WO2018223822A1 - 基于剪枝和蒸馏的卷积神经网络压缩方法 - Google Patents

基于剪枝和蒸馏的卷积神经网络压缩方法 Download PDF

Info

Publication number
WO2018223822A1
WO2018223822A1 PCT/CN2018/087063 CN2018087063W WO2018223822A1 WO 2018223822 A1 WO2018223822 A1 WO 2018223822A1 CN 2018087063 W CN2018087063 W CN 2018087063W WO 2018223822 A1 WO2018223822 A1 WO 2018223822A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
network
pruning
convolutional neural
neural network
Prior art date
Application number
PCT/CN2018/087063
Other languages
English (en)
French (fr)
Inventor
江帆
单羿
Original Assignee
北京深鉴智能科技有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 北京深鉴智能科技有限公司 filed Critical 北京深鉴智能科技有限公司
Publication of WO2018223822A1 publication Critical patent/WO2018223822A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/082Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology

Definitions

  • the present invention relates to convolutional neural networks, and more particularly to a convolutional neural network compression method based on pruning and distillation.
  • Convolutional Neural Network is widely used in various fields of computer vision, such as image recognition, object detection, image segmentation, etc. It is a research hotspot and key technology in the field of computer vision and multimedia. Research significance and practical value. Convolutional neural network is one of the representative network structures in deep learning technology. It has achieved great success in the field of image processing. Many successful models are based on convolutional neural networks on the international standard ImageNet dataset. Compared with the traditional image processing algorithms, the convolutional neural network avoids the complicated pre-processing process (extracting artificial features, etc.) and can directly input the original image, which is an end-to-end model.
  • CNN Convolutional Neural Network
  • the convolutional neural network has achieved good results, the convolutional neural network model has a large amount of parameters and many calculations.
  • the operation of the convolutional neural network is generally accelerated by the GPU, and the application of the convolutional neural network is also More in the cloud, facing huge challenges in the terminal landing.
  • FIG. 1 is a schematic illustration of pruning a convolutional neural network.
  • the pruning method can effectively reduce the parameters and calculation amount of the model, but since some elements and modules in the original model are deleted, the accuracy of the compressed model also has a certain degree of loss.
  • Another type of model compression method is called a distillation algorithm. See, for example, G. Hinton, O. Vinyals, J. Dean (2015) Distillering the knowledge in a neural network. ar Xiv preprint arXiv: 1503.02531, the entire disclosure of which is hereby incorporated by reference.
  • the distillation algorithm the "knowledge" of the complex model is refined or distilled to the small model, which guides the training of the small model, thereby also realizing the compression of the large model.
  • This kind of method can improve the performance of small models and reduce the computational complexity of large models, but the performance of small models after distillation is often not ideal.
  • the invention proposes a convolutional neural network compression method based on pruning and distillation, which can achieve better model precision while effectively compressing the model.
  • the accuracy of the model tends to decrease.
  • distillation the original model is used as a large model, and the pruned model is used as a small model.
  • the output of the small network fit the output of the large network to achieve the purpose of distillation.
  • the present invention combines two conventional network compression methods to compress the convolutional neural network model more effectively.
  • a convolutional neural network compression method comprising: pruning a raw convolutional neural network model to obtain a pruned model; fine-tuning a parameter of the pruned model;
  • the convolutional neural network model is used as the teacher network of the distillation algorithm.
  • the pruned model with fine-tuned parameters is used as the student network of the distillation algorithm.
  • the student network is used to guide the student network for training. Students who are trained by the distillation algorithm will be trained.
  • the network acts as a compressed convolutional neural network model.
  • the pruning operation may include retaining important connections or modules in the original convolutional neural network model, and deleting the remaining connections or modules.
  • the pruning criterion may be a magnitude of an absolute value of an element of the network model, the element having an absolute value greater than a threshold being retained, the element having an absolute value less than a threshold being set to zero.
  • the pruning criterion may be a size of a sum of absolute values of elements in a convolution kernel of the network model, the convolution kernel whose sum of absolute values is greater than a threshold is retained, and the sum of the absolute values is less than a threshold The convolution kernel is removed.
  • guiding the student network to perform training through the teacher network may include: adjusting the parameters of the student network by fitting the output of the teacher network with the output of the student network.
  • Figure 1 is a schematic illustration of pruning a convolutional neural network.
  • FIG. 2 is a schematic diagram of a distillation algorithm for a convolutional neural network.
  • FIG. 3 is a schematic diagram of a compression method in accordance with an embodiment of the present invention.
  • FIG. 4 is a flow chart of a convolutional neural network compression method in accordance with an embodiment of the present invention.
  • Figure 1 is a schematic illustration of pruning a convolutional neural network.
