list_clear():
        x = [1, 3, 4]
        x.clear()
        return x
    
    x = test_list_clear()
    print('x:{}'.format(x))
    

    The results are as follows:

    x:[]
    

  • List.extend

    Basic syntax: list_object.extend(target).

    Basic semantics: Insert all elements inside the ‘target’ to the end of the ‘List’ object ‘list_object’.

    The supported types for ‘target’ are ‘Tuple’, ‘List’, and ‘Tensor’. Among them, if the ‘target’ type is ‘Tensor’, the ‘Tensor’ will be converted to ‘List’ before inserting it.

    Examples are as follows:

    import mindspore as ms
    
    @ms.jit()
    def test_list_extend():
        x1 = [1, 2, 3]
        x1.extend((4, "a"))
        x2 = [1, 2, 3]
        x2.extend(ms.Tensor([4, 5]))
        return x1, x2
    
    output1, output2 = test_list_extend()
    print('output1:{}'.format(output1))
    print('output2:{}'.format(output2))
    

    The results are as follows:

    output1:[1, 2, 3, 4, 'a']
    output2:[1, 2, 3, Tensor(shape=[1], dtype=Int64, value= [4]), Tensor(shape=[1], dtype=Int64, value= [5])]
    

  • List.pop

    Basic syntax: pop_element = list_object.pop(index=-1).

    Basic semantics: Remove the ‘index’ element of the ‘List’ object ‘list_object’ from the ‘list_object’ and return the element.

    The ‘index’ requires that it must be a constant ‘int’, and when ‘list_object’ has a length of ‘list_obj_size’, ‘index’ has a value range of ‘[-list_obj_size,list_obj_size-1]’. ‘index’ is a negative number representing the number of digits from back to front. When no ‘index’ is entered, the default value is -1, i.e. the last element is removed.

    import mindspore as ms
    
    @ms.jit()
    def test_list_pop():
        x = [1, 2, 3]
        b = x.pop()
        return b, x
    
    pop_element, res_list = test_list_pop()
    print('pop_element:{}'.format(pop_element))
    print('res_list:{}'.format(res_list))
    

    The results are as follows:

    pop_element:3
    res_list:[1, 2]
    

  • List.reverse

    Basic syntax: list_object.reverse().

    Basic semantics: Reverse the order of the elements of the ‘List’ object ‘list_object’.

    Examples are as follows:

    import mindspore as ms
    
    @ms.jit()
    def test_list_reverse():
        x = [1, 2, 3]
        x.reverse()
        return x
    
    output = test_list_reverse()
    print('output:{}'.format(output))
    

    The results are as follows:

    output1:[3, 2, 1]
    

  • List.insert

    Basic syntax: list_object.insert(index, target_obj).

    Basic semantics: insert ‘target_obj’ into the ‘index’ bit of ‘list_object’.

    The ‘index’ requirement must be a constant ‘int’. If the length of ‘list_object’ is ‘list_obj_size’. When ‘index < -list_obj_size’, insert the first place in ‘List’. When ‘index >= -list_obj_size’, insert at the end of ‘List’. A negative ‘index’ represents the number of digits from back to front.

    Examples are as follows:

    import mindspore as ms
    
    @ms.jit()
    def test_list_insert():
        x = [1, 2, 3]
        x.insert(3, 4)
        return x
    
    output = test_list_insert()
    print('output:{}'.format(output))
    

    The results are as follows:

    output:[1, 2, 3, 4]
    

Tuple

Tuple can be constructed on the network, that is, the syntax y = (1, 2, 3) is supported. The elements of the tuple Tuple cannot be modified, but indexed access to elements in the tuple Tuple is supported, and concatenated combinations of tuples are supported.

• Supported index values

  Support accessing elements in the tuple Tuple using square brackets plus subscripted indexes. The index value can be int, slice, Tensor, and multi-level index value. That is, the syntax data = tuple_x[index0][index1]... is supported.

  Restrictions on the index value Tensor are as follows:

  • Tuple stores Cell. Each Cell must be defined before a tuple is defined. The number of input parameters, input parameter type, and input parameter shape of each Cell must be the same. The number of outputs of each Cell must be the same. The output type must be the same as the output shape.

  • The index Tensor is a scalar Tensor whose dtype is int32. The value range is [-tuple_len, tuple_len), and negative index is not supported in Ascend backend.

  • CPU, GPU and Ascend backend is supported.

  An example of the int and slice indexes is as follows:

  import numpy as np
  import mindspore as ms
  
  t = ms.Tensor(np.array([1, 2, 3]))
  
  @ms.jit()
  def test_index():
      x = (1, (2, 3, 4), 3, 4, t)
      y = x[1][1]
      z = x[4]
      m = x[1:4]
      n = x[-4]
      return y, z, m, n
  
  y, z, m, n = test_index()
  print('y:{}'.format(y))
  print('z:{}'.format(z))
  print('m:{}'.format(m))
  print('n:{}'.format(n))
  

  The result is as follows:

  y:3
  z:[1 2 3]
  m:((2, 3, 4), 3, 4)
  n:(2, 3, 4)
  

  An example of the Tensor index is as follows:

  import mindspore as ms
  from mindspore import nn, set_context
  
  set_context(mode=ms.GRAPH_MODE)
  
  class Net(nn.Cell):
      def __init__(self):
          super(Net, self).__init__()
          self.relu = nn.ReLU()
          self.softmax = nn.Softmax()
          self.layers = (self.relu, self.softmax)
  
      def construct(self, x, index):
          ret = self.layers[index](x)
          return ret
  
  x = ms.Tensor([-1.0], ms.float32)
  
  net = Net()
  ret = net(x, 0)
  print('ret:{}'.format(ret))
  

  The result is as follows:

  ret:[0.]
  

