CN110188879B

CN110188879B - Operation method, device and related product

Info

Publication number: CN110188879B
Application number: CN201910471370.6A
Authority: CN
Inventors: 不公告发明人
Original assignee: Anhui Cambricon Information Technology Co Ltd
Current assignee: Anhui Cambricon Information Technology Co Ltd
Priority date: 2019-05-31
Filing date: 2019-05-31
Publication date: 2021-01-15
Anticipated expiration: 2039-05-31
Also published as: CN110188879A

Abstract

The disclosure relates to an operation method, an operation device and a related product. The product includes a chip, other kit components including but not limited to: a memory device, an interface apparatus and a control device; the storage device is connected with the chip through a bus and used for storing data; the interface device is electrically connected with the chip and is used for realizing data transmission between the chip and external equipment; the control device is electrically connected with the chip and used for monitoring the state of the chip. Products according to the present disclosure may reduce duplicate codes.

Description

Operation method, device and related product

Technical Field

The present disclosure relates to the field of information processing technologies, and in particular, to an operation method, an operation device, and a related product.

Background

In the technical field of artificial intelligence, a neural network algorithm is a very popular machine learning algorithm in recent years, and has a very good effect in various fields, such as image recognition, voice recognition, natural language processing and the like. Along with the development of neural network algorithms, the complexity of the algorithms is higher and higher, and in order to improve the recognition degree, the scale of the model is gradually increased.

Disclosure of Invention

In view of this, the present disclosure provides a method and an apparatus for creating an operation of a neural network node.

According to an aspect of the present disclosure, there is provided a method of creating an operation of a neural network node, the method comprising:

traversing nodes in a neural network, and determining a creating function corresponding to the operation of the nodes, wherein the creating function comprises a creating template function corresponding to the operation type to which the operation of the nodes belongs;

acquiring an input tensor of the operation based on the creation template function, analyzing parameters of the node, and creating an output tensor of the operation according to the operation, the input tensor and the parameters;

creating the operation from the input tensor and the output tensor of the operation and the parameters.

In one possible implementation, determining a creation function corresponding to the operation of the node includes:

and determining a creating function corresponding to the operation of the node according to the first corresponding relation, wherein the first corresponding relation records the association relation between the operation and the creating function corresponding to the operation.

In one possible implementation, obtaining an input tensor of the operation based on the creating template function, parsing parameters of the node, and creating a template function according to the operation, the input tensor, and the parameters, wherein the creating template function includes:

acquiring an input tensor of the operation according to the context information of the node;

analyzing the node according to the variable parameter of the created template function to obtain a parameter corresponding to the variable parameter; the parameters corresponding to the variable parameters can be changed according to different operations;

and creating an output tensor of the operation according to the operation, the input tensor of the operation and the parameters corresponding to the variable parameters.

In one possible implementation, the method further includes:

and establishing a second corresponding relation among the nodes, the created operation and the created output tensor, and storing the second corresponding relation as context information.

In one possible implementation, the method further includes:

in the process of operating the calculation graph corresponding to the neural network, aiming at each node in the calculation graph, determining the created operation corresponding to the node according to the second corresponding relation;

determining a calculation template function corresponding to the node according to the created operation;

and acquiring the input data of the created operation according to the context information of the node based on the calculation template function, and calculating according to the created operation and the input data to obtain the output data of the node.

In one possible implementation, the method further includes:

before traversing the nodes of the neural network according to the topological sequence, judging whether the nodes are preset nodes or not;

and if the node is not the preset node, determining the operation of the node.

In one possible implementation, the method further includes:

and creating a creating template function corresponding to the operation according to the operation type to which the operation belongs, wherein the creating template function comprises variable parameters, and the parameters corresponding to the variable parameters can be changed according to different operations.

In one possible implementation, the operation types include one or more of the following types: single input single output, single input multiple output, double input single output, multiple input single output.

According to another aspect of the present disclosure, there is provided an apparatus for creating an operation of a neural network node, the apparatus including:

the device comprises a first determining module, a second determining module and a third determining module, wherein the first determining module is used for traversing nodes in a neural network and determining a creating function corresponding to the operation of the nodes, and the creating function comprises a creating template function corresponding to the operation type to which the operation of the nodes belongs;

a first creating module, configured to obtain an input tensor of the operation based on the creating template function, analyze parameters of the node, and create an output tensor of the operation according to the operation, the input tensor, and the parameters;

a second creation module to create the operation from the input tensor and the output tensor of the operation and the parameters.

In one possible implementation manner, the first determining module includes:

the determining unit is used for determining a creating function corresponding to the operation of the node according to a first corresponding relation, wherein the first corresponding relation records the association relation between the operation and the creating function corresponding to the operation.

