CN110874553A - Recognition model training method and device - Google Patents

Abstract

The embodiments of the present application provide a recognition model training method and device. The recognition model training method includes: acquiring sequence samples; inputting the sequence samples into a recognition model to obtain a first probability of a preset forward target sequence and a preset backward target sequence The second probability of The forward and backward target sequence is obtained, and the third probability of the forward and backward target sequence is calculated; the target function is calculated according to the first probability, the second probability and the third probability; and the recognition model is trained by using a preset training algorithm according to the target function. Through this solution, the real-time recognition of the recognition model can be realized.

Description

A recognition model training method and device

technical field

The present application relates to the technical field of machine learning, and in particular, to a recognition model training method and device.

Background technique

With the development of artificial intelligence technology, machine learning, as the core technology of artificial intelligence, has been widely used in object detection and tracking, behavior detection and recognition, speech recognition, etc. As an emerging field in machine learning research, DNN (Deep Neural Network) analyzes data by imitating the mechanism of the human brain, and is an intelligent model that analyzes and learns by establishing and simulating the human brain.

In traditional DNN, such as CNN (Convolutional Neural Network, Convolutional Neural Network), the network model establishes the mapping relationship between the input data and the output result, and the output result is obtained by inputting the input data into the network model. The output results obtained are independent of each other. However, in some special application scenarios, such as speech recognition, video target tracking and other scenarios, the data at each moment has a greater correlation with the data at other moments. RNN (Recurrent Neural Network, Recurrent Neural Network) is a DNN that implements cyclic sequence operations. The operation of RNN on each input data depends on the operation results of input data at other times.

When training the recognition model based on RNN, the method of forward calculation is mostly used. The process of forward calculation is to introduce the operation result of the past moment into the operation of the current moment. The model obtained by training tends to use as much future information as possible, so that the operation results at each moment are often delayed, resulting in the recognition model unable to meet the requirements of real-time recognition.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present application is to provide a recognition model training method and apparatus, so as to realize real-time recognition of the recognition model. The specific technical solutions are as follows:

In a first aspect, an embodiment of the present application provides a method for training a recognition model, the method comprising:

Get sequence samples;

Inputting the sequence samples into a recognition model to obtain a first probability of a preset forward target sequence and a second probability of a preset backward target sequence;

According to the preset forward target sequence and the preset backward target sequence, according to the same position, the target in the preset backward target sequence is first, and the target in the preset forward target sequence is behind. order, arrange to obtain the forward and backward target sequence, and calculate the third probability of the forward and backward target sequence;

calculating an objective function according to the first probability, the second probability and the third probability;

According to the objective function, the recognition model is trained by using a preset training algorithm.

Optionally, the recognition model includes a recurrent neural network and a connectionist time series classification algorithm;

The inputting the sequence samples into the recognition model to obtain the first probability of the preset forward target sequence and the second probability of the preset backward target sequence, including:

Inputting the sequence samples into the cyclic neural network, and through the forward calculation of the cyclic neural network, a first probability sequence composed of output probabilities of each feature in the sequence samples is obtained, and according to the first probability sequence, using the connectionist time series classification algorithm to calculate the first probability of the preset forward target sequence;

Through the backward calculation of the cyclic neural network, a second probability sequence composed of the output probabilities of each feature in the sequence sample is obtained, and according to the second probability sequence, the connectionist time series classification algorithm is used to calculate the preset The second probability of the backward target sequence.

Optionally, the calculating the third probability of the forward and backward target sequence includes:

According to the first probability sequence and the second probability sequence, calculate the mean value of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence to obtain a third probability sequence;

According to the third probability sequence, using the connectionist temporal classification algorithm, a third probability of the forward and backward target sequence is calculated.

Optionally, the calculating the objective function according to the first probability, the second probability and the third probability includes:

According to the first probability, the second probability and the third probability, use the objective function calculation formula to calculate the objective function, wherein the objective function calculation formula is:

g=-log(P _f )-log(P _b )-log(P _fb )

The g is the objective function, the P _f is the first probability, the P _b is the second probability, and the P _fb is the third probability.