• Support connection combinations

  Similar to the string String, tuples support combining using + and * to get a new tuple Tuple, for example:

  import mindspore as ms
  
  @ms.jit()
  def test_index():
      x = (1, 2, 3)
      y = (4, 5, 6)
      return x + y, x * 2
  
  out1, out2 = test_index()
  print('out1:{}'.format(out1))
  print('out2:{}'.format(out2))
  

  The result is as follows:

  out1:(1, 2, 3, 4, 5, 6)
  out2:(1, 2, 3, 1, 2, 3)
  

Dictionary

Dictionary can be constructed on the network. Each key value key:value is separated by a colon :, and each key value pair is separated by a comma ,. The entire dictionary contains the key-value pairs using curly braces {}. That is, the syntax y = {"a": 1, "b": 2} is supported.

The key is unique, and if there are multiple identical keys in the dictionary, the duplicate keys are finalized with the last one and the value value can be non-unique. The key key needs to be guaranteed to be immutable. Currently, the key can be String, Number, constant Tensor, or Tuple that contains these types. The value can be Number, Tuple, Tensor, List or Dictionary.

• Supported APIs

  keys: extracts all key values from dict to form Tuple and return it.

  values: extracts all value values from dict to form Tuple and return it.

  items: extracts Tuple composed of each pair of value values and key values in dict to form List and return it.

  get: dict.get(key[, value]) returns the value value corresponding to the specified key, if the specified key does not exist, the default value None or the set default value value is returned .

  clear: removes all elements in dict.

  has_key: dict.has_key(key) determines whether the specified key exists in dict.

  update: dict1.update(dict2) updates the elements in dict2 to dict1.

  fromkeys: dict.fromkeys(seq([, value])) is used to create a new Dictionary, using the elements in the sequence seq as the key of the Dictionary, and the value is initial value corresponding to all key.

  The example is as follows, where the ‘x’ and ‘new_dict’ in the return value are a ‘Dictionary’, and the support is extended under the JIT syntax support level option LAX in graph mode, for more advanced use of Dictionary, please refer to the Supporting the high-level usage of Dictionary section of this article.

  import mindspore as ms
  import numpy as np
  
  x = {"a": ms.Tensor(np.array([1, 2, 3])), "b": ms.Tensor(np.array([4, 5, 6])), "c": ms.Tensor(np.array([7, 8, 9]))}
  
  @ms.jit()
  def test_dict():
      x_keys = x.keys()
      x_values = x.values()
      x_items = x.items()
      value_a = x.get("a")
      check_key = x.has_key("a")
      y = {"a": ms.Tensor(np.array([0, 0, 0]))}
      x.update(y)
      new_dict = x.fromkeys("abcd", 123)
      return x_keys, x_values, x_items, value_a, check_key, x, new_dict
  
  x_keys, x_values, x_items, value_a, check_key, new_x, new_dict = test_dict()
  print('x_keys:{}'.format(x_keys))
  print('x_values:{}'.format(x_values))
  print('x_items:{}'.format(x_items))
  print('value_a:{}'.format(value_a))
  print('check_key:{}'.format(check_key))
  print('new_x:{}'.format(new_x))
  print('new_dict:{}'.format(new_dict))
  

  The result is as follows:

  x_keys:('a', 'b', 'c')
  x_values:(Tensor(shape=[3], dtype=Int64, value= [1, 2, 3]), Tensor(shape=[3], dtype=Int64, value= [4, 5, 6]), Tensor(shape=[3], dtype=Int64, value= [7, 8, 9]))
  x_items:[('a', Tensor(shape=[3], dtype=Int64, value= [1, 2, 3])), ('b', Tensor(shape=[3], dtype=Int64, value= [4, 5, 6])), ('c', Tensor(shape=[3], dtype=Int64, value= [7, 8, 9]))]
  value_a:[1 2 3]
  check_key:True
  new_x:{'a': Tensor(shape=[3], dtype=Int64, value= [0, 0, 0]), 'b': Tensor(shape=[3], dtype=Int64, value= [4, 5, 6]), 'c': Tensor(shape=[3], dtype=Int64, value= [7, 8, 9])}
  new_dict:{'a': 123, 'b': 123, 'c': 123, 'd': 123}
  

MindSpore User-defined Data Types

Currently, MindSpore supports the following user-defined data types: Tensor, Primitive, and Cell.

Tensor

For details of Tensor, click Tensor API document.

Supporting creating and using Tensor. The code case is as follows, Tensor(1, dtype=mstype.int32) is extended support under the graph mode JIT syntax support level option LAX.

import mindspore.nn as nn
import mindspore as ms

class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()

    def construct(self):
        return ms.Tensor(1, dtype=ms.int32)

ms.set_context(mode=ms.GRAPH_MODE)

net = Net()
print(net())

1

The above example uses the interface of ‘Tensor’ class to create a ‘Tensor’. In some cases, it may be necessary to create a runtime ‘Tensor’, that is, the ‘Tensor’ data that cannot get the value at compile time, in which case you can use the above class ‘ms.Tensor’ interface to create ‘Tensor’, or tensor function interfaceto create a Tensor. The code example is shown below.

