JPWO2020202313A1 - Neural network data compression device and data compression method - Google Patents

Abstract

The data compression device 701 is a data compression device for a neural network that reduces the amount of data of the output data of the partial neural network created by dividing one neural network, and is a storage in which the output data of the partial neural network is stored. The read circuit 711 is provided with a read circuit 711 that reads data from the unit and a data compression circuit 712 that compresses the read data. Read the data of.

Description

The present invention relates to a data compression device and a data compression method for a neural network that reduces the amount of data of the output data of a partial neural network created by dividing one neural network.

Neural networks are used in many fields. Since the amount of calculation in a neural network, particularly in a multi-layer neural network having a large number of hidden layers, is enormous, the calculation is executed by a server installed in a data center having high computing power or a cloud server.

However, there are problems such as congestion in the traffic between the user and the server and high power consumption of the server. Therefore, it is devised to use edge computing. Specifically, the neural network is divided, and the edge device executes the calculation of one partial neural network (referred to as the first half neural network) generated by the division. For example, the cloud server performs the operation of the other partial neural network (second half neural network). (See, for example, Non-Patent Document 1). In that case, the feature amount data which is the output of the final layer in the first half neural network is transferred from the edge device to the cloud server via the communication network. In Non-Patent Document 1, the neural network is a convolutional neural network (CNN), and the final layer in the first half neural network is a certain pooling layer.

Non-Patent Document 1 describes that the feature amount data is subjected to data compression processing in order to reduce the amount of data transferred from the edge device to the cloud server. In the data compression process, first, feature data in which multiple channels of small-sized channel image data exist are arranged in a tile so that coding based on the HEVC (High Efficiency Video Coding) standard can be performed, and the large-sized channel image data is arranged in a tile. A process of converting channel image data into image data having one channel or three channels is performed. After that, coding based on the HEVC standard is performed.

FIG. 15 is an explanatory diagram for explaining the data compression process of the feature amount data described in Non-Patent Document 1. FIG. 15A shows how the feature data composed of a plurality of channels are arranged in chronological order for each of the three frames fi, fi + 1, and fi + 2. For example, it is assumed that the size of the output window is n × n (n: a natural number of 2 or more) and the number of channels is K (K: a natural number of 2 or more). Here, the size n of the output window may be unsuitable for the data compression processing method. For example, in Non-Patent Document 1, since the HEVC encoder uses the CTU size as the minimum unit for coding, it has been pointed out that the HEVC encoder cannot be applied as it is when the feature amount data size n is smaller than the CTU size. There is. As a method for dealing with such a case, Non-Patent Document 1 discloses a method of combining a plurality of channel images into a single image in a tile shape.

FIG. 15B shows a state in which output windows for 9 channels are arranged in tiles for each frame. Multiple output windows arranged in tiles are considered as one frame. Then, the coding is executed for the frame. Hereinafter, the data before being encoded, that is, the output data of the first half neural network is referred to as intermediate data.

T. Mitani et al., "Compression and Aggregation for Optimizing Information Transmission in Distributed CNN", IEEE, 2017 fifth International Symposium on Computing and Networking, 19-22 Nov. 2017

FIG. 16 is an explanatory diagram showing one frame in which output windows for 9 channels are arranged in a tile. In FIG. 16, the arrows indicate the coding order. However, FIG. 16 shows the coding order only for the first to third output channels (output channels with numbers 1 to 3 in FIG. 16B). Further, the case where the size of the output window is 3 × 3 is illustrated.

For example, the coding order is the first line of the first output channel → the first line of the second output channel → the first line of the third output channel → the second line of the first output channel → ... Taking CNN as an example, the output of the convolution layer and the output of the pooling layer are stored in the memory as feature data for the next layer. RAM (Random Access Memory) is used as the memory. Generally, the order of the storage addresses is the data of the first output channel → the data of the second output channel → ... → the data of the Kth output channel.

That is, the data storage order and the coding order in the memory are not the same. Then, it is required to execute the process of matching the order of writing data to the memory or the order of reading data from the memory to the coding order. Since such processing is executed, the processing time in the first half neural network increases. In addition, memory management becomes complicated. In other words, the efficiency of the processing (including the coding processing) of the first half neural network is reduced.

An object of the present invention is to efficiently reduce the amount of output data of a partial neural network created by dividing one neural network.

