CN112115724A - Optimization method and system for fine adjustment of multi-domain neural network in vertical domain - Google Patents
Optimization method and system for fine adjustment of multi-domain neural network in vertical domain Download PDFInfo
- Publication number
- CN112115724A CN112115724A CN202010718332.9A CN202010718332A CN112115724A CN 112115724 A CN112115724 A CN 112115724A CN 202010718332 A CN202010718332 A CN 202010718332A CN 112115724 A CN112115724 A CN 112115724A
- Authority
- CN
- China
- Prior art keywords
- domain
- model
- basic
- vertical
- cutting
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/40—Processing or translation of natural language
- G06F40/58—Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Complex Calculations (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an optimization method and a system for fine tuning of a multi-field neural network in a vertical field, wherein the method comprises the following steps: training a basic domain model; cutting the basic domain model to obtain a cut basic domain model; generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model; and fine-tuning the cutting basic field model based on the vertical field data and the mask matrix to obtain the multi-field neural network model with the vertical field capability. According to the method, the useless branches in the basic network are trimmed and cut by using the vertical domain data and the mask matrix, and the effect of the vertical domain can be improved as an incremental learning mode, so that the finally obtained multi-domain model has all the performances of the basic domain model and the capability of processing the vertical domain, and the problem that the effect of the multi-domain neural network is sharply reduced in the basic domain after the multi-domain neural network is trimmed in the vertical domain can be solved.
Description
Technical Field
The invention relates to the technical field of machine translation, in particular to an optimization method and system for fine tuning of a multi-field neural network in a vertical field.
Background
At present, data volume of a vertical field is often small, direct training cannot be achieved, Fine-Tuning (Fine-Tuning) technology is mostly used for optimizing effects of the vertical field in the prior art, Fine-Tuning is performed on a basic field model by using vertical field data, for example, Fine-Tuning is performed on a spoken language translation model by using the textual field data in a machine translation task to optimize textual translation, and the method is one of the most effective methods for optimizing effects of the vertical field at present.
The method has the disadvantages that Fine-Tuning can effectively improve the translation effect of the vertical field, but has great influence on the effect of the pre-training model, which can cause the rapid reduction of the effect of the basic field, for example, in machine translation, the model of the spoken language field is used as the model of the basic field to optimize the translation task of the spoken language, and although the effect of the spoken language translation can be improved, the translation effect of the spoken language field is rapidly reduced; in addition, training in multiple domains simultaneously using a Multi-Task Learning (Multi-Task Learning) method can alleviate the rapid decrease in the effect of the basic domain, but cannot fundamentally solve the mutual influence between the multiple domains.
Disclosure of Invention
The invention provides an optimization method and system for fine tuning of a multi-field neural network in a vertical field, which are used for solving the problem that the effect of the multi-field neural network is sharply reduced in a basic field after the multi-field neural network is subjected to fine tuning in the vertical field.
The invention provides an optimization method for fine tuning of a multi-field neural network in a vertical field, which comprises the following steps:
step 1: training a basic domain model;
step 2: cutting the basic domain model to obtain a cut basic domain model;
and step 3: generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model;
and 4, step 4: and fine-tuning the cutting basic domain model based on the vertical domain data and the mask matrix to obtain a multi-domain neural network model with the vertical domain capability.
Further, the step 1: training the basic domain model includes:
and training the basic domain model to obtain the basic domain model which achieves the optimal effect in the basic domain.
Further, the step 2: cutting the basic domain model to obtain a cut basic domain model, wherein the step of cutting the basic domain model comprises the following steps:
and reserving useful branches in the basic domain model, and removing useless branches in the basic domain model to obtain the cutting basic domain model.
Further, the step 3: based on the basic cutting field model, generating a mask matrix corresponding to each layer in the basic cutting field model comprises the following steps:
and assigning the weight of the useful branch in the cutting basic domain model as 1, and assigning the weight of the useless branch in the cutting basic domain model as 0 to obtain a mask matrix corresponding to each layer in the cutting basic domain model.