The difference between it and the ‘Tensor’ class interface lies in the addition of the Annotation Type tag inside, which can specify the output ‘Tensor dtype’ during type inference stage to avoid the generation of ‘AnyType’, with dtype set. When dynamically creating ‘Tensor’ at runtime, we recommend using this method to create ‘Tensor’ and hope that users can pass in the expected dtype type to avoid the generation of ‘AnyType’.

import mindspore as ms
import mindspore.nn as nn

class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()

    @ms.jit
    def construct(self, x):
        return ms.tensor(x.asnumpy(), dtype=ms.float32)

ms.set_context(mode=ms.GRAPH_MODE)
net = Net()
x = ms.Tensor(1, dtype=ms.int32)
print(net(x))

1.0

Primitive

Currently, Primitive and its subclass instances can be constructed in construct.

However, during call, the parameter can be specified only in position parameter mode, and cannot be specified in the key-value pair mode.

For example:

import mindspore as ms
from mindspore import nn, ops, Tensor, set_context
import numpy as np

set_context(mode=ms.GRAPH_MODE)

class Net(nn.Cell):
    def __init__(self):
        super().__init__()

    def construct(self, x):
        reduce_sum = ops.ReduceSum(True) #`Primitive` and its subclass instances can be constructed in construct.
        ret = reduce_sum(x, axis=2)
        return ret

x = Tensor(np.random.randn(3, 4, 5, 6).astype(np.float32))
net = Net()
ret = net(x)
print('ret.shape:{}'.format(ret.shape))

In the network defined above, the parameters of reduce_sum(x, axis=2) cannot be specified in the key-value pair mode. The parameter can be specified only in position parameter mode, that is, reduce_sum(x, 2).

The error is reported as follows:

TypeError: Only supported positional parameter type for python primitive, but got keyword parameter type.

Currently, the attributes and APIs related to Primitive and its subclasses cannot be called on the network.

For details about the defined Primitive, click Primitive API document.

Cell

Currently, Cell and its subclass instances can be constructed on the network. That is, the syntax cell = Cell(args...) is supported.

However, during call, the parameter can be specified only in position parameter mode, and cannot be specified in the key-value pair mode. That is, the syntax cell = Cell(arg_name=value) is not supported.

Currently, the attributes and APIs related to Cell and its subclasses cannot be called on the network unless they are called through self in construct of Cell.

For details about the definition of Cell, click Cell API document.

Parameter

Parameter is a variable tensor, indicating the parameters that need to be updated during network training.

For details about the definition of Parameter, click Parameter API document.

Operators

Arithmetic operators and assignment operators support the Number and Tensor operations, as well as the Tensor operations of different dtype. For more details, please refer to Operators

Primaries

Primaries represent the most tightly bound operations of the language.

Attribute References and Attribute Modification

An attribute reference is a primary followed by a period and a name.

Using attribute references as l-values in Cell instances of MindSpore requires the following requirements:

• The modified attribute belongs to this cell object, i.e. it must be self.xxx.

• The attribute is initialized in Cell’s ‘init’ function and is of type Parameter.

When the JIT syntax support level option is ‘LAX’, can support attribute modification in more situations, see Support Attribute Setting and Modification.

Examples are as follows:

import mindspore as ms
from mindspore import nn, set_context

set_context(mode=ms.GRAPH_MODE)

class Net(nn.Cell):
    def __init__(self):
        super().__init__()
        self.weight = ms.Parameter(ms.Tensor(3, ms.float32), name="w")
        self.m = 2

    def construct(self, x, y):
        self.weight = x  # The conditions are met, they can be modified
        # self.m = 3     # self.m is not of type Parameter and modification is prohibited
        # y.weight = x   # y is not self, modification is prohibited
        return x

net = Net()
ret = net(1, 2)
print('ret:{}'.format(ret))

The results are as follows:

ret:1

Index Value

Index value of a sequence Tuple, List, Dictionary, Tensor which called subscription in Python.

Index value of Tuple refers to chapter Tuple of this page.

Index value of List refers to chapter List of this page.

Index value of Dictionary refers to chapter Dictionary of this page.

Index value of Tensor refers to Tensor index value document.

Calls

A call calls a callable object (e.g., Cell or Primitive) with a possibly empty series of arguments.

For example:

import mindspore as ms
from mindspore import nn, ops, set_context
import numpy as np

set_context(mode=ms.GRAPH_MODE)

class Net(nn.Cell):
    def __init__(self):
        super().__init__()
        self.matmul = ops.MatMul()

    def construct(self, x, y):
        out = self.matmul(x, y)  # A call of Primitive
        return out

x = ms.Tensor(np.ones(shape=[1, 3]), ms.float32)
y = ms.Tensor(np.ones(shape=[3, 4]), ms.float32)
net = Net()
ret = net(x, y)
print('ret:{}'.format(ret))

The result is as follows:

ret:[[3. 3. 3. 3.]]