The data compression device of the neural network according to the present invention is a data compression device of the neural network that reduces the amount of data of the output data of the partial neural network created by dividing one neural network, and is the output data of the partial neural network. The read circuit includes a read circuit for reading data from the storage unit in which the data is stored and a data compression circuit for compressing the read data. Read the data of the channel of.

The data compression device of the neural network of another aspect according to the present invention is a data compression device of the neural network that reduces the amount of data of the output data of the partial neural network created by dividing one neural network, and is a partial neural. The data compression circuit includes a read circuit for reading data from a storage unit in which output data of the network is stored and a data compression circuit for compressing the read data. Encodes the difference between the data of one similar channel and the data of another similar channel and the coding unit that compresses the data of one of the two or more similar channels that are similar. It has a difference coding unit.

The data compression method of the neural network according to the present invention is a data compression method of the neural network that reduces the amount of data of the output data of the partial neural network created by dividing one neural network, and is the output data of the partial neural network. Data is read from the storage unit in which the data is stored, data is compressed from the read data, and when the data is read from the storage unit, after all the data in one channel has been read out, the data in the other channel is completed. Is read.

Another aspect of the data compression method for a neural network according to the present invention is a data compression method for a neural network that reduces the amount of output data of a partial neural network created by dividing one neural network, and is a partial neural network. When data is read from the storage unit where the output data of the network is stored, the data of the read data is compressed, and the data is compressed, two or more of the output data are similar, which is indicated by the similar channel information. Data compression of the data of one of the similar channels in is performed, and the difference between the data of the other similar channels and the data of one similar channel is encoded.

The data compression program of the neural network according to the present invention is a data compression program of the neural network that reduces the amount of data of the output data of the partial neural network created by dividing one neural network, and is a data compression program of the partial neural network to a computer. When the process of reading data from the storage unit in which the output data of the above is stored and the process of compressing the read data are executed and the data is read from the storage unit, all the data in one channel is read out. Is completed, and then the data of other channels is read out.

Another aspect of the neural network data compression program according to the present invention is a neural network data compression program that reduces the amount of output data of a partial neural network created by dividing one neural network into a computer. , The process of reading data from the storage unit in which the output data of the partial neural network is stored and the process of compressing the read data are executed, and when the data is compressed, it is indicated by similar channel information. Encoding the difference between the process of compressing the data of one of the two or more similar channels whose output data is similar and the data of the other similar channel with the data of one similar channel. To execute the processing to be performed.

According to the present invention, the amount of data of the output data of the partial neural network created by dividing one neural network is efficiently reduced.

It is a block diagram which shows an example of a neural network system. It is explanatory drawing which shows an example of CNN. It is a block diagram which shows the structural example of the data compression part in 1st Embodiment. It is explanatory drawing for demonstrating the coding order in 1st Embodiment. It is a flowchart which shows an example of the operation of the data compression part in 1st Embodiment. It is a block diagram which shows the structural example of the data compression part in 2nd Embodiment. It is explanatory drawing for demonstrating the coding order in 2nd Embodiment. It is a flowchart which shows an example of the operation of the data compression part in 2nd Embodiment. It is a block diagram which shows the structural example of the data compression part in 3rd Embodiment. It is a flowchart which shows an example of the operation of the data compression part in 3rd Embodiment. It is a block diagram which shows the main part of the data compression apparatus of a neural network. It is a block diagram which shows the main part of the data compression apparatus of the neural network of another aspect. It is a block diagram which shows the main part of the data compression apparatus of the neural network of the aspect of. It is a block diagram which shows an example of the computer which has a CPU. It is explanatory drawing for demonstrating the data compression processing of the conventional feature quantity data. It is explanatory drawing which shows one frame in the state that the output windows for 9 channels are arranged in a tile shape.

FIG. 1 is a block diagram showing an example of a neural network system. The neural network system includes a neural network device, a communication network 500 and a cloud server 600. The neural network device includes a first half neural network 200, a data compression unit 100, and a transmission unit 400.

The first half neural network 200 includes one or more convolution layers. Here, CNN is taken as an example as a neural network. The data compression unit 100 executes a process for reducing the amount of intermediate data created by the first half neural network 200. The transmission unit 400 transmits the data output by the data compression unit 100 to the cloud server 600 via the communication network 500.

The cloud server 600 is taken as an example of the calculation means for executing the calculation of the latter half neural network, but the calculation means is not limited to the cloud server 600. The arithmetic means may be, for example, an on-premises server.