Optionally, the preset training algorithm includes: a backpropagation algorithm;

According to the objective function, using a preset training algorithm to train the recognition model, including:

According to the objective function, determine the error between the predicted sequence obtained after the sequence sample is input into the recognition model and a preset target sequence, where the preset target sequence is the preset forward target sequence or the preset target sequence. the preset backward target sequence;

According to the error, the recognition model is trained by adjusting the parameters of the recognition model by using the back-propagation algorithm.

In a second aspect, an embodiment of the present application provides an apparatus for training a recognition model, the apparatus comprising:

The acquisition module is used to acquire sequence samples;

an identification module for inputting the sequence samples into an identification model to obtain a first probability of a preset forward target sequence and a second probability of a preset backward target sequence;

an arrangement module for, according to the preset forward target sequence and the preset backward target sequence, according to the same position, the target in the preset backward target sequence is in the front, and the preset forward target sequence is in the same position The order of the target in the rear, arrange to get the front and rear target sequence;

a calculation module, configured to calculate the third probability of the forward and backward target sequence; calculate the target function according to the first probability, the second probability and the third probability;

A training module is used to train the recognition model by using a preset training algorithm according to the objective function.

Optionally, the recognition model includes a recurrent neural network and a connectionist time series classification algorithm;

The identification module is specifically used for:

Inputting the sequence samples into the cyclic neural network, and through the forward calculation of the cyclic neural network, a first probability sequence composed of output probabilities of each feature in the sequence samples is obtained, and according to the first probability sequence, using the connectionist time series classification algorithm to calculate the first probability of the preset forward target sequence;

Through the backward calculation of the cyclic neural network, a second probability sequence composed of the output probabilities of each feature in the sequence sample is obtained, and according to the second probability sequence, the connectionist time series classification algorithm is used to calculate the preset The second probability of the backward target sequence.

Optionally, the computing module is specifically used for:

According to the first probability sequence and the second probability sequence, calculate the mean value of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence to obtain a third probability sequence;

According to the third probability sequence, using the connectionist temporal classification algorithm, a third probability of the forward and backward target sequence is calculated.

Optionally, the computing module is specifically used for:

According to the first probability, the second probability and the third probability, use the objective function calculation formula to calculate the objective function, wherein the objective function calculation formula is:

g=-log(P _f )-log(P _b )-log(P _fb )

The g is the objective function, the P _f is the first probability, the P _b is the second probability, and the P _fb is the third probability.

Optionally, the preset training algorithm includes: a backpropagation algorithm;

The training module is specifically used for:

According to the objective function, determine the error between the predicted sequence obtained after the sequence sample is input into the recognition model and a preset target sequence, where the preset target sequence is the preset forward target sequence or the preset target sequence. the preset backward target sequence;

According to the error, the recognition model is trained by adjusting the parameters of the recognition model by using the back-propagation algorithm.

In a recognition model training method and device provided by the embodiments of the present application, by acquiring a sequence sample and inputting the sequence sample into a recognition model, a first probability of a preset forward target sequence and a second probability of a preset backward target sequence are obtained , according to the preset forward target sequence and the preset backward target sequence, according to the order of the target in the preset backward target sequence at the same position first, and the target in the preset forward target sequence after the sequence, the front and rear targets are arranged to obtain sequence, and calculate the third probability of the forward and backward target sequence, calculate the target function according to the first probability, the second probability and the third probability, and use the preset training algorithm to train the recognition model according to the target function. By rearranging the preset forward target sequence and the preset backward target sequence, in the forward and backward target sequence, the target in the preset backward target sequence at each position is constrained to be in the front and the preset forward target sequence in the forward target sequence. The target is in the latter order. In this way, constraints on the decoding positions of the forward calculation and the backward calculation are added to the calculated target function, that is, for the target decoding of each position, the backward calculation is earlier than the forward calculation. The forward calculation will be delayed, and the backward calculation will be advanced. In this way, by restricting the decoding position, the result of the forward calculation in the trained recognition model is not delayed, and the result of the backward calculation is not advanced, and the real-time recognition of the recognition model is realized. .