In one possible implementation manner, the first creating module includes:

the acquisition unit is used for acquiring the input tensor of the operation according to the context information of the node;

the analysis unit is used for analyzing the node according to the variable parameters of the created template function to obtain the parameters corresponding to the variable parameters; the parameters corresponding to the variable parameters can be changed according to different operations;

and the creating unit is used for creating an output tensor of the operation according to the operation, the input tensor of the operation and the parameters corresponding to the variable parameters.

In one possible implementation, the apparatus further includes:

and the establishing module is used for establishing a second corresponding relation among the nodes, the created operation and the created output tensor, and storing the second corresponding relation as context information.

In one possible implementation, the apparatus further includes:

a second determining module, configured to determine, for each node in the computational graph, a created operation corresponding to the node according to the second correspondence in a process of operating the computational graph corresponding to the neural network;

the third determining module is used for determining a calculation template function corresponding to the node according to the created operation;

and the computing module is used for acquiring the input data of the created operation according to the context information of the node based on the computing template function, and computing according to the created operation and the input data to obtain the output data of the node.

In one possible implementation, the apparatus further includes:

the judging module is used for judging whether the node is a preset node or not before traversing the nodes of the neural network according to the topological sequence;

a fourth determining module, configured to determine an operation of the node if the node is not a preset node.

In one possible implementation, the apparatus further includes:

and the third creating module is used for creating a creating template function corresponding to the operation according to the operation type to which the operation belongs, wherein the creating template function comprises variable parameters, and the parameters corresponding to the variable parameters can be changed according to different operations.

According to another aspect of the present disclosure, there is provided a creation apparatus of an operation of a neural network node, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to perform the above method.

According to another aspect of the present disclosure, there is provided a non-transitory computer readable storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement the above-described method.

By setting the creation template function corresponding to the operation type to which the operation belongs in the creation function of the operation, the same creation template function creation operation can be adopted for the operations belonging to the same operation type, and repeated codes can be reduced.

Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments, which proceeds with reference to the accompanying drawings.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.

Fig. 1 shows a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure.

Figure 2 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure.

Fig. 3 shows a flowchart of the method of step S12 according to an embodiment of the present disclosure.

Figure 4 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure.

Figure 5 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure.

Figure 6 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure.

Fig. 7 shows a block diagram of a creation apparatus of a neural network node operation according to an embodiment of the present disclosure.

Fig. 8 shows a block diagram of a creation apparatus of a neural network node operation according to an embodiment of the present disclosure.

Fig. 9 shows a block diagram of a creation apparatus of a neural network node operation according to an embodiment of the present disclosure.

Fig. 10 shows a block diagram of a creation apparatus of a neural network node operation according to an embodiment of the present disclosure.

Detailed Description

Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. In the drawings, like reference numbers can indicate functionally identical or similar elements. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

Furthermore, in the following detailed description, numerous specific details are set forth in order to provide a better understanding of the present disclosure. It will be understood by those skilled in the art that the present disclosure may be practiced without some of these specific details. In some instances, methods, means, elements and circuits that are well known to those skilled in the art have not been described in detail so as not to obscure the present disclosure.

Because the deep learning has a large amount of calculation and high requirement on the calculation speed, the actual application scene of the deep learning has higher requirements on the performance and the function of hardware for processing data, and the hardware with high calculation speed and low power consumption is very necessary. Neural network accelerators (artificial intelligence processors) are such hardware that have high computational speed and low power consumption.

In order to fully exert the performance of hardware, a programming interface of a hardware platform is provided for a user, and a high-performance neural network computing library is produced. The operation module of the high-performance neural network calculation library provides interfaces for basic operation and tensor calculation of the neural network and more perfect support for deep learning operation, and users can splice the neural network designed by themselves by using the basic operation.

The method has the advantages that the high-performance neural network calculation library is used for building the neural network, developers are required to clearly know the operation module, the data module and the equipment management module, the operational flow convention of the whole high-performance neural network calculation library is known, and when different neural networks are built, the code amount is large and the code repetition degree is high.

In order to solve the problem that a neural network is difficult to build based on a high-performance neural network computer library, a neural network execution framework is provided, so that a network structure can be quickly built by development and testing personnel of the high-performance neural network computer library without knowing details of the high-performance neural network computer library.

The function module in the high-performance neural network computer library is the place which is modified most after the whole system is built, and on one hand, the function module is the interface adaptation which is forced to be carried out by the interface modification of the high-performance neural network computer library; and on the other hand, the reconstruction is performed for multiple times in order to reduce the maintenance difficulty. Firstly, only a function of create _ op (create function) is used, a search from op to create function is realized by a series of if judgment, and then a computing interface of an operation provided by the high-performance neural network computing library is changed from a uniform parent interface into an interface different from each child interface, so that the computation of the operation also needs to be registered in a function module. When a neural network execution framework is built, registration of functions requires many repeated codes, and maintenance cost is increased.

In order to solve the above technical problem, the present disclosure provides a method for creating a neural network node operation. The method may be applied to a processor, which may be a general-purpose processor, for example, a central Processing unit (cpu), a graphics Processing unit (gpu), and the like. The processor may also be an artificial intelligence processor for performing artificial intelligence operations, which may include machine learning operations, brain-like operations, and the like. The machine learning operation comprises neural network operation, k-means operation, support vector machine operation and the like. The artificial intelligence processor may, for example, include one or a combination of an NPU (Neural-Network Processing Unit), a DSP (Digital Signal Processing), and a Field Programmable Gate Array (FPGA) chip. The artificial intelligence processor may include a plurality of arithmetic units, and the plurality of arithmetic units may perform operations in parallel.

Fig. 1 shows a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure. As shown in fig. 1, the method may include:

step S11, traversing nodes in a neural network, and determining a creating function corresponding to the operation of the nodes, wherein the creating function comprises a creating template function corresponding to the operation type to which the operation of the nodes belongs;

step S12, obtaining an input tensor of the operation based on the created template function, analyzing parameters of the node, and creating an output tensor of the operation according to the operation, the input tensor, and the parameters;

step S13, creating the operation according to the input tensor and the output tensor of the operation and the parameters.

And setting a creation template function corresponding to the operation type to which the operation belongs in the creation function of the operation, so that the same creation template function creation operation can be adopted for the operations belonging to the same operation type, and repeated codes can be reduced.

Specifically, the create function and the compute function of the operation provided by the high performance neural network computational library may be analyzed to classify the operation into different operation types, for example, the operation types may include: single input single output, single input multiple output, double input single output, multiple input single output, etc. The single-input single-output mode is that one input tensor is operated, one output tensor is obtained after the one input tensor is processed by operation, the single-input multi-output mode is that one input tensor is operated, a plurality of output tensors are obtained after the one input tensor is processed by operation, the double-input single-output mode is that two input tensors are operated, one output tensor is obtained after the two input tensors are processed by operation, and the multi-input single-output mode is that a plurality of input tensors are operated, and one output tensor is obtained after the plurality of input tensors are processed by operation.

The difference between different operations belonging to the same operation type may be that the additional parameters required for the operations are different, and the additional parameters (hereinafter, referred to as parameters) may refer to parameters other than the input tensor and the output tensor. For example, the creating function of the transpose operation is createtranship (input, output, perm), and other operations will exchange perm for their parameters, such as createSoftmax (input, output, dim). The same create template function may be provided for such different operations belonging to the same operation type.

In one possible implementation, the method of the present disclosure may further include: and creating a creating template function corresponding to the operation according to the operation type to which the operation belongs, wherein the creating template function can comprise variable parameters, and the parameters corresponding to the variable parameters can be changed according to different operations and can support the introduction of any parameter in a parameter variable mode, so that repeated codes in the function registering process are reduced, the maintenance difficulty is reduced, and the workload is reduced.

In one possible implementation manner, after the creation of the template function is completed, an association relationship between the operation type and the creation of the template function may be established. For example, as described above, a creation template function corresponding to the operation type of the operation may be set in the creation function of the operation to establish an association relationship between the operation type and the creation template function. Different function templates may have different identifications, where the identification may refer to the name, number, etc. of the created template function. Therefore, when a certain operation calls the creating function corresponding to the operation, the creating template function corresponding to the operation type of the operation can be called.

Table 1 shows an example of creating a template function according to an embodiment of the present disclosure

TABLE 1

The Context is Context information and is an abstraction of the connection relationship of the computation graph. Because each layer in the neural network is responsible for creating its own output tensor, the input tensor of each layer is created by its predecessor nodes, so the context information is needed to obtain the input tensor of each layer. A Node is an abstraction of a Node in a computation graph, a Node class may be created for each Node in the computation graph, information related to the Node is stored in the Node class, each Node includes all parameters describing computation (operation) of a current layer, for example, the parameters may include the above-mentioned additional parameters, attributes of the operation, and the like, and each Node may further include information of a predecessor Node and a successor Node of the Node, that is, context information, for example, the predecessor Node in _ Node and the successor Node out _ Node may be included.