Further, the step 4: fine-tuning the cutting basic domain model based on the vertical domain data and the mask matrix to obtain a multi-domain neural network model with the vertical domain capability, and executing the following steps:
step S41: fine tuning training is carried out on the cutting basic domain model by utilizing vertical domain data;
step S42: loading the mask matrix in the fine tuning training process;
step S43: and keeping the branch with the weight value of 1 fixed according to the mask matrix, and training the branch with the weight value of 0 to obtain the multi-domain neural network model with the vertical domain capability.
Further, the method further comprises the steps of:
and 5: and processing a language reasoning task according to the multi-domain neural network model with the vertical domain capability and the mask matrix.
Further, the step 5: according to the multi-domain neural network model with the vertical domain capability and the mask matrix, processing a language reasoning task and executing the following steps:
step S51: when an inference task in a basic field is processed, combining the multi-field neural network model with the vertical field capability with the mask matrix to obtain a basic field inference model, and performing inference by using the basic field inference model;
step S52: and when the inference task in the vertical field is processed, the multi-field neural network model with the vertical field capability is used for reasoning.
The optimization method for fine tuning of the multi-field neural network in the vertical field provided by the embodiment of the invention has the following beneficial effects: the vertical domain data and the mask matrix are used for trimming useless branches in the basic network, and as an incremental learning mode, the effect of the vertical domain can be improved, so that the finally obtained multi-domain model has all the performances of the basic domain model and has the capability of processing the vertical domain, and the problem that the effect of the multi-domain neural network is sharply reduced in the basic domain after the multi-domain neural network is trimmed in the vertical domain can be solved.
The invention also provides an optimization system for fine tuning of the multi-field neural network in the vertical field, which comprises the following steps:
the basic domain model training module is used for training a basic domain model;
the basic field model generating module is used for generating basic field models;
the mask matrix generating module is used for generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model;
and the trimming module of the cutting basic field model is used for trimming the cutting basic field model based on the vertical field data and the mask matrix to obtain the multi-field neural network model with the vertical field capability.
Further, the basic domain model training module is specifically configured to:
and training the basic domain model to obtain the basic domain model which achieves the optimal effect in the basic domain.
Further, the cutting basic domain model generation module is specifically configured to:
and reserving useful branches in the basic domain model, and removing useless branches in the basic domain model to obtain the cutting basic domain model.
The optimization system for fine tuning of the multi-field neural network in the vertical field provided by the embodiment of the invention has the following beneficial effects: the vertical domain data and the mask matrix are used for trimming useless branches in the basic network, and as an incremental learning mode, the effect of the vertical domain can be improved, so that the finally obtained multi-domain model has all the performances of the basic domain model and has the capability of processing the vertical domain, and the problem that the effect of the multi-domain neural network is sharply reduced in the basic domain after the multi-domain neural network is trimmed in the vertical domain can be solved.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a schematic flow chart illustrating an optimization method for fine tuning of a multi-domain neural network in a vertical domain according to an embodiment of the present invention;
fig. 2 is a block diagram of an optimization system for fine tuning of a multi-domain neural network in a vertical domain according to an embodiment of the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
The embodiment of the invention provides an optimization method for fine tuning of a multi-domain neural network in a vertical domain, as shown in fig. 1, the method comprises the following steps:
step 1: training a basic domain model;
step 2: cutting the basic domain model to obtain a cut basic domain model;
and step 3: generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model;
and 4, step 4: and fine-tuning the cutting basic domain model based on the vertical domain data and the mask matrix to obtain a multi-domain neural network model with the vertical domain capability.
The working principle of the technical scheme is as follows: because the network models have sparsity of different degrees, model cutting on the basic domain model does not affect the performance of the model; the trimmed model is finely adjusted, so that the performance of the basic domain model can be completely recovered during reasoning; since the effect of the basic domain can be reduced rapidly by directly fine-tuning the basic model by using the vertical domain data, the inventor finds that the useless branches in the basic domain can be useful for the vertical domain through research, so that the useless branches in the basic network can be fine-tuned and cut by using the vertical domain data and the mask matrix in the invention, and the effect of the vertical domain can be improved as an incremental learning mode, so that the finally obtained multi-domain model has all the performances of the basic domain model and the capability of processing the vertical domain.
Specifically, a basic domain model is trained first; cutting the basic domain model to obtain a cut basic domain model; generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model; and fine-tuning the cutting basic field model based on the vertical field data and the mask matrix to obtain the multi-field neural network model with the vertical field capability.