  • pruning we set a pruning criterion, and the elements or modules that satisfy the pruning criteria are deleted for each layer of the model.
  • the purpose of pruning is to preserve the more important connections or elements.
  • the pruning criterion is generally the absolute value of the element. Elements with a large absolute value are reserved, and elements with a small absolute value are set to 0.
  • Another method of pruning is to pruning the convolution kernel of the network, retaining the convolution kernel with the larger sum of the absolute values of the elements in the convolution kernel, and deleting the other convolution kernels.
  • the former pruning method can effectively reduce the amount of calculation and the size of the model.
  • the latter pruning method can compress the model size and accelerate the operation on specific hardware.
  • Distillation is a technique that refines "knowledge" in large networks into small networks.
  • 2 is a schematic diagram of a distillation algorithm for a convolutional neural network.
  • the teacher network is a large network model
  • the student network is a small network model
  • the teacher network guides the training of the student network.
  • the goal of training is to bring the output of the student network closer to the output of the teacher network. Since the output of the teacher network contains more information than the label of the picture, the distillation technique can better train the network than the traditional network training method.
  • the distillation we use the complete network as the teacher network, and the pruned network as the student model.
  • the loss function of the training is the cross entropy loss or the two norm loss of the output of the teacher network full connection layer and the output of the student network full connection layer. The convergence of distillation is often much faster than training from scratch.
  • FIG. 3 is a schematic diagram of a compression method in accordance with an embodiment of the present invention.
  • the convolutional neural network model (original model) is pruned, the more important connections or modules are retained, and the remaining connections or modules are deleted to obtain the pruned model.
  • the model size after pruning is compressed, but the performance is poor.
  • the fine-tuned pruning model obtained in the previous step is used as the student model in the distillation, and the original network model is used to guide the pruning network model training, and the parameters are Adjustments to further enhance the performance of the pruning model.
  • FIG. 4 is a flow chart of a convolutional neural network compression method in accordance with an embodiment of the present invention.
  • the convolutional neural network compression method 400 begins in step S401, in which a pruning operation is performed on the original convolutional neural network model to obtain a pruned model.
  • the pruning operation described in step S401 may include retaining important connections or modules in the original convolutional neural network model, and deleting the remaining connections or modules. More specifically, based on the pruning criteria, which connections or modules are reserved or deleted are determined.
  • the pruning criterion may be the magnitude of an absolute value of an element of the network model, the element having an absolute value greater than a threshold being retained, the element having an absolute value less than the threshold being set to zero.
  • the pruning criterion may be a magnitude of a sum of absolute values of elements in a convolution kernel of a network model, and a convolution kernel having a sum of absolute values greater than a threshold is retained, A convolution kernel whose sum of absolute values is less than the threshold is deleted.
  • step S403 parameter fine-tuning is performed on the pruned model.
  • step S405 the original convolutional neural network model is used as the teacher network of the distillation algorithm, and the pruned model after fine tuning of the parameters is used as the student network of the distillation algorithm, and the student network is used to guide the training through the teacher network according to the distillation algorithm.
  • the parameters of the student network are adjusted by fitting the output of the teacher network with the output of the student network.
  • step S407 the student network trained by the distillation algorithm is used as a compressed convolutional neural network model. Method 400 then ends.
  • the present invention proposes a compression method based on a neural network model of pruning and distillation.
  • the network model is first compressed by pruning, and then the distillation algorithm is used to adjust the parameters of the compression model to improve the performance of the compression model.
  • the present invention combines two conventional pruning methods, and the performance of the compressed model can be comparable to that of the original model.
  • the present invention is not a simple combination of the pruning method and the distillation method.
  • the accuracy of the model obtained by the traditional pruning method is often reduced, and additional measures are needed to improve the accuracy.
  • the pruned model is used as a small network to fit the large network.
  • distillation convergence is often much faster than de novo training. That is to say, by using the combination of the pruning method and the distillation method in the present invention, not only the scale of the network model is greatly compressed by the pruning method, but also the convergence speed of the distillation algorithm is improved, and the performance of the compressed network model can be improved.
  • the goal of compressing the convolutional neural network by those skilled in the art is fully achieved, and it is more advantageous to be applied in practice.