Statements

Currently supported Python statements include raise statement, assert statement, pass statement, return statement, break statement, continue statement, if statement, for statement, while statement, with statement, list comprehension, generator expression and function definition statement. For more details, please refer to Statements

Python Built-in Functions

Currently supported Python built-in functions include int, float, bool, str, list, tuple, getattr, hasattr, len, isinstance, all, any, round, max, min , sum, abs, partial, map, range, enumerate, super, pow, filter. The use of built-in functions in graph mode is similar to the corresponding Python built-in functions. For more details, please refer to Python Built-in Functions.

Network Definition

Network Input parameters

The input parameters of the outermost network only can be lool, int, float, Tensor, None, mstype.number(mstype.bool, mstype.int, mstype.float, mstype.uint), List or Tuple that contains these types, and Dictionary whose values are these types.

If you want to use other types of input for the network, please transfer them to the network while initializing network, and save them as network attributes, then use in the construct. The input parameters of inner network do not have this restriction.

The code example is shown below. In the defined Net network, a flag parameter of string type is transferred as the network attribute self.flag during initialization, and then the attribute self.flag is used in the construct.

import mindspore as ms
from mindspore import nn

ms.set_context(mode=ms.GRAPH_MODE)

class Net(nn.Cell):
    def __init__(self, flag):
        super(Net, self).__init__()
        self.flag = flag

    def construct(self, x):
        if self.flag == "ok":
            return x + 1
        return x - 1

flag = "ok"
net = Net(flag)
ret = net(ms.Tensor([5]))
print('ret:{}'.format(ret))

The result is as follows:

ret:[6]

While calculating gradient for outermost network, only Tensor input could be calculated, input of other type will be ignored.

The code example is shown below. Among the input parameter (x, y,  z) of outermost network, x and z are Tensor type but y is not. While grad_net calculating gradient of the input parameters (x, y, z) for the network, gradient of y is automatically ignored. Only gradients of x and z are calculated, and (grad_x, grad_y) is returned.

import numpy as np
import mindspore as ms
from mindspore import nn

ms.set_context(mode=ms.GRAPH_MODE)

class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()

    def construct(self, x, y, z):
        return x + y + z

class GradNet(nn.Cell):
    def __init__(self, net):
        super(GradNet, self).__init__()
        self.forward_net = net

    def construct(self, x, y, z):
        return ms.grad(self.forward_net, grad_position=(0, 1, 2))(x, y, z)

input_x = ms.Tensor([1])
input_y = 2
input_z = ms.Tensor([3])

net = Net()
grad_net = GradNet(net)
ret = grad_net(input_x, input_y, input_z)
print('ret:{}'.format(ret))

The result is as follows:

ret:(Tensor(shape=[1], dtype=Int64, value= [1]), Tensor(shape=[1], dtype=Int64, value= [1]))

Syntax Constraints of Basic Syntaxes

The execution graph in graph mode is converted from source code, and not all Python syntax can support it. The following describes some of the syntax constraints that exist under the basic syntax. More network compilation problems can be found in Network compilation.

1. When an undefined class member is used in the construct function, AttributeError exception will be thrown. For example:

   import mindspore as ms
   from mindspore import nn, set_context
   
   set_context(mode=ms.GRAPH_MODE)
   
   class Net(nn.Cell):
       def __init__(self):
           super(Net, self).__init__()
   
       def construct(self, x):
           return x + self.y
   
   net = Net()
   net(1)
   

   The result is as follows:

   AttributeError: External object has no attribute y
   

2. Class methods modified by classmethod in nn.Cell are not supported. For example:

   import mindspore as ms
   
   ms.set_context(ms.GRAPH_MODE)
   
    class Net(ms.nn.Cell):
    @classmethod
    def func(cls, x, y):
        return x + y
   
    def construct(self, x, y):
        return self.func(x, y)
   
    net = Net()
    out = net(ms.Tensor(1), ms.Tensor(2))
    print(out)
   

   The result is as follows:

   TypeError: too many positional arguments
   

3. In graph mode, some Python syntax is difficult to convert to intermediate MindIR in graph mode. For Python keywords, there are some keywords that are not supported in graph mode: AsyncFunctionDef, ClassDef, Delete, AnnAssign, AsyncFor, AsyncWith, Match, Try, Import, ImportFrom, Nonlocal, NamedExpr, Set, SetComp, DictComp, Await, Yield, YieldFrom, Starred. If the relevant syntax is used in graph mode, an error message will alert the user.

   If you use the Try statement, the following example is used:

   import mindspore as ms
   
   @ms.jit
   def test_try_except(x, y):
       global_out = 1
       try:
           global_out = x / y
       except ZeroDivisionError:
           print("division by zero, y is zero.")
       return global_out
   
   test_try_except_out = test_try_except(1, 0)
   print("out:", test_try_except_out)
   

   The result is as follows:

   RuntimeError: Unsupported statement 'Try'.
   

4. Benchmarking Python built-in data types, except for Built-in Python Data Types supported in the current graph mode, complex ‘complex’ and collection ‘set’ types are not supported. Some high-level uses of the list ‘list’ and dictionary ‘dictionary’ are not supported in the basic syntax scenario, and need to be supported when the JIT syntax support level option ‘jit_syntax_level’ is ‘LAX’, please refer to the Extended Syntaxes (LAX level) section of this article for more information.

5. In the basic syntax scenario, in addition to the Python Built-in Functions supported in the current graph mode, there are still some built-in functions that are not supported in graph mode. For example: basestring, bin, bytearray, callable, chr, cmp, compile, delattr, dir, divmod, eval, execfile, file, frozenset, hash, hex, id, input, issubclass, iter, locals, long, memoryview, next, object, oct, open, ord, property, raw_input, reduce, reload, repr, reverse, set, slice, sorted, unichr, unicode, vars, xrange, __import__.

6. Python provides a number of third-party libraries that usually need to be called via import statements. In graph mode, when the JIT syntax support level is ‘STRICT’, you cannot directly use third-party libraries. If you need to use the data types of third-party libraries in graph mode or call methods of third-party libraries, you need to support them only if the JIT syntax support level option ‘jit_syntax_level’ is ‘LAX’, please refer to the Calling the Third-party Libraries section in Extended Syntaxes (LAX level) of this article.

7. In graph mode, when the JIT syntax support level is ‘STRICT’, you cannot directly use objects, properties, and methods of custom classes. If you need to use custom classes in graph mode, refer to the Supporting the Use of Custom Classes section of Extended Syntaxes (LAX level) in this article.

   For example:

   import mindspore as ms
   
   ms.set_context(mode=ms.GRAPH_MODE)
   
   class GetattrClass():
       def __init__(self):
           self.attr1 = 99
           self.attr2 = 1
   
       def method1(self, x):
           return x + self.attr2
   
   class GetattrClassNet(ms.nn.Cell):
       def __init__(self):
           super(GetattrClassNet, self).__init__()
           self.cls = GetattrClass()
   
       def construct(self):
           return self.cls.method1(self.cls.attr1)
   
   net = GetattrClassNet()
   out = net()
   assert out == 100
   

   There will be related errors:

   TypeError: Do not support to convert <class '__main__.GetattrClass'> object into graph node.
   

Extended Syntaxes (LAX level)

The following mainly introduces the static graph syntax supported by the current extension.

Calling the Third-party Libraries

• Third-party libraries

  1. Python built-in modules and Python standard libraries, such as os, sys, math, time and other modules.

  2. Third-party code libraries. Their module paths are under the site-packages directory of the Python installation directory, which need to be installed first and then imported, such NumPy and Scipy. It should be noted that MindSpore suites such as mindyolo and mindflow are not treated as third-party libraries.

  3. Modules specified by the environment variable MS_JIT_IGNORE_MODULES. In contrast, there is the environment variable MS_JIT_MODULES. For more details, please refer to Environment Variables.

• Supporting data types of third-party libraries, allowing calling and returning objects of third-party libraries.

  The code example is as follows.

  import numpy as np
  import mindspore as ms
  
  @ms.jit
  def func():
      a = np.array([1, 2, 3])
      b = np.array([4, 5, 6])
      out = a + b
      return out
  
  print(func())
  

  The result is as follows:

  [5 7 9]
  

• Supporting calling methods of third-party libraries.

  The code example is as follows.

  from scipy import linalg
  import mindspore as ms
  
  @ms.jit
  def func():
      x = [[1, 2], [3, 4]]
      return linalg.qr(x)
  
  out = func()
  print(out[0].shape)
  

  The result is as follows:

  (2, 2)
  

• Supporting creating Tensor instances by using the data types of the third-party library NumPy.

  The code example is as follows.

  import numpy as np
  import mindspore as ms
  
  @ms.jit
  def func():
      x = np.array([1, 2, 3])
      out = ms.Tensor(x) + 1
      return out
  
  print(func())
  

  The result is as follows:

  [2, 3, 4]
  

• The assignment of subscripts for data types in third-party libraries is not currently supported.

  The code example is as follows.

  import numpy as np
  import mindspore as ms
  
  @ms.jit
  def func():
      x = np.array([1, 2, 3])
      x[0] += 1
      return ms.Tensor(x)
  
  res = func()
  print("res: ", res)
  

  The error message is reported as follows:

  RuntimeError: For operation 'setitem', current input arguments types are <External, Number, Number>. The 1-th argument type 'External' is not supported now.
  

Supporting the Use of Custom Classes

Custom classes can be used in graph mode, and classes can be instantiated and object properties and methods can be used.

For example, where ‘GetattrClass’ is a user-defined class that does not use the ‘@ms_class’ decoration and does not inherit ‘nn. Cell`。

import mindspore as ms

ms.set_context(mode=ms.GRAPH_MODE)

class GetattrClass():
    def __init__(self):
        self.attr1 = 99
        self.attr2 = 1

    def method1(self, x):
        return x + self.attr2

class GetattrClassNet(ms.nn.Cell):
    def __init__(self):
        super(GetattrClassNet, self).__init__()
        self.cls = GetattrClass()

    def construct(self):
        return self.cls.method1(self.cls.attr1)

net = GetattrClassNet()
out = net()
assert out == 100

Basic Operators Support More Data Type

In the syntax of graph mode, the following basic operators in the list is overloaded: [‘+’, ‘-’, ‘*’, ‘/’, ‘//’, ‘%’, ‘**’, ‘<<’, ‘>>’, ‘&’, ‘|’, ‘^’, ‘not’, ‘==’, ‘!=’, ‘<’, ‘>’, ‘<=’, ‘>=’, ‘in’, ‘not in’, ‘y=x[0]’]. For more details, please refer to Operators. When getting unsupported input type, those operators need to use extended static graph syntax to support, and make the output consistent with the output in the pynative mode.

The code example is as follows.

import mindspore as ms
import mindspore.nn as nn
from mindspore import Tensor
ms.set_context(mode=ms.GRAPH_MODE)

class InnerClass(nn.Cell):
    def construct(self, x, y):
        return x.asnumpy() + y.asnumpy()

net = InnerClass()
ret = net(Tensor([4, 5]), Tensor([1, 2]))
print(ret)

[5 7]

In the example above, since the output of x.asnumpy() is numpy.ndarray and is an unsupported input type of + in the graph mode, x.asnumpy() + y.asnumpy() will be supported by static graph syntax.

In another example:

class InnerClass(nn.Cell):
    def construct(self):
        return (None, 1) in ((None, 1), 1, 2, 3)

net = InnerClass()
print(net())

True

tuple in tupleis an unsupported operation in original graph mode，and will be supported by static graph syntax.

Base Type

Use the JIT Fallback feature to extend support for Python’s native data types ‘List’, ‘Dictionary’, ‘None’.

Supporting List Inplace Modification Operations

The list ‘List’ and tuple ‘Tuple’ are the most basic sequential built-in types in Python, and the core difference between ‘List’ and ‘Tuple’ is that ‘List’ is an object that can be changed, while ‘Tuple’ cannot be changed. This means that once ‘Tuple’ is created, it cannot be changed without changing the object address. ‘List’, on the other hand, can modify an object without changing its address through a series of inplace operations. For example:

a = [1, 2, 3, 4]
a_id = id(a)
a.append(5)
a_after_id = id(a)
assert a_id == a_after_id

In the above example code, when you change the ‘List’ object through the ‘append’ inplace syntax, the address of the object is not changed. ‘Tuple’ does not support this kind of inplace. With ‘JIT_SYNTAX_LEVEL’ set to ‘LAX’, static graph mode can support the inplace operation of some ‘List’ objects.

The specific usage scenarios are as follows:

• Support for getting the original ‘List’ object from a global variable

  In the following example, the static graph gets the ‘List’ object, performs the inplace operation ‘list.reverse()’ supported by graph mode on the original object, and returns the original object. It can be seen that the object returned by the graph mode has the same ID as the original global variable object, that is, the two are the same object. If ‘JIT_SYNTAX_LEVEL’ is set to the ‘STRICT’ option, the returned ‘List’ object and the global object are two different objects.

  import mindspore as ms
  
  global_list = [1, 2, 3, 4]
  
  @ms.jit
  def list_func():
      global_list.reverse()
      return global_list
  
  output = list_func()  # output: [4, 3, 2, 1]
  assert id(global_list) == id(output)
  

• Inplace operations on input ‘List’ objects are not supported

  When List’ is imported as a static graph, the ‘List’ object is copied once, and subsequent calculations are performed using the copied object, so it is not possible to perform an inplace operation on the original input object. For example:

  import mindspore as ms
  
  list_input = [1, 2, 3, 4]
  
  @ms.jit
  def list_func(x):
      x.reverse()
      return x
  
  output = list_func()  # output: [4, 3, 2, 1]  list_input: [1, 2, 3, 4]
  assert id(output) != id(list_input)
  

  As shown in the above use case, the ‘List’ object cannot be inplaced on the original object when input as a graph mode. The object returned by the graph mode is different from the object ID entered.

• Support for in-place modification of some ‘List’ built-in functions

  With ‘JIT_SYNTAX_LEVEL’ set to ‘LAX’, the graph mode section ‘List’ built-in function supports inplace. In cases where ‘JIT_SYNTAX_LEVEL’ is ‘STRICT’, none of the methods support the inplace operation.

  Currently, the built-in methods for ‘List’ in-place modification supported by graph mode are ‘extend’, ‘pop’, ‘reverse’, and ‘insert’. The built-in methods ‘append’, ‘clear’ and index assignment do not support in-place modification at the moment, and will be supported in subsequent versions.

  Examples are as follows:

  import mindspore as ms
  
  list_input = [1, 2, 3, 4]
  
  @ms.jit
  def list_func():
      list_input.reverse()
      return list_input
  
  output = list_func()  # output: [4, 3, 2, 1]  list_input: [4, 3, 2, 1]
  assert id(output) == id(list_input)
  

Supporting the High-Level Usage of Dictionary

• Support Top Graph Return Dictionary

  Examples are as follows:

  import mindspore as ms
  
  @ms.jit()
  def test_dict():
      x = {'a': 'a', 'b': 'b'}
      y = x.get('a')
      z = dict(y=y)
      return z
  
  out = test_dict()
  print("out:", out)
  

  The results are as follows:

  out:{'y': 'a'}
  

• Supporting Dictionary Index Value Retrieval and Assignment

  Examples are as follows:

  import mindspore as ms
  import numpy as np
  
  x = {"a": ms.Tensor(np.array([1, 2, 3])), "b": ms.Tensor(np.array([4, 5, 6])), "c": ms.Tensor(np.array([7, 8, 9]))}
  