The beneficial effects of the above technical scheme are: the vertical domain data and the mask matrix are used for trimming useless branches in the basic network, and as an incremental learning mode, the effect of the vertical domain can be improved, so that the finally obtained multi-domain model has all the performances of the basic domain model and has the capability of processing the vertical domain, and the problem that the effect of the multi-domain neural network is sharply reduced in the basic domain after the multi-domain neural network is trimmed in the vertical domain can be solved.
In one embodiment, the step 1: training the basic domain model includes:
and training the basic domain model to obtain the basic domain model which achieves the optimal effect in the basic domain.
The working principle of the technical scheme is as follows: the base domain model M1 is trained to optimize in base domain effect.
The beneficial effects of the above technical scheme are: specific methods of training a base domain model are provided.
In one embodiment, the step 2: cutting the basic domain model to obtain a cut basic domain model, wherein the step of cutting the basic domain model comprises the following steps:
and reserving useful branches in the basic domain model, and removing useless branches in the basic domain model to obtain the cutting basic domain model.
The working principle of the technical scheme is as follows: model clipping is performed on the basic domain model M1, namely, useful branches in the M1 are reserved, and useless branches are removed, so that a clipped model M2 (called a clipping basic domain model) is obtained, the effect of the clipping basic domain model M2 is equivalent to that of the basic domain model M1, and in the subsequent steps, the clipping basic domain M2 is used as a new basic model to perform effect optimization of a vertical domain.
The beneficial effects of the above technical scheme are: and providing a specific method for cutting the basic field model to obtain a cut basic field model.
In one embodiment, the step 3: based on the basic cutting field model, generating a mask matrix corresponding to each layer in the basic cutting field model comprises the following steps:
and assigning the weight of the useful branch in the cutting basic domain model as 1, and assigning the weight of the useless branch in the cutting basic domain model as 0 to obtain a mask matrix corresponding to each layer in the cutting basic domain model.
The working principle of the technical scheme is as follows: and assigning the weight of the useful branch in the clipping basic domain model M2 to be 1, and assigning the weight of the useless branch in the clipping basic domain model M2 to be 0, so that a mask matrix corresponding to each layer in the clipping basic domain model can be obtained, wherein the mask matrix is a matrix consisting of 0 and 1, and the mask matrix is introduced, so that the performance of the basic model can be completely recovered during reasoning.
The beneficial effects of the above technical scheme are: a specific method for generating a mask matrix corresponding to each layer in the clipping basic domain model based on the clipping basic domain model is provided.
In one embodiment, the step 4: fine-tuning the cutting basic domain model based on the vertical domain data and the mask matrix to obtain a multi-domain neural network model with the vertical domain capability, and executing the following steps:
step S41: fine tuning training is carried out on the cutting basic domain model by utilizing vertical domain data;
step S42: loading the mask matrix in the fine tuning training process;
step S43: and keeping the branch with the weight value of 1 fixed according to the mask matrix, and training the branch with the weight value of 0 to obtain the multi-domain neural network model with the vertical domain capability.
The working principle of the technical scheme is as follows: firstly, performing Fine-Tuning on a cutting basic domain model M2 by using vertical domain data; then loading the mask matrix obtained in the step 3 in the course of Fine-Tuning training; and finally, according to the mask matrix, fixing the weight of the branch with the value of 1 in the clipping basic field model to be unchanged, training the weight of the branch with the value of 0 in the clipping basic field model to finally obtain a multi-field neural network model M3, wherein the multi-field neural network model M3 has the capability of processing the vertical field.
The beneficial effects of the above technical scheme are: the method comprises the specific steps of fine-tuning a cutting basic field model based on vertical field data and a mask matrix to obtain a multi-field neural network model with the vertical field capability.
In one embodiment, the method further comprises the steps of:
and 5: and processing a language reasoning task according to the multi-domain neural network model with the vertical domain capability and the mask matrix.