Abstract

一种基于剪枝和蒸馏的卷积神经网络压缩方法(400),包括:对原始卷积神经网络模型进行剪枝操作,得到剪枝后的模型(S401);对剪枝后的模型进行参数微调(S403);利用原始卷积神经网络模型作为蒸馏算法的老师网络,将经过参数微调的剪枝后的模型作为蒸馏算法的学生网络,根据蒸馏算法,通过老师网络来指导学生网络进行训练(S405);将经过蒸馏算法训练的学生网络作为压缩后的卷积神经网络模型(S407)。该方法通过将两个传统的网络压缩方法联合使用,更有效地压缩了卷积神经网络模型。

Description

基于剪枝和蒸馏的卷积神经网络压缩方法 技术领域
本发明涉及卷积神经网络,更具体涉及基于剪枝和蒸馏的卷积神经网络压缩方法。
背景技术
卷积神经网络(Convolutional Neural Network,CNN)目前被广泛地应用在计算机视觉的各个领域,如图像识别、物体检测、图像分割等等,是计算机视觉和多媒体领域的研究热点和关键技术,具有重要的研究意义和实用价值。卷积神经网络是深度学习技术中具有代表性的网络结构之一,在图像处理领域取得了很大的成功,在国际标准的ImageNet数据集上,许多成功的模型都是基于卷积神经网络。相较于传统的图像处理算法,卷积神经网络避免了对图像复杂的前期预处理过程(提取人工特征等),可以直接输入原始的图像,是一种端到端的模型。
卷积神经网络虽然取得了较好的效果,但由于卷积神经网络模型的参数量较大,运算次数较多,目前卷积神经网络的运算一般采用GPU进行加速,卷积神经网络的应用也多在云端,在终端落地面临巨大的挑战。
为了减少网络模型的冗余,科研人员提出对模型进行剪枝的压缩。例如,参见S.Han,J.Pool,J.Tran,W.J.DalIy(2015)Learning both Weights and Connections for Efficient Neural Network.In Advances in Neural Information Processing Systems,其公开的全部内容通过援引加入进来。另外,也可以参见和S.Han,H.Mao,W.J.Dally(2016)Deep compression:Compressing deep neural network with pruning,trained quantization and huffman coding.In International Conference on Learning Representations,其公开的全部内容通过援引加入进来。在剪枝中,通过一定的准则,保留比较重要的连接或元素,并将其余的连接或者元素删除,得到压缩的网络。然后,对压缩后的网络进 行参数微调。图1是对卷积神经网络进行剪枝的示意图。该剪枝方法能够有效地减小模型的参数和运算量,但由于删除了原始模型中的部分元素和模块,压缩后模型的准确率也有一定程度的损失。
另一类模型压缩方法被称为蒸馏算法。例如,参见G.Hinton,O.Vinyals,J.Dean(2015)Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531,其公开的全部内容通过援引加入进来。在蒸馏算法中,将复杂模型的“知识”提炼或者蒸馏到小模型,指导小模型的训练,从而也实现了对大模型的压缩。该类方法能够提升小模型的性能,减小大模型的运算量,但蒸馏后的小模型的性能往往不够理想。
考虑到针对卷积神经网络应用的现状,希望提出一种新型的卷积神经网络压缩方法。
发明内容
本发明提出了一种基于剪枝和蒸馏的卷积神经网络压缩方法,该方法在有效压缩模型的同时能达到较好的模型精度。对网络模型进行剪枝后,模型精度往往会有所降低,我们首先对该模型进行参数微调来恢复模型的精度。进一步地,我们采用蒸馏的方法将原始模型中的知识提炼到压缩模型中,提升压缩模型的性能。在蒸馏中,原始的模型作为大模型,剪枝后的模型作为小模型。训练时,我们使小网络的输出去拟合大网络的输出,从而达到蒸馏的目的。相对于剪枝方法和蒸馏方法,本发明将两个传统的网络压缩方法联合使用,更有效地压缩卷积神经网络模型。
具体地说,根据本发明,提供了卷积神经网络压缩方法,包括:对原始卷积神经网络模型进行剪枝操作,得到剪枝后的模型;对剪枝后的模型进行参数微调;利用原始卷积神经网络模型作为蒸馏算法的老师网络,将经过参数微调的剪枝后的模型作为蒸馏算法的学生网络,根据蒸馏算法,通过老师网络来指导学生网络进行训练;将经过蒸馏算法训练的学生网络作为压缩后的卷积神经网络模型。
优选地,所述的剪枝操作可以包括:保留原始卷积神经网络模型中重要的连接或模块,删除其余的连接或模块。
在根据本发明的方法中,优选地,基于剪枝准则来确定保留或删除哪些连接或模块。
优选地,所述剪枝准则可以是网络模型的元素的绝对值的大小,所述绝对值大于阈值的元素得以保留,所述绝对值小于阈值的元素被设置为0。
优选地,所述剪枝准则可以是网络模型的卷积核中的元素绝对值之和的大小,所述绝对值之和大于阈值的卷积核得以保留,所述绝对值之和小于阈值的卷积核被删除。
优选地,所述的根据蒸馏算法,通过老师网络来指导学生网络进行训练可以包括:通过用学生网络的输出去拟合老师网络的输出,对学生网络的参数进行调整。
附图说明
下面参考附图结合实施例说明本发明。在附图中:
图1是对卷积神经网络进行剪枝的示意图。
图2是对卷积神经网络进行蒸馏算法的示意图。