Description of drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

Fig. 1 is the speech recognition block diagram based on RNN of the prior art;

2 is a schematic flowchart of a recognition model training method according to an embodiment of the present application;

3 is a schematic structural diagram of a recognition model training apparatus according to an embodiment of the present application;

FIG. 4 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

As shown in Figure 1, it is a block diagram of speech recognition based on RNN, and the input speech feature sequence is x=[x ₁ ,x ₂ ,x ₃ ,...,x _T ], RNN is a sequence learning network model, which can be LSTM ( Long Short Term Memory) network, GRU (Gated Recurrent Unit, gated recurrent unit) network, etc.

In the currently commonly used RNN, the input feature sequence is usually processed by bidirectional calculation, or the input feature sequence is processed by forward calculation. If two-way computation is used to process the input feature sequence, the output result at any time is related to the entire input feature sequence, which can only be used for offline recognition and cannot be used for real-time recognition; if forward computing is used to process the input feature sequence, the RNN network The model is also trained by forward computing. The final trained network model will have different delays in recognition, and the recognition model cannot meet the requirements of real-time recognition.

In order to realize real-time recognition of recognition models, embodiments of the present application provide a recognition model training method, apparatus, electronic device, and machine-readable storage medium.

Below, the recognition model training method provided by the embodiment of the present application is first introduced.

The execution subject of the recognition model training method provided by the embodiment of the present application may be an electronic device that executes an intelligent algorithm, and the electronic device may be an intelligent device with functions such as target detection and tracking, behavior detection and recognition, or speech recognition, for example For remote computers, remote servers, smart cameras, smart voice devices, etc., the executive body should at least include a processor equipped with a core processing chip. A manner of implementing the recognition model training method provided by the embodiments of the present application may be at least one manner of software, hardware circuits, and logic circuits provided in the execution body.

As shown in FIG. 2, a recognition model training method provided by the embodiment of the present application may include the following steps:

S201, obtain a sequence sample.

This embodiment can be applied to scenarios such as speech recognition and target tracking. Therefore, the sequence samples may be speech sequences, video frame sequences, text sequences, and the like.

S202, the sequence samples are input into the recognition model to obtain the first probability of the preset forward target sequence and the second probability of the preset backward target sequence.

The recognition model is a DNN model that realizes functions such as speech recognition and target recognition. In order to realize the cyclic operation of the sequence, the recognition model includes RNN and an algorithm unit for probability calculation. The algorithm for probability calculation can include CTC (Connectionism Temporal Classification, Connectionism temporal classification), HMM (Hidden MarkovModel, Hidden Markov Model), Attention (attention) mechanism, etc.

The operation process of RNN includes forward calculation and backward calculation. When the forward calculation is performed on the characteristics of a certain moment, the input of the operation needs to consider the operation state of each moment before the moment in addition to the characteristics of the moment; When the direction calculation is to perform an operation on the feature of a certain time, the input of the operation needs to consider the operation state at each time after the time in addition to the feature of the time. The preset forward target sequence is the target sequence expected to be obtained during the forward calculation, and the preset backward target sequence is the target sequence expected to be obtained during the backward calculation. Usually, the preset forward target sequence Same as preset backward target sequence. The first probability is the probability of obtaining the preset forward target sequence through forward calculation, and the second probability is the probability of obtaining the preset backward target sequence through backward calculation.

Optionally, the recognition model may include RNN and CTC algorithms.

RNN is a powerful sequence learning model, but it requires the input to be pre-segmented data, so the application is greatly limited. Combining with CTC can avoid the requirement of pre-segmenting the data. The basic idea is to interpret the network output of RNN as the probability distribution of all possible class sequences. Given this distribution, the objective function is to maximize the probability of the preset target sequence.