The working principle of the technical scheme is as follows: the step 5: according to the multi-domain neural network model with the vertical domain capability and the mask matrix, processing a language reasoning task and executing the following steps:
step S51: when an inference task in a basic field is processed, combining the multi-field neural network model with the vertical field capability with the mask matrix to obtain a basic field inference model, and performing inference by using the basic field inference model;
step S52: and when the inference task in the vertical field is processed, the multi-field neural network model with the vertical field capability is used for reasoning.
During reasoning, when the basic domain is processed, combining the multi-domain neural network model M3 with the mask matrix obtained in the step 3, wherein the obtained model (called basic domain reasoning model) is completely the same as the basic domain model M2, and the basic domain reasoning model has the capability of processing the basic domain; when processing the vertical domain, the multi-domain neural network model M3 model can be directly used, and the model has the capability of processing the vertical domain.
The beneficial effects of the above technical scheme are: the specific steps of processing the language reasoning task according to the multi-domain neural network model with the vertical domain capability and the mask matrix are provided.
As shown in fig. 2, an embodiment of the present invention provides an optimization system for fine tuning of a multi-domain neural network in a vertical domain, including:
a basic domain model training module 201 for training a basic domain model;
a cutting basic domain model generating module 202, configured to cut the basic domain model to obtain a cutting basic domain model;
a mask matrix generating module 203, configured to generate, based on the clipping basic domain model, a mask matrix corresponding to each layer in the clipping basic domain model;
and the cutting basic field model fine-tuning module 204 is used for fine-tuning the cutting basic field model based on the vertical field data and the mask matrix to obtain the multi-field neural network model with the vertical field capability.
The working principle of the technical scheme is as follows: because the network models have sparsity of different degrees, the model cutting performed on the basic domain model by the cutting basic domain model generating module 202 does not affect the performance of the model; the trimming module 204 for trimming the trimmed model of the basic domain can completely recover the performance of the basic domain model during reasoning; since the effect of the basic domain can be reduced rapidly by directly fine-tuning the basic model by using the vertical domain data, the inventor finds that the useless branches in the basic domain can be useful for the vertical domain through research, so that the trimming module 204 for trimming the basic domain model in the invention uses the vertical domain data and the mask matrix to fine-tune the useless branches in the trimming basic network as an incremental learning mode, the effect of the vertical domain can be improved, and the finally obtained multi-domain model has all the performances of the basic domain model and the capability of processing the vertical domain.
The beneficial effects of the above technical scheme are: the vertical domain data and the mask matrix are used for trimming useless branches in the basic network, and as an incremental learning mode, the effect of the vertical domain can be improved, so that the finally obtained multi-domain model has all the performances of the basic domain model and has the capability of processing the vertical domain, and the problem that the effect of the multi-domain neural network is sharply reduced in the basic domain after the multi-domain neural network is trimmed in the vertical domain can be solved.
In one embodiment, the basic domain model training module 201 is specifically configured to:
and training the basic domain model to obtain the basic domain model which achieves the optimal effect in the basic domain.
The working principle of the technical scheme is as follows: the base domain model training module 201 trains the base domain model M1 to achieve the optimum in the base domain effect.
The beneficial effects of the above technical scheme are: the base domain model may be trained by means of a base domain model training module 201.
In an embodiment, the cropping basis domain model generation module 202 is specifically configured to:
and reserving useful branches in the basic domain model, and removing useless branches in the basic domain model to obtain the cutting basic domain model.
The working principle of the technical scheme is as follows: the clipping basic domain model generation module 202 performs model clipping on the basic domain model M1, that is, useful branches in the M1 are reserved, and useless branches are removed, so as to obtain a clipped model M2 (referred to as a clipping basic domain model), the effect of the clipping basic domain model M2 is equivalent to that of the basic domain model M1, and the clipping basic domain model fine-tuning module 204 performs vertical domain effect optimization by using the clipping basic domain M2 as a new basic model.
The beneficial effects of the above technical scheme are: by means of the basic field model cutting generation module, cutting of the basic field model can be achieved, and the basic field model can be obtained.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.
Claims (10)
1. A method for optimizing fine tuning of a multi-domain neural network in a vertical domain is characterized by comprising the following steps:
step 1: training a basic domain model;
step 2: cutting the basic domain model to obtain a cut basic domain model;
and step 3: generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model;
and 4, step 4: and fine-tuning the cutting basic domain model based on the vertical domain data and the mask matrix to obtain a multi-domain neural network model with the vertical domain capability.