The first half neural network 200 includes a product-sum calculation unit 210, a storage unit 220, and a selection unit 230. The product-sum calculation unit 210 executes the operation of the convolution layer using the feature amount data stored in the storage unit 220 and the weighting coefficient, and stores the calculation result in the storage unit 220. A RAM is used as the storage unit 220.

The selection unit 230 realizes a pooling layer. That is, the selection unit 230 selects one calculation result for each predetermined number of calculation results of the product-sum calculation unit 210 stored in the storage unit 220. The selection unit 230 selects, for example, the maximum value among a predetermined number of calculation results. The selection unit 230 may select, for example, an average value of a predetermined number of calculation results.

FIG. 2 is an explanatory diagram showing an example of CNN. In the example shown in FIG. 2, the CNN is a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a third convolution layer, a fourth convolution layer, and a fifth convolution layer for input data. , And a fifth pooling layer.

The first half neural network 200, for example, has a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a third convolution layer, a fourth convolution layer, a fifth convolution layer, and a fifth pooling. Consists of layers. Further, the latter half neural network 300 includes a plurality of layers after the fifth pooling layer.

Embodiment 1.
FIG. 3 is a block diagram showing a configuration example of the data compression unit according to the first embodiment. The data compression unit 100A shown in FIG. 3 has a read circuit 111 and a coding unit 112. The data compression unit 100A corresponds to the data compression unit 100 in the neural network device shown in FIG.

The read circuit 111 reads the feature amount data from the storage unit 220. The coding unit 112 encodes the read feature amount data. The coding unit 112 performs coding based on, for example, the HEVC standard. In the present embodiment, the coding unit 112 executes the coding process based on the HEVC standard, but the coding based on the HEVC standard is an example, and the coding unit 112 uses another coding method. You may. For example, the next-generation coding method studied by JVET (Joint Video Experts Team) may be used.

FIG. 4 is an explanatory diagram for explaining the coding order in the present embodiment. In the present embodiment, the output data of the first half neural network 200 is regarded as an independent moving image for each channel. FIG. 4A shows how the feature data composed of a plurality of channels are arranged in chronological order for each of the three frames fi, fi + 1, and fi + 2. For example, it is assumed that the size of the output window is n × n (n: a natural number of 2 or more) and the number of channels is K (K: a natural number of 2 or more). Here, the size n of the output window may be unsuitable for the data compression processing method. As described above, since the HEVC encoder uses the CTU size as the minimum unit for coding, the HEVC encoder cannot be applied as it is when the size n of the feature amount data is smaller than or equal to the CTU size.

Therefore, as shown in FIG. 4B, the data read from the storage unit 220 are arranged in a strip shape in the horizontal direction. There are a plurality of strips (4 in the example shown in FIG. 4B). In FIG. 4B, the numbers indicate the strip numbers. Data of the same channel in a certain frame is set in one strip. Also, when the arrangement of data in a strip is completed, the next channel data is arranged.

Next, the operation of the data compression unit 100A will be described. FIG. 5 is a flowchart showing an example of the operation of the data compression unit 100A. The number of frames is F (F: a natural number of 2 or more).

First, the read circuit 111 sets the variable f regarding the number of frames to 1 (step S101) and sets the variable k regarding the number of channels to 1 (step S102).

The read circuit 111 reads the data of the kth channel of the fth frame from the storage unit 220 (step S103). Taking the case where the size of the output window is 3 × 3 as an example, the data reading order is as follows: 1st line of the 1st output channel of the fth frame → 2nd line of the 1st output channel → 1st output channel. The third line of → ...

Then, the read circuit 111 stores the read data so that the data can be arranged in a strip shape in the temporary storage unit (not shown) as a storage circuit. Then, the read circuit 111 increments the value of k by 1 (step S104). The temporary storage unit is built in the reading circuit 111, or is provided between the reading circuit 111 and the coding unit 112.

If the value of k exceeds K (step S105), the reading of the data of all channels is completed, so that the coding unit 112 encodes the data stored in the temporary storage unit. Is executed (step S106). If the value of k does not exceed K, the process returns to step S103.

After the process of step S106 is executed, the read circuit 111 increments the value of f by 1 (step S107). When the value of f does not exceed F (step S108), the process returns to step S102.