For the input of each time t of the RNN, the network output layer has L+1 (L is the number of category sets) nodes, where the output of the first L nodes is the probability of observing each category at time t, and the L+ The output of 1 node is the probability of observing a space. Adding spaces to the output enables CTC to handle the case where the adjacent target classes in the ground-truth preset target sequence are the same. By combining the values of the output layer at all times, the probability of any target sequence can be calculated. CTC comprehensively considers all alignment cases, so there is no need to pre-segment the data.

Correspondingly, S202 may specifically be:

Input the sequence sample into the RNN, and through the forward calculation of the RNN, obtain a first probability sequence composed of the output probabilities of each feature in the sequence sample, and use the CTC algorithm to calculate the first probability sequence of the preset forward target sequence according to the first probability sequence. probability;

Through the backward calculation of the RNN, a second probability sequence composed of output probabilities of each feature in the sequence sample is obtained, and according to the second probability sequence, the CTC algorithm is used to calculate the second probability of the preset backward target sequence.

Through the forward calculation of RNN, the first probability sequence [p _f1 ,p _f2 ,p _f3 ,...,p _fT ] can be obtained, where n∈[1,T] is the moment of inputting the feature sequence. The forward target sequence is [π _f1 ,π _f2 ,π _f3 ,…,π _fU ], and the first probability of the preset forward target sequence calculated by the CTC algorithm is P _f ; through the backward calculation of the RNN, the second probability can be obtained Sequence [p _b1 ,p _b2 ,p _b3 ,…,p _bT ], set the preset backward target sequence as [π _b1 ,π _b2 ,π _b3 ,…,π _bU ], and calculate the preset backward direction through the CTC algorithm The second probability of the target sequence is P _b .

S203, according to the preset forward target sequence and the preset backward target sequence, according to the order in which the target in the preset backward target sequence at the same position is at the front and the target in the preset forward target sequence is at the rear, arrange to obtain the forward and backward direction target sequence, and calculate the third probability of the forward and backward target sequence.

In the operation of the RNN, the forward calculation will be delayed and the backward calculation will be advanced. In order to prevent the forward calculation from being delayed and the backward calculation not to be advanced, this embodiment uses the decoding results of the forward and backward RNNs to constrain each other, that is, According to the preset forward target sequence and the preset backward target sequence, according to the order of the target in the preset backward target sequence at the same position first and the target in the preset forward target sequence at the back, the front and rear target sequence is obtained by arranging , as above, set the preset forward target sequence as [π _f1 ,π _f2 ,π _f3 ,…,π _fU ] and the preset backward target sequence as [π _b1 ,π _b2 ,π _b3 ,…,π _bU ], the obtained forward and backward target sequence is [π _b1 ,π _f1 ,π _b2 ,π _f2 ,π _b3 ,π _f3 ,…,π _bU ,π _fU ]. The third probability of the forward and backward target sequence can be calculated through algorithm mechanisms such as CTC, HMM, and Attention.

Optionally, the step of calculating the third probability of the forward and backward target sequence in S203 may specifically be:

According to the first probability sequence and the second probability sequence, calculate the mean value of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence to obtain a third probability sequence;

According to the third probability sequence, the CTC algorithm is used to calculate the third probability of the forward and backward target sequence.

Before using the CTC algorithm to calculate the third probability, it is necessary to calculate the third probability sequence that is mutually constrained in the forward and backward directions. Add the output probabilities of , and divide by two (that is, calculate the mean of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence), for example, the first probability sequence is [p _f1 ,p _f2 ,p _f3 ,…,p _fT ], the second probability sequence is [p _b1 ,p _b2 ,p _b3 ,…,p _bT ], then the third probability sequence is [(p _f1 +p _b1 )/2,(p _f2 + p _b2 )/2,(p _f3 +p _b3 )/2,…,(p _fT +p _bT )/2], using the CTC algorithm, calculate the third probability of the forward and backward target sequence as P _fb .

S204, calculate the objective function according to the first probability, the second probability and the third probability.