2. The method of claim 1, wherein the step 1: training the basic domain model includes:
and training the basic domain model to obtain the basic domain model which achieves the optimal effect in the basic domain.
3. The method of claim 1, wherein step 2: cutting the basic domain model to obtain a cut basic domain model, wherein the step of cutting the basic domain model comprises the following steps:
and reserving useful branches in the basic domain model, and removing useless branches in the basic domain model to obtain the cutting basic domain model.
4. The method of claim 3, wherein step 3: based on the basic cutting field model, generating a mask matrix corresponding to each layer in the basic cutting field model comprises the following steps:
and assigning the weight of the useful branch in the cutting basic domain model as 1, and assigning the weight of the useless branch in the cutting basic domain model as 0 to obtain a mask matrix corresponding to each layer in the cutting basic domain model.
5. The method of claim 4, wherein the step 4: fine-tuning the cutting basic domain model based on the vertical domain data and the mask matrix to obtain a multi-domain neural network model with the vertical domain capability, and executing the following steps:
step S41: fine tuning training is carried out on the cutting basic domain model by utilizing vertical domain data;
step S42: loading the mask matrix in the fine tuning training process;
step S43: and keeping the branch with the weight value of 1 fixed according to the mask matrix, and training the branch with the weight value of 0 to obtain the multi-domain neural network model with the vertical domain capability.
6. The method of claim 1, further comprising the steps of:
and 5: and processing a language reasoning task according to the multi-domain neural network model with the vertical domain capability and the mask matrix.
7. The method of claim 1, wherein the step 5: according to the multi-domain neural network model with the vertical domain capability and the mask matrix, processing a language reasoning task and executing the following steps:
step S51: when an inference task in a basic field is processed, combining the multi-field neural network model with the vertical field capability with the mask matrix to obtain a basic field inference model, and performing inference by using the basic field inference model;
step S52: and when the inference task in the vertical field is processed, the multi-field neural network model with the vertical field capability is used for reasoning.
8. A multi-domain neural network optimization system for vertical domain fine tuning, comprising:
the basic domain model training module is used for training a basic domain model;
the basic field model generating module is used for generating basic field models;
the mask matrix generating module is used for generating a mask matrix corresponding to each layer in the cutting basic domain model based on the cutting basic domain model;
and the trimming module of the cutting basic field model is used for trimming the cutting basic field model based on the vertical field data and the mask matrix to obtain the multi-field neural network model with the vertical field capability.
9. The system of claim 8, wherein the base domain model training module is specifically configured to:
and training the basic domain model to obtain the basic domain model which achieves the optimal effect in the basic domain.
10. The system of claim 8, wherein the clipping base domain model generation module is specifically configured to:
and reserving useful branches in the basic domain model, and removing useless branches in the basic domain model to obtain the cutting basic domain model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010718332.9A CN112115724B (en) | 2020-07-23 | 2020-07-23 | Optimization method and system for fine adjustment of multi-domain neural network in vertical domain |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010718332.9A CN112115724B (en) | 2020-07-23 | 2020-07-23 | Optimization method and system for fine adjustment of multi-domain neural network in vertical domain |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112115724A true CN112115724A (en) | 2020-12-22 |
| CN112115724B CN112115724B (en) | 2023-10-20 |
Family
ID=73798847
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010718332.9A Active CN112115724B (en) | 2020-07-23 | 2020-07-23 | Optimization method and system for fine adjustment of multi-domain neural network in vertical domain |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112115724B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112861549A (en) * | 2021-03-12 | 2021-05-28 | 云知声智能科技股份有限公司 | Method and equipment for training translation model |