图3是根据本发明实施例的压缩方法的示意图。
图4是根据本发明实施例的卷积神经网络压缩方法的流程图。
具体实施方式
附图仅用于示例说明,不能理解为对本专利的限制;下面结合附图和实施例对本发明的技术方案做进一步的说明。
为了对卷积神经网络模型进行压缩,我们首先对网络进行剪枝操作。图1是对卷积神经网络进行剪枝的示意图。
在剪枝时,我们设定一个剪枝准则,对模型的每一层的参数,满足剪枝准则的元素或者模块被删除掉。剪枝的目的是保留比较重要的连接或者元素。
在对网络的元素剪枝时,剪枝准则一般是元素的绝对值的大小。绝对值大的元素被保留,绝对值小的元素被置为0。另一种剪枝方法是对网络的卷积核进行剪枝,将卷积核中的元素绝对值之和较大的卷积核保留,将其他的卷积核删除。
前一种剪枝方法可以有效减小运算量以及模型的尺寸,后一种剪枝方法可以压缩模型尺寸,在特定硬件上可以加速运算。
通过剪枝,我们得到了一个压缩后的小模型。由于原始模型的部分模块被删除,此时小模型的精度较差。我们对剪枝后的小模型进行参数微调。通常在参数微调后,模型的性能会得到较大的改善,但一般会低于原始网络模型的性能。在此基础上,我们通过蒸馏技术,进一步提升压缩模型的性能。
蒸馏是一种将大网络中的“知识”提炼到小网络中的技术。图2是对卷积神经网络进行蒸馏算法的示意图。蒸馏算法中有两个网络,即老师网络和学生网络。老师网络为大网络模型,学生网络为小的网络模型,老师网络指导学生网络的训练。训练的目标是使学生网络的输出逼近教师网络的输出。由于老师网络的输出相对于图片的标签包含更多的信息,因此蒸馏技术相对于传统的网络训练方法能够更好地训练网络。在蒸馏时,我们利用完整的网络作为教师网络,剪枝后的网络作为学生模型。训练的损失函数为教师网络全连接层的输出与学生网络全连接层的输出的交叉熵损失或者二范数损失。蒸馏的收敛往往比从头训练要快很多。
以上的过程可用图3来表示。图3是根据本发明实施例的压缩方法的示意图。
如图3中所示,首先,对卷积神经网络模型(原始模型)进行剪枝操作,保留比较重要的连接或者模块,删除其余的连接或者模块,得到剪枝后的模 型。剪枝后的模型尺寸得到压缩,但是性能较差。
接下来,对剪枝后的模型进行参数微调,提升剪枝模型的性能,得到微调后的剪枝模型。微调后的剪枝模型性能有了较大的提升,但还是会略低于完整的模型。
然后,利用原始的卷积神经网络作为蒸馏算法的老师网络,将上一步骤中得到的微调后的剪枝模型作为蒸馏中的学生模型,通过原始网络模型指导剪枝网络模型训练,对参数进行调整,进一步提升剪枝模型的性能。
最后,蒸馏训练后的学生模型即为最终的压缩模型。
为清楚起见,申请人绘制了该方法的流程图。图4是根据本发明实施例的卷积神经网络压缩方法的流程图。
如图4中所示,根据本发明的卷积神经网络压缩方法400开始于步骤S401,在此步骤,对原始卷积神经网络模型进行剪枝操作,得到剪枝后的模型。
根据本发明的优选实施例,步骤S401所述的剪枝操作可以包括:保留原始卷积神经网络模型中重要的连接或模块,删除其余的连接或模块。更具体地说,基于剪枝准则来确定保留或删除哪些连接或模块。
根据本发明的一个优选实施例,所述剪枝准则可以是网络模型的元素的绝对值的大小,所述绝对值大于阈值的元素得以保留,所述绝对值小于阈值的元素被设置为0。
根据本发明的另一优选实施例,所述剪枝准则可以是网络模型的卷积核中的元素绝对值之和的大小,所述绝对值之和大于阈值的卷积核得以保留,所述绝对值之和小于阈值的卷积核被删除。
接下来,在步骤S403,对剪枝后的模型进行参数微调。
然后,在步骤S405,利用原始卷积神经网络模型作为蒸馏算法的老师网络,将经过参数微调的剪枝后的模型作为蒸馏算法的学生网络,根据蒸馏算法,通过老师网络来指导学生网络进行训练。通过用学生网络的输出去拟合老师网络的输出,对学生网络的参数进行调整。
在步骤S407,将经过蒸馏算法训练的学生网络作为压缩后的卷积神经网络模型。然后,方法400结束。
综上所述,本发明提出了一种基于剪枝和蒸馏的神经网络模型的压缩方法。该方法中,首先利用剪枝压缩网络模型,然后利用蒸馏算法来对压缩模型的参数进行调整,提升压缩模型的性能。本发明将两种传统的剪枝方法联合使用,压缩后的模型性能可以与原始模型的性能相当。
本领域普通技术人员通过以上的详细描述应该理解,本发明并非剪枝法与蒸馏法的简单组合。一方面,传统的剪枝法得到的模型精度往往有所降低,需要通过额外的手段来提升精度;另一方面,蒸馏算法中,通过使用剪枝后模型作为小网络,去拟合大网络的输出,蒸馏的收敛往往比从头训练要快得多。也就是说,通过本发明这样将剪枝法和蒸馏法的联合使用,不仅通过剪枝法大大压缩了网络模型的规模,提高了蒸馏算法的收敛速度,而且使得压缩后的网络模型的性能可以与原始模型的性能相当,从而充分达到了本领域技术人员压缩卷积神经网络的目标,更有利于在实践中应用。
上面已经描述了本发明的各种实施例和实施情形。但是,本发明的精神和范围不限于此。本领域技术人员将能够根据本发明的教导而做出更多的应用,而这些应用都在本发明的范围之内。
也就是说,本发明的上述实施例仅仅是为清楚说明本发明所做的举例,而非对本发明实施方式的限定。对于所属领域的普通技术人员来说,在上述说明的基础上还可以做出其他不同形式的变化或变动。这里无需也无法对所 有的实施方式予以穷举。凡在本发明的精神和原则之内所作的任何修改、替换或改进等,均应包含在本发明权利要求的保护范围之内。

Claims (6)

  1. 一种卷积神经网络压缩方法,包括:
    对原始卷积神经网络模型进行剪枝操作,得到剪枝后的模型;
    对剪枝后的模型进行参数微调;
    利用原始卷积神经网络模型作为蒸馏算法的老师网络,将经过参数微调的剪枝后的模型作为蒸馏算法的学生网络,根据蒸馏算法,通过老师网络来指导学生网络进行训练;
    将经过蒸馏算法训练的学生网络作为压缩后的卷积神经网络模型。
  2. 根据权利要求1所述的方法,其中,所述的剪枝操作包括:保留原始卷积神经网络模型中重要的连接或模块,删除其余的连接或模块。
  3. 根据权利要求2所述的方法,其中,基于剪枝准则来确定保留或删除哪些连接或模块。
  4. 根据权利要求3所述的方法,其中,所述剪枝准则是网络模型的元素的绝对值的大小,所述绝对值大于阈值的元素得以保留,所述绝对值小于阈值的元素被设置为0。
  5. 根据权利要求3所述的方法,其中,所述剪枝准则是网络模型的卷积核中的元素绝对值之和的大小,所述绝对值之和大于阈值的卷积核得以保留,所述绝对值之和小于阈值的卷积核被删除。
  6. 根据权利要求1所述的方法,其中,所述的根据蒸馏算法,通过老师网络来指导学生网络进行训练包括:通过用学生网络的输出去拟合老师网络的输出,对学生网络的参数进行调整。
PCT/CN2018/087063 2017-06-07 2018-05-16 基于剪枝和蒸馏的卷积神经网络压缩方法 WO2018223822A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710421225.8A CN108334934B (zh) 2017-06-07 2017-06-07 基于剪枝和蒸馏的卷积神经网络压缩方法
CN201710421225.8 2017-06-07

Publications (1)

Publication Number Publication Date
WO2018223822A1 true WO2018223822A1 (zh) 2018-12-13

Family

ID=62923049

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/087063 WO2018223822A1 (zh) 2017-06-07 2018-05-16 基于剪枝和蒸馏的卷积神经网络压缩方法

Country Status (2)

Country Link
CN (1) CN108334934B (zh)
WO (1) WO2018223822A1 (zh)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111695483A (zh) * 2020-06-05 2020-09-22 腾讯科技(深圳)有限公司 一种车辆违章检测方法、装置、设备及计算机存储介质
CN111695375A (zh) * 2019-03-13 2020-09-22 上海云从企业发展有限公司 基于模型蒸馏的人脸识别模型压缩算法、介质及终端
CN112215353A (zh) * 2020-09-29 2021-01-12 电子科技大学 一种基于变分结构优化网络的通道剪枝方法
CN112487899A (zh) * 2020-11-19 2021-03-12 武汉高德飞行器科技有限公司 基于无人机的目标识别方法、系统、存储介质及电子设备
WO2021068180A1 (en) * 2019-10-11 2021-04-15 Beijing Didi Infinity Technology And Development Co., Ltd. Method and system for continual meta-learning
RU2749970C1 (ru) * 2019-10-24 2021-06-21 Бейдзин Сяоми Интиллиджент Текнолоджи Ко., ЛТД. Способ сжатия модели нейронной сети, а также способ и устройство для перевода языкового корпуса
US11403523B2 (en) 2019-06-07 2022-08-02 Tata Consultancy Services Limited Sparsity constraints and knowledge distillation based learning of sparser and compressed neural networks
CN114881227A (zh) * 2022-05-13 2022-08-09 北京百度网讯科技有限公司 模型压缩方法、图像处理方法、装置和电子设备
CN115496207A (zh) * 2022-11-08 2022-12-20 荣耀终端有限公司 神经网络模型压缩方法、装置和系统
US11775841B2 (en) 2020-06-15 2023-10-03 Cognizant Technology Solutions U.S. Corporation Process and system including explainable prescriptions through surrogate-assisted evolution
US11783195B2 (en) 2019-03-27 2023-10-10 Cognizant Technology Solutions U.S. Corporation Process and system including an optimization engine with evolutionary surrogate-assisted prescriptions
CN117011617A (zh) * 2023-10-07 2023-11-07 之江实验室 基于二阶段教师-学生框架的肺结节检测装置及构建方法

Families Citing this family (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109087303B (zh) * 2018-08-15 2022-04-01 中山大学 基于迁移学习提升语义分割模型效果的框架
CN109190521B (zh) * 2018-08-17 2022-03-04 北京亮亮视野科技有限公司 一种基于知识提纯的人脸识别模型的构建方法及应用
CN110929836B (zh) * 2018-09-20 2023-10-31 北京市商汤科技开发有限公司 神经网络训练及图像处理方法和装置、电子设备、介质
CN110929839B (zh) * 2018-09-20 2024-04-16 深圳市商汤科技有限公司 训练神经网络的方法和装置、电子设备和计算机存储介质
CN109409500B (zh) * 2018-09-21 2024-01-12 清华大学 基于知识蒸馏与非参数卷积的模型加速方法及装置
CN110163236B (zh) * 2018-10-15 2023-08-29 腾讯科技(深圳)有限公司 模型的训练方法和装置、存储介质、电子装置
CN109472360B (zh) * 2018-10-30 2020-09-04 北京地平线机器人技术研发有限公司 神经网络的更新方法、更新装置和电子设备
CN109522949B (zh) * 2018-11-07 2021-01-26 北京交通大学 目标识别模型建立方法及装置
CN111178115B (zh) * 2018-11-12 2024-01-12 北京深醒科技有限公司 对象识别网络的训练方法及系统
CN109598340A (zh) * 2018-11-15 2019-04-09 北京知道创宇信息技术有限公司 卷积神经网络的裁剪方法、装置及存储介质
CN109816636B (zh) * 2018-12-28 2020-11-27 汕头大学 一种基于智能终端的裂缝检测方法
CN110070183B (zh) * 2019-03-11 2021-08-20 中国科学院信息工程研究所 一种弱标注数据的神经网络模型训练方法及装置
CN110097084B (zh) * 2019-04-03 2021-08-31 浙江大学 通过投射特征训练多任务学生网络的知识融合方法
CN110059740A (zh) * 2019-04-12 2019-07-26 杭州电子科技大学 一种针对嵌入式移动端的深度学习语义分割模型压缩方法
CN110119811B (zh) * 2019-05-15 2021-07-27 电科瑞达(成都)科技有限公司 一种基于熵重要性准则模型的卷积核裁剪方法
CN110097177B (zh) * 2019-05-15 2022-11-29 电科瑞达(成都)科技有限公司 一种基于伪孪生网络的网络剪枝方法
US20230119593A1 (en) * 2019-06-21 2023-04-20 One Connect Smart Technology Co., Ltd. Method and apparatus for training facial feature extraction model, method and apparatus for extracting facial features, device, and storage medium
CN110348572B (zh) * 2019-07-09 2022-09-30 上海商汤智能科技有限公司 神经网络模型的处理方法及装置、电子设备、存储介质
CN110555417A (zh) * 2019-09-06 2019-12-10 福建中科亚创动漫科技股份有限公司 一种基于深度学习的视频图像识别系统及方法
CN111312271B (zh) * 2020-02-28 2023-03-21 云知声智能科技股份有限公司 一种提高收敛速度和处理性能的模型压缩方法和系统
CN111429415B (zh) * 2020-03-18 2020-12-08 东华大学 基于网络协同剪枝的产品表面缺陷高效检测模型构建方法
CN111553169B (zh) * 2020-06-25 2023-08-25 北京百度网讯科技有限公司 语义理解模型的剪枝方法、装置、电子设备和存储介质
CN112070233B (zh) * 2020-08-25 2024-03-22 北京百度网讯科技有限公司 模型联合训练方法、装置、电子设备和存储介质
WO2022198606A1 (zh) * 2021-03-26 2022-09-29 深圳市大疆创新科技有限公司 深度学习模型的获取方法、系统、装置及存储介质
CN113487028A (zh) * 2021-07-09 2021-10-08 平安科技(深圳)有限公司 知识蒸馏方法、装置、终端设备及介质
CN113837308B (zh) * 2021-09-29 2022-08-05 北京百度网讯科技有限公司 基于知识蒸馏的模型训练方法、装置、电子设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010042256A2 (en) * 2008-05-30 2010-04-15 The University Of Memphis Research Foundation Methods of improved learning in simultaneous recurrent neural networks
CN105894847A (zh) * 2016-06-27 2016-08-24 华南理工大学 一种云平台环境下无监督学习的实时公交动态调度系统及方法
CN106355248A (zh) * 2016-08-26 2017-01-25 深圳先进技术研究院 一种深度卷积神经网络训练方法及装置
CN106779068A (zh) * 2016-12-05 2017-05-31 北京深鉴智能科技有限公司 调整人工神经网络的方法和装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9916538B2 (en) * 2012-09-15 2018-03-13 Z Advanced Computing, Inc. Method and system for feature detection
US20160189026A1 (en) * 2014-12-26 2016-06-30 Universidad De Santiago De Chile Running Time Prediction Algorithm for WAND Queries
KR102492318B1 (ko) * 2015-09-18 2023-01-26 삼성전자주식회사 모델 학습 방법 및 장치, 및 데이터 인식 방법
CN106548234A (zh) * 2016-11-17 2017-03-29 北京图森互联科技有限责任公司 一种神经网络剪枝方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010042256A2 (en) * 2008-05-30 2010-04-15 The University Of Memphis Research Foundation Methods of improved learning in simultaneous recurrent neural networks
CN105894847A (zh) * 2016-06-27 2016-08-24 华南理工大学 一种云平台环境下无监督学习的实时公交动态调度系统及方法
CN106355248A (zh) * 2016-08-26 2017-01-25 深圳先进技术研究院 一种深度卷积神经网络训练方法及装置
CN106779068A (zh) * 2016-12-05 2017-05-31 北京深鉴智能科技有限公司 调整人工神经网络的方法和装置

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111695375B (zh) * 2019-03-13 2021-04-20 上海云从企业发展有限公司 基于模型蒸馏的人脸识别模型压缩方法、介质及终端
CN111695375A (zh) * 2019-03-13 2020-09-22 上海云从企业发展有限公司 基于模型蒸馏的人脸识别模型压缩算法、介质及终端
US11783195B2 (en) 2019-03-27 2023-10-10 Cognizant Technology Solutions U.S. Corporation Process and system including an optimization engine with evolutionary surrogate-assisted prescriptions
US11403523B2 (en) 2019-06-07 2022-08-02 Tata Consultancy Services Limited Sparsity constraints and knowledge distillation based learning of sparser and compressed neural networks
WO2021068180A1 (en) * 2019-10-11 2021-04-15 Beijing Didi Infinity Technology And Development Co., Ltd. Method and system for continual meta-learning
RU2749970C1 (ru) * 2019-10-24 2021-06-21 Бейдзин Сяоми Интиллиджент Текнолоджи Ко., ЛТД. Способ сжатия модели нейронной сети, а также способ и устройство для перевода языкового корпуса
US11556723B2 (en) 2019-10-24 2023-01-17 Beijing Xiaomi Intelligent Technology Co., Ltd. Neural network model compression method, corpus translation method and device
CN111695483B (zh) * 2020-06-05 2022-04-08 腾讯科技(深圳)有限公司 一种车辆违章检测方法、装置、设备及计算机存储介质
CN111695483A (zh) * 2020-06-05 2020-09-22 腾讯科技(深圳)有限公司 一种车辆违章检测方法、装置、设备及计算机存储介质
US11775841B2 (en) 2020-06-15 2023-10-03 Cognizant Technology Solutions U.S. Corporation Process and system including explainable prescriptions through surrogate-assisted evolution
CN112215353B (zh) * 2020-09-29 2023-09-01 电子科技大学 一种基于变分结构优化网络的通道剪枝方法
CN112215353A (zh) * 2020-09-29 2021-01-12 电子科技大学 一种基于变分结构优化网络的通道剪枝方法
CN112487899A (zh) * 2020-11-19 2021-03-12 武汉高德飞行器科技有限公司 基于无人机的目标识别方法、系统、存储介质及电子设备
CN114881227B (zh) * 2022-05-13 2023-07-04 北京百度网讯科技有限公司 模型压缩方法、图像处理方法、装置和电子设备
CN114881227A (zh) * 2022-05-13 2022-08-09 北京百度网讯科技有限公司 模型压缩方法、图像处理方法、装置和电子设备
CN115496207B (zh) * 2022-11-08 2023-09-26 荣耀终端有限公司 神经网络模型压缩方法、装置和系统
CN115496207A (zh) * 2022-11-08 2022-12-20 荣耀终端有限公司 神经网络模型压缩方法、装置和系统
CN117011617A (zh) * 2023-10-07 2023-11-07 之江实验室 基于二阶段教师-学生框架的肺结节检测装置及构建方法
CN117011617B (zh) * 2023-10-07 2024-03-22 之江实验室 基于二阶段教师-学生框架的肺结节检测装置及构建方法

Also Published As

Publication number Publication date
CN108334934A (zh) 2018-07-27
CN108334934B (zh) 2021-04-13

Similar Documents

Publication Publication Date Title
WO2018223822A1 (zh) 基于剪枝和蒸馏的卷积神经网络压缩方法
CN111340814B (zh) 一种基于多模态自适应卷积的rgb-d图像语义分割方法
US11657264B2 (en) Content-specific neural network distribution
CN112070209B (zh) 基于w距离的稳定可控图像生成模型训练方法
WO2020015330A1 (zh) 基于增强的神经网络的图像复原方法、存储介质及系统
WO2020237646A1 (zh) 图像处理方法、设备及计算机可读存储介质
CN110163815A (zh) 基于多阶段变分自编码器的低照度还原方法
KR20220130630A (ko) 이미지 처리 방법, 안면 인식 모델 훈련 방법, 장치 및 기기
CN107784296A (zh) 一种低分辨率图像的人脸识别方法
CN111222532A (zh) 具有分类精度保持和带宽保护的边云协同深度学习模型训练方法
WO2022262660A1 (zh) 针对超分辨率网络的剪枝量化压缩方法、系统及介质
WO2023050738A1 (zh) 基于知识蒸馏的模型训练方法、装置、电子设备
CN105913001A (zh) 基于聚类的在线式多人脸图像处理的方法
CN111797707A (zh) 一种基于聚类的镜头关键帧提取方法
CN114973229A (zh) 文本识别模型训练、文本识别方法、装置、设备及介质
CN115797835A (zh) 一种基于异构Transformer的无监督视频目标分割算法
WO2022077903A1 (zh) 基于二值神经网络的局部激活方法与系统
CN111479286B (zh) 一种边缘计算系统减少通信流量的数据处理方法
CN111291663B (zh) 一种利用时空信息的快速视频目标物体分割方法
CN110163489B (zh) 一种戒毒运动锻炼成效评价方法
CN110958417B (zh) 一种基于语音线索的视频通话类视频去除压缩噪声的方法
CN109714629A (zh) 一种定格动画的生成方法和生成系统
US20220122281A1 (en) Depth information processing method, apparatus, and storage medium
CN115049055A (zh) 基于动态双可训练界限的超分神经网络的量化方法
CN113129869B (zh) 语音识别模型的训练与语音识别的方法、装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18813181

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18813181

Country of ref document: EP

Kind code of ref document: A1