Setting the creation template function corresponding to the operation type of the operation in the creation function of the operation may mean that each operation has a corresponding creation function, and the creation template function corresponding to the operation type of the operation is set in the creation function of the operation to obtain parameters of an input tensor, a creation output tensor, and an analysis node. For example, create _ op ()

create_simple_11p<OP>(Context*ctx,Node*node,Args&…param)

}

Assuming that the creation template function corresponding to the operation type of a certain operation op is create _ single _11p, the setting may be performed in the manner described above. It should be noted that the create _ op () may further include other instructions required for creating operation, and for simplicity, the present disclosure does not list specific contents.

Therefore, when the operation of creating the nodes of the neural network is needed, the nodes of the neural network can be traversed, the corresponding creating template function is called according to the operation type of each node, and the operation corresponding to the node is created according to the creating template function and the parameters of the node.

Specifically, for step S11, the node being traversed and for which the corresponding operation is to be created may be referred to as the currently traversed node. Figure 2 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure. As shown in fig. 2, the corresponding operations may be created for the nodes in the computation graph one by one according to the topological order of the computation graph, and thus, the currently traversed node may be the node which is being traversed and for which the corresponding operation is to be created.

In a possible implementation manner, the correspondence between the operation and the creation function corresponding to the operation may be saved by a map, and the creation function corresponding to the operation may be determined according to the operation and the correspondence.

As mentioned above, the create _ op function is a search from op to create function implemented by a series of if judgments, and then the computing interface of the operation provided by the high-performance neural network computing library is changed from a uniform parent interface to a different interface for each child interface. According to the method for establishing the function corresponding to the determining operation, the condition judging logic does not need to be searched, so that the condition judging logic for searching the establishing function does not need to be written, repeated codes in the function registering process are reduced, the maintenance difficulty is reduced, and the workload is reduced.

In one possible implementation, the map may be used to store the correspondence between the identifier of the operation and the identifier of the creation function corresponding to the operation. The identification of the operation may refer to information that can uniquely represent the operation, for example, the name, number, and the like of the operation; also, the identification of the creating function may refer to information that can uniquely represent the creating function, for example, the name, number, and the like of the creating function. For example, for a certain node, the operation of the node may be determined to be power (identification of the operation) according to the parameter op of the node, and the creation function corresponding to the operation may be determined to be create _ power (identification of the creation function) according to the identification "power" of the operation and the correspondence. The create function create _ power may call a create template function corresponding to the operation type of the operation power, for example, create _ simple _11p < > (Context ctx, Node, arms & … param):

create_power(){

create_simple_11p<>(Context*ctx,Node*node,Args&…param)

}

it should be noted that the correspondence between the operations and the creation functions corresponding to the operations may also be established in other manners, and the present disclosure is not limited to the manner in the above example.

For step S12, a template function creating implementation corresponding to the operation type of the operation of the currently traversed node may be called, and the specific process is as follows.

Fig. 3 shows a flowchart of the method of step S12 according to an embodiment of the present disclosure. As shown in fig. 3, in one possible implementation, step S12 may include:

step S121, obtaining an input tensor of the operation according to the context information of the node;

step S122, analyzing the node according to the variable parameter of the created template function to obtain a parameter corresponding to the variable parameter; the parameters corresponding to the variable parameters can be changed according to different operations;

and step S123, creating an output tensor of the operation according to the operation, the input tensor of the operation, and the parameters corresponding to the variable parameters.

As described above, the context information may be abstracted according to a connection relationship between nodes in the computation graph, the context information may be stored in the Node class, the context information may include a predecessor Node of the Node and a successor Node of the Node, and the context information may further include information such as an operation of the Node, an output tensor of the Node, and an address of an output tensor of the Node.

Referring to fig. 2, when the processor executes the creating method of the present disclosure, a predecessor node of a currently traversed node may be obtained according to context information of the currently traversed node, where an output tensor of the predecessor node is an input tensor of an operation. In a possible implementation manner, the parameter corresponding to the variable parameter may be an additional parameter of the operation, and thus, after the input tensor of the operation is obtained, the currently traversed node may be analyzed to obtain the additional parameter based on creating the template function.

An output tensor for the operation can be created from the operation, the input tensor for the operation, and the variable parameters based on creating the template function.

After the input tensor, the output tensor, and the additional parameters of the operation are obtained, the creation template function may be called according to step S13 to create the operation according to the input tensor, the output tensor, and the additional parameters of the operation. That is, the end point of the recursion is still a function of the creation of the high performance neural network computational library.

In one possible implementation, an identification of the operation may be passed to the create template function, which invokes an interface of the high performance neural network computational library creation operation according to the identification of the operation.

Through the process, the corresponding creating template function can be called according to the operation type of each node, and the operation corresponding to the node is created according to the creating template function and the parameter of the node, so that the same creating template function creating operation can be adopted for the operations belonging to the same operation type, and repeated codes can be reduced.

Figure 4 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure. As shown in fig. 4, in one possible implementation, the method may further include:

step S14, establishing a second correspondence between the node, the created operation, and the created output tensor, and saving the second correspondence as context information.

After the output tensor and the operation of the currently traversed node are created, a second corresponding relation between the created output tensor and the created operation and the node can be established and stored in the context information.

In one possible implementation, a Context class may be created to store the second correspondence.

After the second corresponding relationship is established, when the forward calculation and the reverse training of the neural network are carried out, the operation corresponding to the node can be obtained through the second corresponding relationship. In addition, the connection relationship between the nodes in the context information may include a forward connection relationship and a reverse connection relationship, the forward connection relationship may refer to a predecessor node and a successor node of the node during forward calculation, and the reverse connection relationship may refer to the predecessor node and the successor node of the node during reverse training. It will be appreciated that in the forward direction of computation, if a node is a predecessor of another node, then in the reverse direction of training, the node is a successor of the other node.

In constructing the computational graph, the processor may read the input nodes of each node and determine the predecessor and successor nodes of each node from the input nodes of each node. After determining the predecessor Node and successor Node of the Node, a Node class is created, which saves the predecessor Node and successor Node of the Node, for example, in _ Node and out _ Node, in addition to the information of the Node.

In one possible implementation, the Context class may be used as a friend class of the Node class, and the Context class provides an interface for accessing a predecessor Node and a successor Node of the Node externally, where the interface is used for reading Context information.

In one possible implementation, a first interface for a predecessor of a read node and a second interface for a successor of the read node may be provided. The first interface and the second interface provide opposite access in forward calculation and reverse training, and forward calculation and reverse training can be realized by using one calculation graph. Wherein, the access provided by the first interface and the second interface in forward calculation and reverse training can be the following in opposite directions: for the first interface: during forward calculation, when reading the context information through the first interface, returning in _ node of the node, and during reverse training, when reading the context information through the first interface, returning out _ node of the node; for the second interface, during forward calculation, when the context information is read through the first interface, the out _ node of the node is returned, and during reverse training, when the context information is read through the first interface, the in _ node of the node is returned.

By the method, only one computation graph is needed to realize the forward computation and the reverse training process of the neural network. In the related art, a forward function and a backward function are provided in a neural network for realizing forward computation and backward training of the neural network, respectively, and in order to perform the forward computation and the backward training, in order to enable the two functions to complete corresponding computations correctly, corresponding computation graphs need to be provided, respectively.

Figure 5 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure. As shown in fig. 5, in one possible implementation, the method may further include:

step S15, before traversing the nodes of the neural network according to the topological order, judging whether the nodes are preset nodes;

step S16, if the node is not a preset node, determining an operation of the node.

As shown in fig. 2, before all nodes of the neural network are traversed, corresponding operations are created for the nodes in the computational graph one by one. The preset node may refer to a node independent of other nodes, for example, the preset node may be a node Const (constant), place holder (Placeholder), or the like, where Const is a constant node, is a starting node in a computational graph, and is used for data incoming, and place holder is also a constant, and place holder may be understood as a type of form parameter, that is, it is not directly available like Const, and requires a user to transfer a constant value.

In a possible implementation manner, the currently traversed node may be compared with the preset nodes one by one to determine whether the currently traversed node is the preset node, or whether the currently traversed node is the preset node may also be determined according to the stored context information, and a specific manner is not limited in the present disclosure.

If the currently traversed node is a preset node, as shown in fig. 2, the input tensor of the node may be acquired, the node may be analyzed to acquire parameters required for creating the output tensor, then the output tensor is created, an operation is created according to the input tensor, the output tensor and the parameters, and a relationship among the node, the operation and the output tensor is established in the context information.

If the currently traversed node is not the preset node, the currently traversed node can be analyzed to determine the identifier of the operation of the node, and then the corresponding creating function can be determined according to the identifier of the operation, and the creating function creating operation is called.

According to the method and the device for creating the template function, different nodes can be classified and processed respectively, for the nodes which do not depend on other nodes, due to the fact that the commonalities among the operations of the nodes are not large, the corresponding creating template function does not need to be created, the corresponding creating function creating operation is called during creating, the different nodes are flexibly processed, and the creating operation time is saved.