| CN117057413A (en) * | 2023-09-27 | 2023-11-14 | 珠高智能科技(深圳)有限公司 | Reinforcement learning model fine tuning method, apparatus, computer device and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108985444A (en) * | 2018-07-12 | 2018-12-11 | 浙江工业大学 | A kind of convolutional neural networks pruning method inhibited based on node |
| WO2019127362A1 (en) * | 2017-12-29 | 2019-07-04 | 清华大学 | Neural network model block compression method, training method, computing device and system |
| CN111160519A (en) * | 2019-12-02 | 2020-05-15 | 上海交通大学 | Convolutional neural network model pruning method based on structure redundancy detection |
| CN111259846A (en) * | 2020-01-21 | 2020-06-09 | 第四范式（北京）技术有限公司 | Text positioning method and system and text positioning model training method and system |
| CN111291883A (en) * | 2018-12-07 | 2020-06-16 | 阿里巴巴集团控股有限公司 | Data processing method and data processing device |
-
2020
- 2020-07-23 CN CN202010718332.9A patent/CN112115724B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019127362A1 (en) * | 2017-12-29 | 2019-07-04 | 清华大学 | Neural network model block compression method, training method, computing device and system |
| CN108985444A (en) * | 2018-07-12 | 2018-12-11 | 浙江工业大学 | A kind of convolutional neural networks pruning method inhibited based on node |
| CN111291883A (en) * | 2018-12-07 | 2020-06-16 | 阿里巴巴集团控股有限公司 | Data processing method and data processing device |
| CN111160519A (en) * | 2019-12-02 | 2020-05-15 | 上海交通大学 | Convolutional neural network model pruning method based on structure redundancy detection |
| CN111259846A (en) * | 2020-01-21 | 2020-06-09 | 第四范式（北京）技术有限公司 | Text positioning method and system and text positioning model training method and system |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112861549A (en) * | 2021-03-12 | 2021-05-28 | 云知声智能科技股份有限公司 | Method and equipment for training translation model |
| CN112861549B (en) * | 2021-03-12 | 2023-10-20 | 云知声智能科技股份有限公司 | Method and equipment for training translation model |
| CN117057413A (en) * | 2023-09-27 | 2023-11-14 | 珠高智能科技(深圳)有限公司 | Reinforcement learning model fine tuning method, apparatus, computer device and storage medium |
| CN117057413B (en) * | 2023-09-27 | 2024-03-15 | 传申弘安智能(深圳)有限公司 | Reinforcement learning model fine tuning method, apparatus, computer device and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112115724B (en) | 2023-10-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10332507B2 (en) | Method and device for waking up via speech based on artificial intelligence | |
| CN107103903B (en) | Acoustic model training method and device based on artificial intelligence and storage medium | |
| US20190189112A1 (en) | Voice recognition processing method, device and computer storage medium | |
| CN112115724A (en) | Optimization method and system for fine adjustment of multi-domain neural network in vertical domain | |
| CN1750124A (en) | Bandwidth extension of band limited audio signals | |
| US11830521B2 (en) | Voice activity detection method and system based on joint deep neural network | |
| CN112800782A (en) | Text semantic feature fused voice translation method, system and equipment | |
| KR20210116923A (en) | Method for Training a Denoising Network, Method and Device for Operating Image Processor | |
| DE102019104304A1 (en) | Dynamic adaptation of speech comprehension systems to acoustic environments | |
| CN114663685B (en) | Pedestrian re-recognition model training method, device and equipment | |
| CN112258557B (en) | Visual tracking method based on space attention feature aggregation | |
| Kim et al. | Accelerating RNN transducer inference via adaptive expansion search | |
| CN114067819B (en) | Speech enhancement method based on cross-layer similarity knowledge distillation | |
| CN115984424A (en) | Electronic component defect image generation method and system based on diffusion model | |
| CN116959477B (en) | Convolutional neural network-based noise source classification method and device | |
| CN111833852B (en) | Acoustic model training method and device and computer readable storage medium | |
| CN110930997B (en) | Method for labeling audio by using deep learning model | |
| CN115527525B (en) | Speech recognition model generation method, speech interaction method, vehicle, and storage medium | |
| CN116071472A (en) | Image generation method and device, computer readable storage medium and terminal | |
| CN116152263A (en) | CM-MLP network-based medical image segmentation method | |
| CN114580659A (en) | BERT model fusion method based on model characteristic information enhancement | |
| CN115719310A (en) | Pretreatment method of fundus image data set and fundus image training model | |
| JP2022088341A (en) | Apparatus learning device and method | |
| CN115796242B (en) | Electronic digital information anti-evidence obtaining method | |
| CN107545583B (en) | Target tracking acceleration method and system based on Gaussian mixture model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |