109 results about "Quantized neural networks" patented technology

A method of generating a fixed-point quantized neural network includes analyzing a statistical distribution for each channel of floating-point parameter values of feature maps and a kernel for each channel from data of a pre-trained floating-point neural network, determining a fixed-point expression of each of the parameters for each channel statistically covering a distribution range of the floating-point parameter values based on the statistical distribution for each channel, determining fractional lengths of a bias and a weight for each channel among the parameters of the fixed-point expression for each channel based on a result of performing a convolution operation, and generating a fixed-point quantized neural network in which the bias and the weight for each channel have the determined fractional lengths.

The invention discloses a neural-network-model compression method, system and device and a computer-readable storage medium. The method includes: utilizing a neural-network clipping method to clip a to-be-clipped neural network model to obtain a to-be-quantized neural network model; utilizing an INQ algorithm to quantize the to-be-quantized neural network model to obtain a to-be-stored neural network model; and using a compression format to store the to-be-stored neural network model. Therefore, according to the neural-network-model compression method provided by the embodiment of the invention, the neural network model is cropped, the INQ algorithm is used at the same time to quantify the same after cropping, thus a model size can be reduced in a case of effectively guaranteeing no loss of model precision after compression, thus the problem of excessive consumed resources can be solved, and calculation can be accelerated.

A neural network method of parameter quantization includes obtaining channel profile information for first parameter values of a floating-point type in each channel included in each of feature maps based on an input in a first dataset to a floating-point parameters pre-trained neural network; determining a probability density function (PDF) type, for each channel, appropriate for the channel profile information based on a classification network receiving the channel profile information as a dataset; determining a fixed-point representation, based on the determined PDF type, for each channel, statistically covering a distribution range of the first parameter values; and generating a fixed-point quantized neural network based on the fixed-point representation determined for each channel.

The invention relates to the field of artificial intelligence. The embodiment of the invention discloses a model generation method and device. The method comprises the following steps: generating a neural network model for executing a deep learning task by sequentially executing multiple iterative operations; wherein the iterative operation comprises the steps of determining a current quantizationmethod of each network structure unit in a quantization method search space corresponding to each network structure unit of a preset neural network model based on a current reward feedback value so as to update a quantization strategy of the preset neural network model; quantifying a preset neural network model based on the updated quantification strategy; obtaining the performance of the quantized neural network model, and updating a reward feedback value; and in response to determining that the reward feedback value reaches a preset convergence condition or the number of times of the iterative operation reaches a preset threshold, determining that the current quantized neural network model is a generated neural network model for executing the deep learning task. The method can reduce the memory space occupied by the neural network model.

A method and processing unit for training a neural network to selectively quantize weights of a filter of the neural network as either binary weights or ternary weights. A plurality of training iterations a performed that each comprise: quantizing a set of real-valued weights of a filter to generate a corresponding set of quantized weights; generating an output feature tensor based on matrix multiplication of an input feature tensor and the set of quantized weights; computing, based on the output feature tensor, a loss based on a regularization function that is configured to move the loss towards a minimum value when either: (i) the quantized weights move towards binary weights, or (ii) the quantized weights move towards a ternary weights; computing a gradient with an objective of minimizing the loss; updating the real-valued weights based on the computed gradient. When the training iterations are complete, a set of weights quantized from the updated real-valued weights is stored as either a set of binary weights or a set of ternary weights.

A method and a device for adjusting an artificial neural network (ANN) are provided. The ANN at least comprises a plurality of layers, and the method comprises the steps: obtaining a to-be-trained neural network model; training the neural network model by using high-bit fixed-point quantization to obtain a trained high-bit fixed-point quantization neural network model; carrying out fine tuning onthe high-bit fixed-point quantization neural network model by using low bits to obtain a trained neural network model with low-bit fixed-point quantization; and outputting the trained neural network model with low-bit fixed-point quantization. According to the neural network training scheme with the bit width gradually reduced, the training and deployment of the neural network are considered, so that the calculation precision comparable to that of a floating point network can be realized under the condition of extremely low bit width.