  @ms.jit()
  def test_dict():
      y = x["b"]
      x["a"] = (2, 3, 4)
      return x, y
  
  out1, out2 = test_dict()
  print('out1:{}'.format(out1))
  print('out2:{}'.format(out2))
  

  The results are as follows:

  out1:{'a': (2, 3, 4), 'b': Tensor(shape=[3], dtype=Int64, value= [4, 5, 6]), 'c': Tensor(shape=[3], dtype=Int64, value= [7, 8, 9])}
  out2:[4 5 6]
  

Supporting the Usage of None

‘None’ is a special value in Python that represents null and can be assigned to any variable. Functions that do not have a return value statement are considered to return ‘None’. At the same time, ‘None’ is also supported as the input parameter or return value of the top graph or subgraph. Support ‘None’ as a subscript of a slice as input to ‘List’, ‘Tuple’, ‘Dictionary’.

Examples are as follows:

import mindspore as ms

@ms.jit
def test_return_none():
    return 1, "a", None

res = test_return_none()
print(res)

The results are as follows:

(1, 'a', None)

For functions with no return value, the ‘None’ object is returned by default.

import mindspore as ms

@ms.jit
def foo():
    x = 3
    print("x:", x)

res = foo()
assert res is None

As in the example below, ‘None’ is used as the default input parameter for the top graph.

import mindspore as ms

@ms.jit
def foo(x, y=None):
    if y is not None:
        print("y:", y)
    else:
        print("y is None")
    print("x:", x)
    return y

x = [1, 2]
res = foo(x)
assert res is None

Built-in Functions Support More Data Types

Extend the support for built-in functions. Python built-in functions perfectly support more input types, such as third-party library data types.

For example, in the following example, ‘x.asnumpy()’ and ‘np.ndarray’ are both types supported by extensions. More support for built-in functions can be found in the Python built-in functions section.

import numpy as np
import mindspore as ms
import mindspore.nn as nn

ms.set_context(mode=ms.GRAPH_MODE)

class Net(nn.Cell):
    def construct(self, x):
        return isinstance(x.asnumpy(), np.ndarray)

x = Tensor(np.array([-1, 2, 4]))
net = Net()
out = net(x)
assert out

Supporting Control Flow

In order to improve the support of Python standard syntax, realize dynamic and static unification, and extend the support for more data types in the use of control flow statements. Control flow statements refer to flow control statements such as ‘if’, ‘for’, and ‘while’. Theoretically, by extending the supported syntax, it is also supported in control flow scenarios. The code use cases are as follows:

import numpy as np
import mindspore as ms

@ms.jit
def func():
    x = np.array(1)
    if x <= 1:
        x += 1
    return ms.Tensor(x)

res = func()
print("res: ", res)

The results are as follows:

res: 2

Supporting Property Setting and Modification

The specific usage scenarios are as follows:

• Set and modify properties of custom class objects and third-party types

In graph mode, you can set and modify the properties of custom class objects, such as:

import mindspore import jit

class AssignClass():
    def __init__(self):
        self.x = 1

obj = AssignClass()

@jit
def foo():
    obj.x = 100
    return

foo()
print(f"obj.x is: {obj.x}")

The result is:

obj.x is: 100

In graph mode, you can set and modify the properties of third-party library objects, such as:

import mindspore import jit
import numpy as np

@jit
def foo():
    a = np.array([1, 2, 3, 4])
    a.shape = (2, 2)
    return a.shape

shape = foo()
print(f"shape is {shape}")

The result is:

shape is (2, 2)

• Make changes to the Cell’s self object, for example:

  import mindspore as ms
  from mindspore import nn, set_context
  set_context(mode=ms.GRAPH_MODE)
  
  class Net(nn.Cell):
    def __init__(self):
        super().__init__()
        self.m = 2
  
    def construct(self):
        self.m = 3
        return
  
  net = Net()
  net()
  print(f"net.m is {net.m}")
  

  The result is:

  net.m is 3
  

  Note that the self object only supports property modification, not property setting, that is, only supports modifying the properties set in the ‘init’ function. If no attribute is defined in ‘init’, it is not allowed to be set in graph mode. For example:

  import mindspore as ms
  from mindspore import nn, set_context
  set_context(mode=ms.GRAPH_MODE)
  
  class Net(nn.Cell):
    def __init__(self):
        super().__init__()
        self.m = 2
  
    def construct(self):
        self.m2 = 3 # self.m2 is not set in the __init__, so it cannot be set in graph mode
        return
  
  net = Net()
  net()
  

• Set and modify Cell objects and jit_class objects in the static graph

  Supporting modifying the properties of the graph mode Cell object, such as:

  import mindspore as ms
  from mindspore import nn, set_context
  set_context(mode=ms.GRAPH_MODE)
  
  class InnerNet(nn.Cell):
      def __init__(self):
          super(InnerNet, self).__init__()
          self.x = 10
  
  class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.inner = InnerNet()
  
    def construct(self):
        self.inner.x = 100
        return
  
  net = Net()
  net()
  print(f"net.inner.x is {net.inner.x}")
  

  The result is:

  net.inner.x is 100
  

  Supporting property modification of objects jit_class graph mode, such as:

  import mindspore as ms
  from mindspore import nn, set_context, jit_class
  set_context(mode=ms.GRAPH_MODE)
  