The objective function is the function based on the recognition model training, such as the gradient function in gradient training, etc., which represents the direction and degree of model parameter adjustment. In this embodiment, the first probability of the preset forward target sequence, The second probability of the backward target sequence and the third probability of the forward and backward target sequence are preset, and by combining these three probabilities, an objective function can be calculated, and the objective function more completely characterizes the direction and degree of model parameter adjustment. The objective function not only guarantees the recognition rate of forward calculation and backward calculation, but also restricts their decoding order.

Optionally, S204 may specifically be:

According to the first probability, the second probability and the third probability, use the objective function calculation formula to calculate the objective function, wherein the objective function calculation formula is:

g=-log(P _f )-log(P _b )-log(P _fb ) (1)

g is the objective function, P _f is the first probability, P _b is the second probability, and P _fb is the third probability.

For training algorithms such as back-propagation algorithm, the relationship between the objective function and the probability is logarithmic. You can calculate the inverse of the logarithm for the first probability, the second probability and the third probability, and then add the three results to remember. to the objective function.

S205, according to the objective function, use a preset training algorithm to train the recognition model.

The preset training algorithm may be a commonly used training algorithm such as a back-propagation algorithm and a gradient algorithm, which is not specifically limited here.

Optionally, the preset training algorithm may include: a back propagation algorithm.

Correspondingly, S205 may specifically be:

According to the objective function, determine the error between the predicted sequence obtained after the sequence samples are input into the recognition model and the preset target sequence, wherein the preset target sequence is a preset forward target sequence or a preset backward target sequence;

According to the error between the predicted sequence and the preset target sequence, the back-propagation algorithm is used to train the recognition model by adjusting the parameters of the recognition model.

The training process is to input the sequence samples into the recognition model to obtain the predicted sequence, calculate the error between the predicted sequence and the preset target sequence, and based on the error, use the back-propagation algorithm to continuously adjust the model parameters of the recognition model. The loop iterations train the recognition model. After training to obtain the final recognition model, when recognizing input such as speech sequences, video sequences, etc., the forward calculation of RNN can be directly used to obtain the recognition results in real time.

Applying this embodiment, by acquiring sequence samples and inputting the sequence samples into the recognition model, the first probability of the preset forward target sequence and the second probability of the preset backward target sequence are obtained. Set the backward target sequence, according to the order of the target in the preset backward target sequence at the same position in the front, and the target in the preset forward target sequence at the back, arrange to obtain the forward and backward target sequence, and calculate the number of the forward and backward target sequence. With three probabilities, an objective function is calculated according to the first probability, the second probability and the third probability, and a preset training algorithm is used to train the recognition model according to the objective function. By rearranging the preset forward target sequence and the preset backward target sequence, in the forward and backward target sequence, the target in the preset backward target sequence at each position is constrained to be in the front and the preset forward target sequence in the forward target sequence. The target is in the latter order. In this way, constraints on the decoding positions of the forward calculation and the backward calculation are added to the calculated target function, that is, for the target decoding of each position, the backward calculation is earlier than the forward calculation. The forward calculation will be delayed, and the backward calculation will be advanced. In this way, by restricting the decoding position, the result of the forward calculation in the trained recognition model is not delayed, and the result of the backward calculation is not advanced, and the real-time recognition of the recognition model is realized. .

Corresponding to the above method embodiments, the embodiments of the present application provide an apparatus for training an identification model. As shown in FIG. 3 , the apparatus for training an identification model may include:

an acquisition module 310, configured to acquire sequence samples;

The identification module 320 is configured to input the sequence samples into the identification model to obtain the first probability of the preset forward target sequence and the second probability of the preset backward target sequence;

The arrangement module 330 is configured to, according to the preset forward target sequence and the preset backward target sequence, according to the same position, the target in the preset backward target sequence is first, and the preset forward target sequence In the order of the target in the back, arrange to get the front and rear target sequence;

A calculation module 340, configured to calculate the third probability of the forward and backward target sequence; calculate the target function according to the first probability, the second probability and the third probability;

The training module 350 is configured to use a preset training algorithm to train the recognition model according to the objective function.

Optionally, the recognition model may include a recurrent neural network and a connectionist time series classification algorithm;

The identification module 320 can be specifically used for:

Inputting the sequence samples into the cyclic neural network, and through the forward calculation of the cyclic neural network, a first probability sequence composed of output probabilities of each feature in the sequence samples is obtained, and according to the first probability sequence, using the connectionist time series classification algorithm to calculate the first probability of the preset forward target sequence;

Through the backward calculation of the cyclic neural network, a second probability sequence composed of the output probabilities of each feature in the sequence sample is obtained, and according to the second probability sequence, the connectionist time series classification algorithm is used to calculate the preset The second probability of the backward target sequence.

Optionally, the computing module 340 can be specifically used for:

According to the first probability sequence and the second probability sequence, calculate the mean value of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence to obtain a third probability sequence;

According to the third probability sequence, using the connectionist temporal classification algorithm, a third probability of the forward and backward target sequence is calculated.

Optionally, the computing module 340 can be specifically used for:

According to the first probability, the second probability and the third probability, use the objective function calculation formula to calculate the objective function, wherein the objective function calculation formula is:

g=-log(P _f )-log(P _b )-log(P _fb )

The g is the objective function, the P _f is the first probability, the P _b is the second probability, and the P _fb is the third probability.

Optionally, the preset training algorithm may include: a backpropagation algorithm;

The training module 350 can be specifically used for:

According to the objective function, determine the error between the predicted sequence obtained after the sequence sample is input into the recognition model and a preset target sequence, where the preset target sequence is the preset forward target sequence or the preset target sequence. the preset backward target sequence;

According to the error, the recognition model is trained by adjusting the parameters of the recognition model by using the back-propagation algorithm.

Applying this embodiment, by acquiring sequence samples and inputting the sequence samples into the recognition model, the first probability of the preset forward target sequence and the second probability of the preset backward target sequence are obtained. Set the backward target sequence, according to the order of the target in the preset backward target sequence at the same position in the front, and the target in the preset forward target sequence at the back, arrange to obtain the forward and backward target sequence, and calculate the number of the forward and backward target sequence. With three probabilities, an objective function is calculated according to the first probability, the second probability and the third probability, and a preset training algorithm is used to train the recognition model according to the objective function. By rearranging the preset forward target sequence and the preset backward target sequence, in the forward and backward target sequence, the target in the preset backward target sequence at each position is constrained to be in the front and the preset forward target sequence in the forward target sequence. The target is in the latter order. In this way, constraints on the decoding positions of the forward calculation and the backward calculation are added to the calculated target function, that is, for the target decoding of each position, the backward calculation is earlier than the forward calculation. The forward calculation will be delayed, and the backward calculation will be advanced. In this way, by restricting the decoding position, the result of the forward calculation in the trained recognition model is not delayed, and the result of the backward calculation is not advanced, and the real-time recognition of the recognition model is realized. .

Corresponding to the foregoing method embodiments, the embodiments of the present application provide an electronic device. As shown in FIG. 4 , the electronic device includes a processor 401 and a memory 402 , wherein,

The memory 402 is used to store computer programs;

The processor 401 is configured to implement any step of the above recognition model training method when executing the computer program stored in the memory 402 .

The above-mentioned memory may include RAM (Random Access Memory, random access memory), and may also include NVM (Non-Volatile Memory, non-volatile memory), for example, at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

The above-mentioned processor may be a general-purpose processor, including a GPU (Graphics Processing Unit, graphics processor), a CPU (Central Processing Unit, central processing unit), an NP (Network Processor, network processor), etc.; it may also be a DSP (Digital Signal Processing Unit) Processor, digital signal processor), ASIC (ApplicationSpecific Integrated Circuit, application specific integrated circuit), FPGA (Field-Programmable Gate Array, field programmable gate array) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components .