The invention discloses an intelligent NIPS (Network Intrusion Prevention System) framework for quantifying a neural network based on a mobile agent (MA) and a learning vector. The NIPS framework comprises a data preprocessing unit, a construction classifier unit, an expert system unit and a knowledge base, wherein the data preprocessing unit is used for collecting network data streams and selecting an input sample and a test sample for the neural network from the collected network data streams; the construction classifier unit is used for making use of an input and learning sample MA-LVQ (Mobile Agent-Learning Vector Quantization) neural network classifier and performing class test to form a knowledge base; the expert system unit is used for interacting with the knowledge base according to a known security policy to compare and classify actions provided by the data streams and action descriptions in the knowledge base so as to determine an output result; and the knowledge base comprises a normal action description and an abnormal action description and is updated by interacting through the expert system unit. By adopting the NIPS framework, a better classifying effect can be achieved by a linear network, and the stronger limit on linear separability of data required by the linear network can be avoided effectively under the action of a competition layer; and the NIPS framework is more practicable and extensive.

The invention discloses a neural network data compression and related calculation method and device. The data compression method comprises the following steps: acquiring a fixed-point neural network model; and replacing specific numbers continuously appearing in the neural network model calculation with a combination of two or more numbers. Therefore, for the inherent attribute that a large numberof identical numbers continuously appear in quantized neural network calculation, a fixed number combination can be adopted for replacement, so that the effect of highly compressing the data is conveniently and simply realized.

Some embodiments of the invention provide a novel method for training a quantized machine-trained network. Some embodiments provide a method of scaling a feature map of a pre-trained floating-point neural network in order to match the range of output values provided by quantized activations in a quantized neural network. A quantization function is modified, in some embodiments, to be differentiable to fix the mismatch between the loss function computed in forward propagation and the loss gradient used in backward propagation. Variational information bottleneck, in some embodiments, is incorporated to train the network to be insensitive to multiplicative noise applied to each channel. In some embodiments, channels that finish training with large noise, for example, exceeding 100%, are pruned.

The invention provides a method and a device for adjusting and deploying a face recognition neural network. The face recognition neural network at least comprises a plurality of convolution layers andat least one full connection layer, the last full connection layer is a classifier for classification, and the method comprises the following steps: obtaining a neural network model to be trained; training the neural network model by using fixed-point quantization to obtain a trained fixed-point quantization neural network model in which the last full connection layer maintains a floating point in a training process; and outputting the trained fixed-point quantized neural network model without the last full connection layer. Therefore, by utilizing the particularity of the face recognition network, the classifier layer unfixed points which greatly influence the overall precision of the network are kept in the training stage and are not input into the network any more, and the classifier layer is not included in the network any more, so that the trained fixed-point neural network can be ensured to have high precision, and extra computing power requirements during network deployment canbe avoided at the same time.

The invention provides a neural network quantification method and device, and an electronic device. The method comprises the steps of: in the iterative training process of a neural network, utilizingscaling factors of all neurons in an input layer, performing quantitative calculation on initial activation values of all the neurons in the input layer in all output channels of the input layer, andoutputting activation values of all the neurons in a next hidden layer of the input layer; taking each hidden layer of the neural network as a current layer one by one; and executing the following quantization operation on each current layer: executing the following quantization operation on each current layer, determining a scaling factor of each neuron in the current layer based on the activation value of each neuron in the current layer, performing quantitative calculation on the activation value of each neuron in the current layer in each output channel of the current layer by utilizing the scaling factor of each neuron in the current layer, and outputting the activation value of each neuron in the next layer of the current layer; and when the iterative training is completed, taking the current neural network as a quantized neural network. The recognition precision of the neural network is improved.

The application relates to the technical field of deep learning, and provides a model construction method and device, an image processing method and device, a hardware platform and a storage medium. The model construction method comprises the steps that a neural network model used for image processing is trained, wherein the neural network model comprises at least one depth separable convolution module, and each depth separable convolution module comprises a layer-by-layer convolution layer, a point-by-point convolution layer, a batch normalization layer and an activation layer which are connected in sequence; and the trained neural network model is quantized to obtain a quantized neural network model. According to the method, firstly, model parameters are quantified, so that the data volume of the parameters is effectively reduced, and the model is suitable for being deployed in NPU equipment. Secondly, the depth separable convolution module in the method is different from the depth separable convolution module in the prior art, and a batch normalization layer and an activation layer are not arranged between a layer-by-layer convolution layer and a point-by-point convolution layer, so that values of model parameters are distributed in a reasonable range, and the model parameters can be quantized with high precision.