Figure 6 illustrates a flow diagram of a method of creation of a neural network node operation in accordance with an embodiment of the present disclosure. As shown in fig. 6, in one possible implementation, the method may further include:

step S17, in the process of running the computational graph corresponding to the neural network, for each node in the computational graph, determining a created operation corresponding to the node according to the second correspondence;

step S18, determining a calculation template function corresponding to the node according to the created operation;

step S19, based on the computation template function, obtaining the input data of the created operation according to the context information of the node, and performing computation according to the created operation and the input data to obtain the output data of the node.

In one possible implementation, the method of the present disclosure further creates a computation template function, and the creating of the computation template function may also classify the operation into different operation types according to the foregoing, for example, the operation types may include: the method comprises the steps of single-input single-output, single-input multi-output, double-input single-output, multi-input single-output and the like, and corresponding calculation template functions are created for each operation type.

In a possible implementation manner, a created operation corresponding to the node may be determined according to the second correspondence, and then the calculation template function may be determined according to an operation type of the created operation. The input parameter of the computation template function may be an identifier of a node, and the identifier of the node may refer to information that can uniquely identify a certain node, and may be, for example, a name, a serial number, and the like of the node. The calculation template function may find the created operation corresponding to the node according to the identifier of the node and the second corresponding relationship, and may also obtain input data of the operation according to context information of the node. After the created operation and the input data corresponding to the node are determined, calculation can be performed according to the created operation and the input data to obtain output data of the node.

And a calculation template function corresponding to the operation is created according to the operation type, so that repeated codes in the calculation registration process of the operation can be reduced, and the maintenance cost is reduced.

The present disclosure also provides an apparatus for creating an operation of a neural network node, which may be applied to a processor, which may be a general-purpose processor, for example, a central Processing unit (cpu), a graphics Processing unit (gpu), and the like. The processor may also be an artificial intelligence processor for performing artificial intelligence operations, which may include machine learning operations, brain-like operations, and the like. The machine learning operation comprises neural network operation, k-means operation, support vector machine operation and the like. The artificial intelligence processor may, for example, include one or a combination of an NPU (Neural-Network Processing Unit), a DSP (Digital Signal Processing), and a Field Programmable Gate Array (FPGA) chip. The artificial intelligence processor may include a plurality of arithmetic units, and the plurality of arithmetic units may perform operations in parallel.

Fig. 7 shows a block diagram of a creation apparatus of a neural network node operation according to an embodiment of the present disclosure. As shown in fig. 7, the apparatus may include:

a first determining module 71, configured to traverse nodes in a neural network, and determine a creating function corresponding to an operation of the node, where the creating function includes a creating template function corresponding to an operation type to which the operation of the node belongs;

a first creating module 72, configured to obtain an input tensor of the operation based on the creating template function, analyze parameters of the node, and create an output tensor of the operation according to the operation, the input tensor, and the parameters;

a second creating module 73 for creating the operation from the input tensor and the output tensor of the operation and the parameters.

Fig. 8 shows a block diagram of a creation apparatus of a neural network node operation according to an embodiment of the present disclosure. As shown in fig. 8, in a possible implementation manner, the first determining module 71 includes:

the determining unit 711 is configured to determine, according to a first correspondence, a creating function corresponding to the operation of the node, where the first correspondence records an association between the operation and the creating function corresponding to the operation.

In one possible implementation, the first creating module 72 may include:

an obtaining unit 721 configured to obtain an input tensor of the operation according to the context information of the node;

the analyzing unit 722 is configured to analyze the node according to the variable parameter of the created template function to obtain a parameter corresponding to the variable parameter; the parameters corresponding to the variable parameters can be changed according to different operations;

a creating unit 723, configured to create an output tensor of the operation according to the operation, the input tensor of the operation, and the parameter corresponding to the variable parameter.

In one possible implementation, the apparatus further includes:

an establishing module 74, configured to establish a second correspondence between the node, the created operation, and the created output tensor, and store the second correspondence as context information.

In one possible implementation, the apparatus further includes:

a second determining module 75, configured to determine, for each node in the computational graph, an operation created corresponding to the node according to the second correspondence in the process of running the computational graph corresponding to the neural network;

a third determining module 76, configured to determine, according to the created operation, a calculation template function corresponding to the node;

a calculating module 77, configured to obtain, based on the computation template function, input data of the created operation according to the context information of the node, and perform calculation according to the created operation and the input data to obtain output data of the node.

In one possible implementation, the apparatus further includes:

a judging module 78, configured to judge whether a node of the neural network is a preset node before traversing the node according to the topological order;

a fourth determining module 79, configured to determine an operation of the node if the node is not a preset node.

In one possible implementation, the apparatus further includes:

a third creating module 70, configured to create, according to the operation type to which the operation belongs, a creating template function corresponding to the operation, where the creating template function includes variable parameters, and parameters corresponding to the variable parameters may be changed according to different operations.