If the value of f exceeds F, the process ends.

In the present embodiment, the data of a plurality of channels are collectively regarded as one image for each frame. In the example shown in FIG. 5, the data of all channels are collectively regarded as one image for each frame, but the data compression unit 100A collects data of a predetermined number of channels smaller than the total number of channels. You may consider them together as one image.

Then, in the present embodiment, the coding process is executed after the data of the same channel read from the storage unit 220 is arranged in a strip shape in the horizontal direction. The order of reading data from the storage unit 220 is consistent with the order of storing data (address order) in the storage unit 220. Therefore, it is not necessary to execute the process of matching the order of writing data to the storage unit 220 or the order of reading data from the storage unit 220 to the coding order. As a result, the decrease in efficiency of the processing (including the coding processing) of the first half neural network is suppressed.

Embodiment 2.
FIG. 6 is a block diagram showing a configuration example of the data compression unit according to the second embodiment. The data compression unit 100B shown in FIG. 6 has a read circuit 113 and a coding unit 112. The data compression unit 100B corresponds to the data compression unit 100 in the neural network device shown in FIG.

The read circuit 113 reads the feature amount data from the storage unit 220. The coding unit 112 encodes the read feature amount data. The coding unit 112 performs coding based on, for example, the HEVC standard. In the present embodiment, the coding unit 112 executes the coding process based on the HEVC standard, but the coding based on the HEVC standard is an example, and the coding unit 112 uses another coding method. You may.

FIG. 7 is an explanatory diagram for explaining the coding order in the present embodiment. In the present embodiment, the output data of the first half neural network 200 is regarded as an independent moving image for each channel. FIG. 7A shows how the output of one frame is regarded as an independent moving image for each channel. For example, if the size of the output window is n × n (n: a natural number of 2 or more) and the number of channels is K (K: a natural number of 2 or more), it is considered that there are K types of n × n videos. Is done. As mentioned above, n needs to be CTU size or larger in order to apply the HEVC encoder. In this embodiment, it is premised that n is CTU size or more.

In this embodiment, as shown in FIG. 7B, the output data of the same channel of all frames is regarded as one video. Then, each of the images is independently encoded. In FIG. 7B, the coded data of each image is represented as a k-th stream (k: 1 to K).

Next, the operation of the data compression unit 100B will be described. FIG. 8 is a flowchart showing an example of the operation of the data compression unit 100B. The number of frames is F (F: a natural number of 2 or more). Further, the case where the first half neural network 200 outputs K channel data is taken as an example.

First, the read circuit 113 sets the variable k related to the number of channels to 1 (step S201), and sets the variable f related to the number of frames to 1 (step S202).

The read circuit 113 reads the data of the fth frame of the kth channel from the storage unit 220 (step S203). Taking the case where the size of the output window is 3 × 3 as an example, the data reading order is the same as in the case of the first embodiment, from the first line of the first output channel of the first frame to the first output. The second line of the channel → the third line of the first output channel → the first line of the first output channel of the second frame → ...

Then, the read circuit 113 increases the value of f by 1 (step S204). When the value of f does not exceed F (step S205), the process returns to step S203.

When the value of f exceeds F, it means that all the data for one channel is stored in the temporary storage unit. The coding unit 112 executes a coding process for the data stored in the temporary storage unit (step S207). That is, the coding unit 112 executes the coding process for all the data of one channel (which is regarded as one video).

Next, the read circuit 113 increments the value of k by 1 (step S208). If the value of k exceeds K, the processing is completed for the data of all channels, and the data compression unit 100B ends the processing (step S209). If the value of k does not exceed K, the process returns to step S202.

In the present embodiment, the feature amount data is treated as a separate moving image for each channel and compressed by a moving image encoder (for example, HEVC encoder). One channel can be regarded as an image obtained by extracting one of the features of the input image. Therefore, it is expected that the moving image (k-th stream) obtained by extracting only the same channel from each frame has a temporal locality (similarity) as in the input image sequence. For example, it is expected that most of the feature image of channel 1 of frame fi and the feature image of channel 1 of frame fi + 1 are the same. Therefore, it is possible to effectively utilize the effect of the algorithm that performs compression with high efficiency by utilizing the temporal locality of the input image sequence possessed by the HEVC encoder.

Embodiment 3.
FIG. 9 is a block diagram showing a configuration example of the data compression unit according to the third embodiment. The data compression unit 100C shown in FIG. 9 includes a read circuit 114, a coding unit 115, and a difference coding unit 116. The data compression unit 100C corresponds to the data compression unit 100 in the neural network device shown in FIG.

In the present embodiment, the data compression unit 100C inputs similar channel information that specifies similar channels. Depending on the type of data handled by the CNN, other channels including output data similar to the output data of one channel can be grasped to a considerable extent with respect to the trained CNN. Hereinafter, a plurality of channels having similar output data will be referred to as similar channels. For example, when image data is handled, a plurality of similar channels corresponding to similar image features can be determined in advance (before the processing of the neural network is performed). Similar channel information is information indicating two or more predetermined similar channels.

The coding unit 115 encodes the data of one of two or more similar channels. The difference coding unit 116 encodes the difference between the data of another similar channel and the data of one channel.

Next, the operation of the data compression unit 100C will be described. FIG. 10 is a flowchart showing an example of the operation of the data compression unit 100B.

The read circuit 114 inputs similar channel information from the outside (step S301). Then, the read circuit 114 reads the data of one of the two or more similar channels indicated by the similar channel information from the storage unit 220 (step S302).

The coding unit 115 encodes the data of one read channel (step S303). The coding unit 115 may use any coding method as long as the amount of data can be reduced. For example, ^{a coding method in which data is quantized to 2 n} (n: natural number) and then frequency-converted can be used. Further, a predictive coding method such as HEVC can be used.

The read circuit 114 reads the data of the other channel of the two or more similar channels from the storage unit 220 (step S304). The difference coding unit 116 calculates the difference between the read data of another channel and the data of one channel. Then, the difference coding unit 116 encodes the difference. The difference coding unit 116 may use any coding method as long as the amount of data can be reduced. For example, the difference coding unit 116 can use run-length coding or Huffman coding.

When the coding processing by the coding unit 115 or the difference coding unit 116 is executed for the output data of all channels, the processing is terminated. If there is still a channel that has not been coded, the process proceeds to step S302.

In the present embodiment, with respect to the data of the channel of the data similar to the data of one channel, the data of the difference from the data of one channel (data of similar channels) is transmitted to the cloud server 600, so that the first method is made. As in the case of the embodiment and the second embodiment, the amount of transmitted data can be effectively reduced.

In each of the above embodiments, CNN is taken as an example, but each embodiment can be applied to a multi-layer neural network other than CNN.

FIG. 11 is a block diagram showing a main part of a data compression device of a neural network. The data compression device 701 shown in FIG. 11 is a data compression device for a neural network that reduces the amount of output data of a partial neural network (for example, the first half neural network 200) created by dividing one neural network. , Read circuit 711 (for example, read circuits 111, 113) that reads data from the storage unit (for example, storage unit 220) in which the output data of the partial neural network is stored, and data compression that performs data compression of the read data. A circuit 712 (for example, a coding unit 112) is provided, and the read circuit 711 reads the data of another channel after the read of all the data of one channel is completed.

FIG. 12 is a block diagram showing a main part of a data compression device of another aspect of a neural network. In the data compression device 701 shown in FIG. 12, the read circuit 711 stores the data of one channel in the strip-shaped area of the storage circuit 713, and the data compression circuit 712 stores the data of the data stored in the strip-shaped area. Perform compression.

FIG. 13 is a block diagram showing a main part of the data compression device of the neural network of still another aspect. The data compression device 702 shown in FIG. 13 is a data compression device for a neural network that reduces the amount of data of the output data of the partial neural network created by dividing one neural network, and the output data of the partial neural network is A read circuit 711 (for example, a read circuit 114) that reads data from the stored storage unit and a data compression circuit 714 that compresses the read data are provided, and the data compression circuit 714 is indicated by similar channel information. The coding unit 715 (for example, the coding unit 115) that compresses the data of one of the two or more similar channels having similar output data, and the data of the other similar channels. , A difference coding unit 716 (for example, a difference coding unit 116) that encodes a difference from the data of one similar channel.

FIG. 14 is a block diagram showing an example of a computer having a CPU (Central Processing Unit). The CPU 1000 realizes each function in the above embodiment by being stored in the storage device 1001 and executing the process according to the data compression program.

That is, the CPU 1000 realizes the functions of the read circuit 111 and the coding unit 112 shown in FIG. Further, the CPU 1000 can realize the functions of the read circuit 113 and the coding unit 112 shown in FIG. Further, the CPU 1000 can realize the functions of the read circuit 114, the coding unit 115, and the difference coding unit 116 shown in FIG. Further, the functions of the data compression devices 701 and 702 (excluding the storage circuit 713) shown in FIGS. 11 to 13 can be realized.

The storage device 1001 is, for example, a non-transitory computer readable medium. Non-temporary computer-readable media include various types of tangible storage media. Specific examples of non-temporary computer-readable media include semiconductor memories (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM).

The memory 1002 is realized by, for example, a RAM (Random Access Memory), and is a storage means for temporarily storing data when the CPU 1000 executes processing. The storage circuit 713 shown in FIG. 12 can be realized by the memory 1002.

Although some or all of the above embodiments may be described as in the appendix below, the configuration of the present invention is not limited to the following configurations.

(Appendix 1) A neural network data compression device that reduces the amount of output data of a partial neural network created by dividing one neural network.
A read circuit that reads data from the storage unit in which the output data of the partial neural network is stored, and
It is equipped with a data compression circuit that compresses the read data.
The read circuit is a data compression device for a neural network, characterized in that data of another channel is read after all data of one channel has been read.

(Appendix 2) The read-out circuit stores data of one channel in a strip-shaped area in the storage circuit.
The data compression circuit is a data compression device for a neural network according to Appendix 1, which compresses data stored in the strip-shaped region.

(Appendix 3) The data compression circuit is a data compression device for a neural network according to Appendix 1, wherein a plurality of data of each channel of a plurality of channels are regarded as one video and data compression is performed.

(Appendix 4) The data compression circuit is a neural network data compression device according to Appendix 2 or Appendix 3 that executes a coding process based on the HEVC standard.

(Appendix 5) A neural network data compression device that reduces the amount of output data of a partial neural network created by dividing one neural network.
A read circuit that reads data from the storage unit in which the output data of the partial neural network is stored, and
It is equipped with a data compression circuit that compresses the read data.
The data compression circuit is
A coding unit that compresses the data of one of two or more similar channels whose output data is similar, which is indicated by the similar channel information, and the data of the similar channel.
A data compression device for a neural network, comprising a difference coding unit that encodes a difference between data of another similar channel and data of the one similar channel.

(Appendix 6) A neural network data compression method that reduces the amount of output data of a partial neural network created by dividing one neural network.
Data is read from the storage unit in which the output data of the partial neural network is stored, and the data is read.
Compress the read data and perform data compression.
A method for compressing data in a neural network, which comprises reading data from another channel after reading all data in one channel when reading data from the storage unit.

(Appendix 7) The data of one channel is stored in a strip-shaped area in the storage circuit.
The data compression method for a neural network according to Appendix 6, which compresses the data stored in the strip-shaped area.

(Appendix 8) The data compression method of the neural network of Appendix 6 in which a plurality of data of each channel of a plurality of channels are regarded as one video and data compression is performed.

(Appendix 9) A neural network data compression method that reduces the amount of output data of a partial neural network created by dividing one neural network.
Data is read from the storage unit in which the output data of the partial neural network is stored, and the data is read.
Compress the read data and perform data compression.
When performing the data compression,
Data compression of the data of one of two or more similar channels with similar output data, which is indicated by the similar channel information, is performed.
A data compression method for a neural network, characterized in that the difference between the data of another similar channel and the data of the one similar channel is encoded.

(Appendix 10) A neural network data compression program that reduces the amount of output data of a partial neural network created by dividing one neural network.
On the computer
The process of reading data from the storage unit in which the output data of the partial neural network is stored, and
The process of compressing the read data is executed, and the data is compressed.
A neural network data compression program for reading data from another channel after the reading of all data in one channel is completed when reading data from the storage unit.

(Appendix 11) To the computer
The process of storing the data of one channel in a strip-shaped area in the storage circuit,
The data compression program of the neural network of Appendix 10 for executing the process of compressing the data stored in the strip-shaped area.

(Appendix 12) To the computer
The data compression program of the neural network of Appendix 10 for executing a process of performing data compression by regarding a plurality of data of each channel of a plurality of channels as one video.

(Appendix 13) A neural network data compression program that reduces the amount of output data of a partial neural network created by dividing one neural network.
On the computer
The process of reading data from the storage unit in which the output data of the partial neural network is stored, and
The process of compressing the read data is executed, and the data is compressed.
When performing the data compression,
The process of compressing the data of one of two or more similar channels whose output data is similar, which is indicated by the similar channel information, and the data of the similar channel.
A neural network data compression program for executing a process of encoding the difference between the data of another similar channel and the data of the one similar channel.

100, 100A, 100B, 100C Data compression unit 111, 113, 114 Read circuit 112, 115 Coding unit 116 Difference coding unit 200 First half Neural network 210 Product sum calculation unit 220 Storage unit 230 Selection unit 400 Transmission unit 500 Communication network 600 Cloud server 701,702 Data compression device 711 Read circuit 712,714 Data compression circuit 713 Storage circuit 715 Coding unit 716 Difference coding unit 1000 CPU
1001 storage device 1002 memory

Claims (13)

A neural network data compression device that reduces the amount of output data of a partial neural network created by dividing one neural network.
A read circuit that reads data from the storage unit in which the output data of the partial neural network is stored, and
It is equipped with a data compression circuit that compresses the read data.
The read circuit is a data compression device for a neural network, characterized in that data of another channel is read after all data of one channel has been read. The read circuit stores the data of one channel in a strip-shaped area in the storage circuit.
The data compression device for a neural network according to claim 1, wherein the data compression circuit compresses data stored in the strip-shaped region. The data compression device for a neural network according to claim 1, wherein the data compression circuit regards a plurality of data of each channel of the plurality of channels as one video and performs data compression. The data compression device for a neural network according to claim 2 or 3, wherein the data compression circuit executes a coding process based on the HEVC standard. A neural network data compression device that reduces the amount of output data of a partial neural network created by dividing one neural network.
A read circuit that reads data from the storage unit in which the output data of the partial neural network is stored, and
It is equipped with a data compression circuit that compresses the read data.
The data compression circuit is
A coding unit that compresses the data of one of two or more similar channels whose output data is similar, which is indicated by the similar channel information, and the data of the similar channel.
A data compression device for a neural network, comprising a difference coding unit that encodes a difference between data of another similar channel and data of the one similar channel. A neural network data compression method that reduces the amount of output data of a partial neural network created by dividing one neural network.
Data is read from the storage unit in which the output data of the partial neural network is stored, and the data is read.
Compress the read data and perform data compression.
A method for compressing data in a neural network, which comprises reading data from another channel after reading all data in one channel when reading data from the storage unit. The data of one channel is stored in a strip-shaped area in the storage circuit,
The data compression method for a neural network according to claim 6, wherein the data stored in the strip-shaped area is compressed. The data compression method for a neural network according to claim 6, wherein a plurality of data of each channel of a plurality of channels are regarded as one video and data compression is performed. A neural network data compression method that reduces the amount of output data of a partial neural network created by dividing one neural network.
Data is read from the storage unit in which the output data of the partial neural network is stored, and the data is read.
Compress the read data and perform data compression.
When performing the data compression,
Data compression of the data of one of two or more similar channels with similar output data, which is indicated by the similar channel information, is performed.
A data compression method for a neural network, characterized in that the difference between the data of another similar channel and the data of the one similar channel is encoded. A neural network data compression program that reduces the amount of output data of a partial neural network created by dividing one neural network.
On the computer
The process of reading data from the storage unit in which the output data of the partial neural network is stored, and
The process of compressing the read data is executed, and the data is compressed.
A neural network data compression program for reading data from another channel after the reading of all data in one channel is completed when reading data from the storage unit. On the computer
The process of storing the data of one channel in a strip-shaped area in the storage circuit,
The data compression program for a neural network according to claim 10, wherein the process of compressing the data stored in the strip-shaped area is executed. On the computer
The data compression program for a neural network according to claim 10, wherein a plurality of data of each channel of a plurality of channels are regarded as one video and a process of performing data compression is executed. A neural network data compression program that reduces the amount of output data of a partial neural network created by dividing one neural network.
On the computer
The process of reading data from the storage unit in which the output data of the partial neural network is stored, and
The process of compressing the read data is executed, and the data is compressed.
When performing the data compression,
The process of compressing the data of one of two or more similar channels whose output data is similar, which is indicated by the similar channel information, and the data of the similar channel.
A neural network data compression program for executing a process of encoding the difference between the data of another similar channel and the data of the one similar channel.

Images