The embodiment of the invention provides a training and testing method, device and equipment for a quantitative neural network model. The training method comprises the following steps: acquiring training data, wherein the training data comprises an image sample and a labe, and the label comprises an identity label of the image sample; inputting the training data into a quantitative neural networkmodel for quantitative processing to obtain quantitative features of the image sample; and training the quantitative neural network model according to the quantitative features to obtain a trained quantitative neural network model. By implementing the embodiment of the invention, the trained quantitative neural network model adopts low-bit quantization processing, the operand is reduced, the operation rate is improved, the method can be more conveniently applied to terminal equipment, the quantitative neural network model is used for face recognition, the quantization error can be minimized, and the recognition precision of the quantitative neural network model is improved.

The invention discloses a structured pruning method for a lightweight neural network. The method comprises the following steps: constructing a similarity matrix according to a three-dimensional tensorFilter in a to-be-pruned target layer in the neural network; clustering the three-dimensional tensor Filter corresponding to the similarity matrix based on spectral clustering to obtain a plurality of clustering clusters; determining the centroid of each clustering cluster and the distance between each three-dimensional tensor Filter in the same clustering cluster and the centroid; and deleting the three-dimensional tensor Filter corresponding to the centroid distance exceeding a set threshold to obtain a target neural network model. As a structured pruning method, the weight matrixes of theneural network subjected to structured pruning do not have an unstructured sparse phenomenon, existing software and hardware can be directly used for acceleration, and the invention can be very naturally compared with other lightweight neural network technologies. Knowledge distillation, weight quantification and the like are combined for use to further reduce the network redundancy.

According to a method and apparatus for neural network quantization, a quantized neural network is generated by performing learning of a neural network, obtaining weight differences between an initial weight and an updated weight determined by the learning of each cycle for each of layers in the first neural network, analyzing a statistic of the weight differences for each of the layers, determining one or more layers, from among the layers, to be quantized with a lower-bit precision based on the analyzed statistic, and generating a second neural network by quantizing the determined one or more layers with the lower-bit precision.

The invention provides a method and a system for switching bit-wide quantized neural network on line, and the method comprises the steps: enabling deep neural networks with different bit widths to beintegrated into a super network, wherein all the networks with different bit widths share the same network architecture; the super network operating at different bit widths, for any bit width, obtaining a corresponding network intermediate layer feature, and processing each network intermediate layer feature by adopting a corresponding batch normalization layer; training the super network throughsupervised learning, and simulating quantization noise in a super-network training stage until a consistency loss function between a low-bit mode and a high-bit mode converges, so as to obtain a trained super network; and extracting the quantized neural network of the target bit from the trained super network by using a preset quantizer to perform low-bit reasoning. While the neural network is nottrained again, the bit width can be switched at will to adapt to different hardware deployment environments.

The invention provides a neural network model quantification method and device, and electronic equipment, and relates to the technical field of machine learning, and the method comprises the steps: obtaining a neural network model, wherein the neural network model comprises a convolution layer, a normalization layer and a quantization activation layer, and the output feature of the quantization activation layer is an integer type feature; converting the parameters of the convolution layer into integer type parameters; combining the normalization layer and the quantization activation layer to obtain a combined layer; and converting the parameters of the merging layer into integer type parameters to obtain a quantized neural network model. According to the method, the operation complexity ofthe neural network model is reduced, so that the quantized neural network model can be accelerated to run on target equipment only supporting fixed-point or low-bit-width operation.

The invention provides an image target detection method and system based on a lightweight neural network model, belongs to the technical field of image target detection, solves the problem of inaccurate image recognition at present, and comprises the following steps: inputting a path of a to-be-detected picture or video; and calculating related confidence coefficients of all classifications in the received current image by using the lightweight neural network model, obtaining a final recognition frame by selecting the highest confidence coefficient, and drawing the recognition frame in the original image to complete the detection process. Under the condition that the model precision is ensured, the operation speed of the model is greatly improved, so that the model can be smoothly deployed and applied in small equipment and a mobile terminal, and the real-time performance and accuracy of smoking detection in a daily scene are met.

The embodiment of the invention provides a neural network quantization method and a system, and the method comprises the steps: (1) obtaining neural network parameters, carrying out the fusion of a convolution layer and a BN layer in the forward propagation of an original network, so as to obtain a new to-be-quantized weight and a to-be-quantized bias, and storing the activation output data of each layer at the same time; (2) quantizing neural network parameters and data distribution obtained after fusion; and (3) forming a new quantization network by taking the quantization parameter obtainedafter quantization processing as a parameter of forward propagation of the quantized new network. According to the method, various problems caused by too large or too small quantization scaling scaleof the neural network can be effectively balanced, and the operation speed is increased while hardware resources are saved.