  @jit_class
  class InnerClass():
      def __init__(self):
          self.x = 10
  
  class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.inner = InnerClass()
  
    def construct(self):
        self.inner.x = 100
        return
  
  net = Net()
  net()
  print(f"net.inner.x is {net.inner.x}")
  

  The result is:

  net.inner.x is 100
  

  Note that if the same property is obtained before modifying the properties of the Cell/jit_class object in the graph mode, the obtained properties will be parsed as constants. This can cause problems when running the same network multiple times, such as:

  import mindspore as ms
  from mindspore import nn, set_context
  set_context(mode=ms.GRAPH_MODE)
  
  class InnerNet(nn.Cell):
      def __init__(self):
          self.x = 1
  
  class Net(nn.Cell):
    def __init__(self):
        super().__init__()
        self.inner = InnerNet()
  
    def construct(self):
        a = self.inner.x
        self.inner.x = a + 1
        return
  
  net = Net()
  value0 = net.inner.x
  net()
  value1 = net.inner.x
  net()
  value2 = net.inner.x
  print(f"value0 is {value0}")
  print(f"value1 is {value1}")
  print(f"value2 is {value2}")
  

  The result is:

  value0 is 1
  value1 is 2
  value2 is 2
  

  But in dynamic graph mode, the value of ‘value2’ should be 3. However, because ‘self.inner.x’ in the statement ‘a = self.inner.x’ is solidified as a constant 2, the value of ‘self.inner.x’ is set to 2 both times. This issue will be resolved in a subsequent release.

Supporting Derivation

The static graph syntax supported by the extension also supports its use in derivation, such as:

import mindspore as ms
from mindspore import ops

@ms.jit
def dict_net(a):
    x = {'a': a, 'b': 2}
    return a, (x, (1, 2))

out = ops.grad(dict_net)(ms.Tensor([1]))
assert out == 2

Annotation Type

For syntax supported by the runtime extensions, nodes are generated that cannot be derived by type, such as dynamically created Tensors, which are called ‘Any’ types. Because this type cannot be inferred correctly at compile time, the Anytype will be operated on with a default maximum precision of Float64 to prevent loss of precision. In order to better optimize performance, it is necessary to reduce the generation of Any type data. When the user can clearly know the specific type that will be generated by the extended syntax, we recommend using Annotation to specify the corresponding Python statement type, thereby determining the type of the interpretation node and avoiding the generation of Any type.

For example, the difference between the Tensor class and the tensor interface lies in the use of the Annotation Type mechanism within the tensor interface. When the dtype of the tensor function is determined, the function uses Annotation to specify the output type, thereby avoiding the generation of Any type. The use of Annotation Type only requires adding a comment #() -> tensor_type[float32] above or after the corresponding Python statement, where tensor_type[float32] after -> indicates the output type of the annotated statement.

The code example is as follows.

import mindspore as ms
import mindspore.nn as nn
from mindspore import ops, Tensor

class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.abs = ops.Abs()

    @ms.jit
    def construct(self, x, y):
        y1 = ms.tensor(x.asnumpy() + y.asnumpy(), dtype=ms.float32)
        y2 = ms.Tensor(x.asnumpy() + y.asnumpy(), dtype=ms.float32) # @jit.typing: () -> tensor_type[float32]
        y3 = Tensor(x.asnumpy() + y.asnumpy())
        y4 = Tensor(x.asnumpy() + y.asnumpy(), dtype=ms.float32)
        return self.abs(y1), self.abs(y2), self.abs(y3), self.abs(y4)

ms.set_context(mode=ms.GRAPH_MODE)
net = Net()
x = ms.Tensor(-1, dtype=ms.int32)
y = ms.Tensor(-1, dtype=ms.float32)
y1, y2, y3, y4 = net(x, y)

print(f"y1 value is {y1}, dtype is {y1.dtype}")
print(f"y2 value is {y2}, dtype is {y2.dtype}")
print(f"y3 value is {y3}, dtype is {y3.dtype}")
print(f"y4 value is {y4}, dtype is {y4.dtype}")

y1 value is 2.0, dtype is Float32
y2 value is 2.0, dtype is Float32
y3 value is 2.0, dtype is Float64
y4 value is 2.0, dtype is Float64

In the above example, you can see the difference related to creating ‘Tensor’. Due to the lack of Annotation indication in the Tensor class, y3 and y4 cannot infer the correct type and can only perform operations in the highest precision Float64. For y2, the corresponding type for JIT Fallback was specified through Annotation during Tensor creation, allowing it to perform operations according to the specified type. y1 created the Tensor using the tensor function interface and passed the dtype parameter as an Annotation indication, avoiding the generation of ‘Any’ type.

Syntax Constraints of Extended Syntaxes

When using the static graph extension support syntax, note the following points:

1. In order to match the support capability of the dynamic graph. That is, it must be within the scope of dynamic graph syntax, including but not limited to data types.

2. When extending the static graph syntax, more syntax is supported, but the execution performance may be affected and is not optimal.

3. When extending the static graph syntax, more syntax is supported, and the ability to import and export cannot be used with MindIR due to use Python.

4. It is not currently supported that the repeated definition of global variables with the same name across Python files, and these global variables are used in the network.