In this embodiment, by reading the computer program stored in the memory, and running the computer program, the processor of the electronic device can achieve: by acquiring a sequence sample, inputting the sequence sample into the recognition model, and obtaining a preset forward target The first probability of the sequence and the second probability of the preset backward target sequence are based on the preset forward target sequence and the preset backward target sequence, and the target in the preset backward target sequence at the same position is in the front and the preset front Arrange the front and rear target sequences in the order in which the targets in the target sequence come after, and calculate the third probability of the front and rear target sequences, calculate the target function according to the first probability, the second probability and the third probability, and according to the target function , using the preset training algorithm to train the recognition model. By rearranging the preset forward target sequence and the preset backward target sequence, in the forward and backward target sequence, the target in the preset backward target sequence at each position is constrained to be in the front and the preset forward target sequence in the forward target sequence. The target is in the latter order. In this way, constraints on the decoding positions of the forward calculation and the backward calculation are added to the calculated target function, that is, for the target decoding of each position, the backward calculation is earlier than the forward calculation. The forward calculation will be delayed, and the backward calculation will be advanced. In this way, by restricting the decoding position, the result of the forward calculation in the trained recognition model is not delayed, and the result of the backward calculation is not advanced, and the real-time recognition of the recognition model is realized. .

In addition, corresponding to the recognition model training method provided by the above embodiments, the embodiments of the present application provide a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor Implement any step of the above recognition model training method.

In this embodiment, the computer readable storage medium executes the computer program of the recognition model training method provided by the embodiment of the present application when it is running, so it can realize: by acquiring a sequence sample, inputting the sequence sample into the recognition model, and obtaining the preset pre- The first probability of the forward target sequence and the second probability of the preset backward target sequence are based on the preset forward target sequence and the preset backward target sequence, and the target in the preset backward target sequence at the same position is in the front and the preset target sequence. Assume that the targets in the forward target sequence are in the latter order, arrange to obtain the forward and backward target sequence, and calculate the third probability of the forward and backward target sequence, and calculate the objective function according to the first probability, the second probability and the third probability. The objective function uses the preset training algorithm to train the recognition model. By rearranging the preset forward target sequence and the preset backward target sequence, in the forward and backward target sequence, the target in the preset backward target sequence at each position is constrained to be in the front and the preset forward target sequence in the forward target sequence. The target is in the latter order. In this way, constraints on the decoding positions of the forward calculation and the backward calculation are added to the calculated target function, that is, for the target decoding of each position, the backward calculation is earlier than the forward calculation. The forward calculation will be delayed, and the backward calculation will be advanced. In this way, by restricting the decoding position, the result of the forward calculation in the trained recognition model is not delayed, and the result of the backward calculation is not advanced, and the real-time recognition of the recognition model is realized. .

For the embodiments of the electronic device and the computer-readable storage medium, since the method contents involved are basically similar to the foregoing method embodiments, the description is relatively simple.

It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any such actual relationship or sequence exists. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

Each embodiment in this specification is described in a related manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the apparatus, electronic device, and computer-readable storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and reference may be made to some descriptions of the method embodiments for related parts.

The above descriptions are only preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application are included in the protection scope of this application.

Claims (10)