Fig. 9 is a block diagram illustrating an apparatus 800 for creation of a neural network node operation, in accordance with an example embodiment. For example, the apparatus 800 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, an exercise device, a personal digital assistant, and the like.

Referring to fig. 9, the apparatus 800 may include one or more of the following components: processing component 802, memory 804, power component 806, multimedia component 808, audio component 810, input/output (I/O) interface 812, sensor component 814, and communication component 816.

The processing component 802 generally controls overall operation of the device 800, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations. The processing components 802 may include one or more processors 820 to execute instructions to perform all or a portion of the steps of the methods described above. Further, the processing component 802 can include one or more modules that facilitate interaction between the processing component 802 and other components. For example, the processing component 802 can include a multimedia module to facilitate interaction between the multimedia component 808 and the processing component 802.

The memory 804 is configured to store various types of data to support operations at the apparatus 800. Examples of such data include instructions for any application or method operating on device 800, contact data, phonebook data, messages, pictures, videos, and so forth. The memory 804 may be implemented by any type or combination of volatile or non-volatile memory devices such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disks.

Power components 806 provide power to the various components of device 800. The power components 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the apparatus 800.

The multimedia component 808 includes a screen that provides an output interface between the device 800 and a user. In some embodiments, the screen may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive an input signal from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 808 includes a front facing camera and/or a rear facing camera. The front camera and/or the rear camera may receive external multimedia data when the device 800 is in an operating mode, such as a shooting mode or a video mode. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

The audio component 810 is configured to output and/or input audio signals. For example, the audio component 810 includes a Microphone (MIC) configured to receive external audio signals when the apparatus 800 is in an operational mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signals may further be stored in the memory 804 or transmitted via the communication component 816. In some embodiments, audio component 810 also includes a speaker for outputting audio signals.

The I/O interface 812 provides an interface between the processing component 802 and peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to: a home button, a volume button, a start button, and a lock button.

The sensor assembly 814 includes one or more sensors for providing various aspects of state assessment for the device 800. For example, the sensor assembly 814 may detect the open/closed status of the device 800, the relative positioning of components, such as a display and keypad of the device 800, the sensor assembly 814 may also detect a change in the position of the device 800 or a component of the device 800, the presence or absence of user contact with the device 800, the orientation or acceleration/deceleration of the device 800, and a change in the temperature of the device 800. Sensor assembly 814 may include a proximity sensor configured to detect the presence of a nearby object without any physical contact. The sensor assembly 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 814 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

The communication component 816 is configured to facilitate communications between the apparatus 800 and other devices in a wired or wireless manner. The device 800 may access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 816 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 816 further includes a Near Field Communication (NFC) module to facilitate short-range communications. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, infrared data association (IrDA) technology, Ultra Wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

In an exemplary embodiment, the apparatus 800 may be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, micro-controllers, microprocessors or other electronic components for performing the above-described methods.

In an exemplary embodiment, a non-transitory computer-readable storage medium, such as the memory 804, is also provided that includes computer program instructions executable by the processor 820 of the device 800 to perform the above-described methods.

Figure 10 is a block diagram illustrating an apparatus 1900 for creation of a neural network node operation, according to an example embodiment. For example, the apparatus 1900 may be provided as a server. Referring to FIG. 10, the device 1900 includes a processing component 1922 further including one or more processors and memory resources, represented by memory 1932, for storing instructions, e.g., applications, executable by the processing component 1922. The application programs stored in memory 1932 may include one or more modules that each correspond to a set of instructions. Further, the processing component 1922 is configured to execute instructions to perform the above-described method.

The device 1900 may also include a power component 1926 configured to perform power management of the device 1900, a wired or wireless network interface 1950 configured to connect the device 1900 to a network, and an input/output (I/O) interface 1958. The device 1900 may operate based on an operating system stored in memory 1932, such as Windows Server, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, or the like.

In an exemplary embodiment, a non-transitory computer readable storage medium, such as the memory 1932, is also provided that includes computer program instructions executable by the processing component 1922 of the apparatus 1900 to perform the above-described methods.

The present disclosure may be systems, methods, and/or computer program products. The computer program product may include a computer-readable storage medium having computer-readable program instructions embodied thereon for causing a processor to implement various aspects of the present disclosure.

The computer readable storage medium may be a tangible device that can hold and store the instructions for use by the instruction execution device. The computer readable storage medium may be, for example, but not limited to, an electronic memory device, a magnetic memory device, an optical memory device, an electromagnetic memory device, a semiconductor memory device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a Static Random Access Memory (SRAM), a portable compact disc read-only memory (CD-ROM), a Digital Versatile Disc (DVD), a memory stick, a floppy disk, a mechanical coding device, such as punch cards or in-groove projection structures having instructions stored thereon, and any suitable combination of the foregoing. Computer-readable storage media as used herein is not to be construed as transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., optical pulses through a fiber optic cable), or electrical signals transmitted through electrical wires.

The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to a respective computing/processing device, or to an external computer or external storage device via a network, such as the internet, a local area network, a wide area network, and/or a wireless network. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers. The network adapter card or network interface in each computing/processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in the respective computing/processing device.

The computer program instructions for carrying out operations of the present disclosure may be assembler instructions, Instruction Set Architecture (ISA) instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider). In some embodiments, the electronic circuitry that can execute the computer-readable program instructions implements aspects of the present disclosure by utilizing the state information of the computer-readable program instructions to personalize the electronic circuitry, such as a programmable logic circuit, a Field Programmable Gate Array (FPGA), or a Programmable Logic Array (PLA).

Various aspects of the present disclosure are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer-readable program instructions.

These computer-readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer-readable program instructions may also be stored in a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer-readable medium storing the instructions comprises an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer, other programmable apparatus or other devices implement the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

Having described embodiments of the present disclosure, the foregoing description is intended to be exemplary, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terms used herein were chosen in order to best explain the principles of the embodiments, the practical application, or technical improvements to the techniques in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method of creating an operation for a neural network node, the method comprising:

traversing nodes in a neural network, and determining a creating function corresponding to the operation of the nodes according to the operation and the corresponding relation of the nodes, wherein the corresponding relation is the corresponding relation between the operation and the creating function corresponding to the operation, the creating function comprises a creating template function corresponding to an operation type to which the operation of the nodes belongs, and the operation type comprises one or more of the following types: the method comprises the following steps of establishing a template function, wherein the template function comprises variable parameters, and the parameters corresponding to the variable parameters are changed according to different operations;

acquiring an input tensor of the operation based on the creation template function, analyzing parameters of the node, and creating an output tensor of the operation according to the operation, the input tensor and the parameters; wherein obtaining an input tensor of the operation based on the creating template function, parsing parameters of the node, and creating a template function according to the operation, the input tensor, and the parameters, and includes: analyzing the node according to the variable parameter of the created template function to obtain a parameter corresponding to the variable parameter;

2. The method of claim 1, wherein determining a creation function to which the operation of the node corresponds comprises:

3. The method of claim 1, wherein obtaining an input tensor for the operation based on the creating a template function, parsing parameters of the node, and creating an output tensor for the operation from the operation, the input tensor, and the parameters creates a template function, further comprising:

and acquiring the input tensor of the operation according to the context information of the node.

4. The method of claim 3, further comprising:

5. The method of claim 4, further comprising:

6. The method according to claim 1 or 2, characterized in that the method further comprises:

and if the node is not the preset node, determining the operation of the node.

7. The method of claim 1, further comprising:

8. An apparatus for creating an operation of a neural network node, the apparatus comprising:

the first determining module is configured to traverse nodes in a neural network, and determine a creating function corresponding to an operation of the node according to the operation and a corresponding relationship of the node, where the corresponding relationship is a corresponding relationship between the operation and the creating function corresponding to the operation, the creating function includes a creating template function corresponding to an operation type to which the operation of the node belongs, and the operation type includes one or more of the following types: the method comprises the following steps of establishing a template function, wherein the template function comprises variable parameters, and the parameters corresponding to the variable parameters are changed according to different operations;

a first creating module, configured to obtain an input tensor of the operation based on the creating template function, analyze parameters of the node, and create an output tensor of the operation according to the operation, the input tensor, and the parameters; wherein the first creating module comprises: the analysis unit is used for analyzing the node according to the variable parameters of the created template function to obtain the parameters corresponding to the variable parameters; the parameters corresponding to the variable parameters can be changed according to different operations;

9. The apparatus of claim 8, wherein the first determining module comprises:

10. The apparatus of claim 8, wherein the first creating module further comprises:

and the acquisition unit is used for acquiring the input tensor of the operation according to the context information of the node.

11. The apparatus of claim 10, further comprising:

12. The apparatus of claim 11, further comprising:

13. The apparatus of claim 8 or 9, further comprising:

14. The apparatus of claim 8, further comprising:

15. An apparatus for creating an operation of a neural network node, comprising:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to carry out the method of any one of claims 1 to 7 when executing the instructions.

16. A non-transitory computer readable storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement the method of any of claims 1 to 7.