1. a recognition model training method, is characterized in that, described method comprises: Get sequence samples; Inputting the sequence samples into a recognition model to obtain a first probability of a preset forward target sequence and a second probability of a preset backward target sequence; According to the preset forward target sequence and the preset backward target sequence, according to the same position, the target in the preset backward target sequence is first, and the target in the preset forward target sequence is behind. order, arrange to obtain the forward and backward target sequence, and calculate the third probability of the forward and backward target sequence; calculating an objective function according to the first probability, the second probability and the third probability; According to the objective function, the recognition model is trained by using a preset training algorithm. 2. The method according to claim 1, wherein the recognition model comprises a recurrent neural network and a connectionism time series classification algorithm; The inputting the sequence samples into the recognition model to obtain the first probability of the preset forward target sequence and the second probability of the preset backward target sequence, including: Inputting the sequence samples into the cyclic neural network, and through the forward calculation of the cyclic neural network, a first probability sequence composed of output probabilities of each feature in the sequence samples is obtained, and according to the first probability sequence, using the connectionist time series classification algorithm to calculate the first probability of the preset forward target sequence; Through the backward calculation of the cyclic neural network, a second probability sequence composed of the output probabilities of each feature in the sequence sample is obtained, and according to the second probability sequence, the connectionist time series classification algorithm is used to calculate the preset The second probability of the backward target sequence. 3. The method according to claim 2, wherein the calculating the third probability of the forward and backward target sequence comprises: According to the first probability sequence and the second probability sequence, calculate the mean value of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence to obtain a third probability sequence; According to the third probability sequence, a third probability of the forward and backward target sequence is calculated using the connectionist temporal classification algorithm. 4. The method according to claim 1, wherein the calculating an objective function according to the first probability, the second probability and the third probability comprises: According to the first probability, the second probability and the third probability, use the objective function calculation formula to calculate the objective function, wherein the objective function calculation formula is: g=-log(P _f )-log(P _b )-log(P _fb ) The g is the objective function, the P _f is the first probability, the P _b is the second probability, and the P _fb is the third probability. 5. The method according to claim 1, wherein the preset training algorithm comprises: a back-propagation algorithm; According to the objective function, using a preset training algorithm to train the recognition model, including: According to the objective function, determine the error between the predicted sequence obtained after the sequence sample is input into the recognition model and a preset target sequence, where the preset target sequence is the preset forward target sequence or the preset target sequence. the preset backward target sequence; According to the error, the recognition model is trained by adjusting the parameters of the recognition model by using the back-propagation algorithm. 6. A recognition model training device, wherein the device comprises: The acquisition module is used to acquire sequence samples; an identification module for inputting the sequence samples into an identification model to obtain a first probability of a preset forward target sequence and a second probability of a preset backward target sequence; an arrangement module for, according to the preset forward target sequence and the preset backward target sequence, according to the same position, the target in the preset backward target sequence is in the front, and the preset forward target sequence is in the same position The order of the target in the rear, arrange to get the front and rear target sequence; a calculation module, configured to calculate the third probability of the forward and backward target sequence; calculate the target function according to the first probability, the second probability and the third probability; A training module is used to train the recognition model by using a preset training algorithm according to the objective function. 7. The device according to claim 6, wherein the recognition model comprises a recurrent neural network and a connectionist time series classification algorithm; The identification module is specifically used for: Inputting the sequence samples into the cyclic neural network, and through the forward calculation of the cyclic neural network, a first probability sequence composed of output probabilities of each feature in the sequence samples is obtained, and according to the first probability sequence, using the connectionist time series classification algorithm to calculate the first probability of the preset forward target sequence; Through the backward calculation of the cyclic neural network, a second probability sequence composed of the output probabilities of each feature in the sequence sample is obtained, and according to the second probability sequence, the connectionist time series classification algorithm is used to calculate the preset The second probability of the backward target sequence. 8. The device according to claim 7, wherein the computing module is specifically used for: According to the first probability sequence and the second probability sequence, calculate the mean value of each output probability in the first probability sequence and the output probability at the same time in the second probability sequence to obtain a third probability sequence; According to the third probability sequence, using the connectionist temporal classification algorithm, a third probability of the forward and backward target sequence is calculated. 9. The device according to claim 6, wherein the computing module is specifically used for: According to the first probability, the second probability and the third probability, use the objective function calculation formula to calculate the objective function, wherein the objective function calculation formula is: g=-log(P _f )-log(P _b )-log(P _fb ) The g is the objective function, the P _f is the first probability, the P _b is the second probability, and the P _fb is the third probability. 10. The apparatus according to claim 6, wherein the preset training algorithm comprises: a back-propagation algorithm; The training module is specifically used for: According to the objective function, determine the error between the predicted sequence obtained after the sequence sample is input into the recognition model and a preset target sequence, where the preset target sequence is the preset forward target sequence or the preset target sequence. the preset backward target sequence; According to the error, the recognition model is trained by adjusting the parameters of the recognition model by using the back-propagation algorithm.

Images