Computing device and method
A computing device, comprising: a computing module, comprising one or more computing units; and a control module, comprising a computing control unit, and used for controlling shutdown of the computing unit of the computing module according to a determining condition. Also provided is a computing method. The computing device and method have the advantages of low power consumption and high flexibility, and can be combined with the upgrading mode of software, thereby further increasing the computing speed, reducing the computing amount, and reducing the computing power consumption of an accelerator.
1. An operation device comprising:
a transformation circuit configured to perform spatial transformation on input data and/or a parameter matrix from first geometric space into second geometric space; and
an operation circuit connected to the transformation circuit and configured to receive transformed input data and parameter matrix, and then perform operations,
wherein the operation device further includes an optimization circuit configured to perform optimization operation on transformed input data and/or a parameter matrix, wherein the optimization operation includes component selection, quantifying, sparsification, or compressing,
wherein the optimization circuit performs component selection according to a component selection condition through a principal component analysis, and
wherein the component selection condition includes:
a threshold condition that with a given threshold, if an absolute value of a coefficient of a component is greater than the threshold, the component is retained, otherwise, if an absolute value of a coefficient of a component is less than the threshold, the component is deleted, and
a quantity condition that with a given quantity for selection, components that satisfy the quantity are selected according to a descending order of coefficients of the components.
2. The operation device of claim 1 , wherein the input data and the parameter matrix are presented by a linear combination of a basis vector of the second geometric space through spatial transformation.
3. The operation device of claim 1 , wherein the input data and the parameter matrix are input data and a parameter matrix used by a convolutional layer, a down-sampling layer, a normalization layer, or a regularization layer.
4. The operation device of claim 1 ,
wherein the first geometric space is a spatial domain, and the second geometric space is a frequency domain, and
wherein a manner of the spatial transformation is an invertible spatial transformation including FFT, DFT, DCT, or DST.
5. The operation device of claim 1 ,
wherein the operation circuit includes:
a multiplier configured to multiply data that are input into the multiplier to obtain output after multiplication, and
one or a plurality of adders configured to add data that are input into the adders to obtain output data, and
wherein the operation circuit further includes an activation function circuit configured to perform activation function operations on data that are input into the activation function circuit to obtain activated output data, wherein an activation function includes sigmoid, tanh, ReLU, or softmax.
6. The operation device of claim 1 ,
wherein the operation circuit further includes a pooling circuit configured to perform pooling operations on data that are input into the pooling circuit to obtain output data after pooling, wherein the pooling operations include: mean-pooling, max-pooling, or median-pooling, and wherein the operation device further includes:
a storage circuit configured to store data and instructions, and
a control circuit configured to control the operation circuit to perform corresponding operations.
7. An operation method comprising:
performing spatial transformation on input data and/or a parameter matrix from first geometric space to second geometric space; and
performing operations on transformed input data and parameter matrix,
wherein prior to the performing operations on the transformed input data and parameter matrix, the method further includes:
performing optimization operations on the input data and/or the parameter matrix in transformation space, wherein the optimization operations include at least one of component selection, quantifying, sparsification, or compressing,
wherein the optimization operations refer to performing component selection according to a component selection condition through a principal component analysis, and
wherein the component selection condition includes:
a threshold condition that with a given threshold, if an absolute value of a coefficient of a component is greater than the threshold, the component is retained, otherwise, if an absolute value of a coefficient of a component is less than the threshold, the component is deleted, and
a quantity condition that with a given quantity for selection, components that satisfy the quantity are selected according to a descending order of coefficients of the components.
8. The operation method of claim 7 , wherein the input data or the parameter matrix are presented by a linear combination of a basis vector of the second geometric space through spatial transformation.
9. The operation method of claim 7 , wherein the input data and the parameter matrix are input data and a parameter matrix used by a convolutional layer, a down-sampling layer, a normalization layer, or a regularization layer.
10. The operation method of claim 7 , wherein the first geometric space is a spatial domain, and the second geometric space is a frequency domain, and
wherein a manner of the spatial transformation is an invertible spatial transformation including FFT, DFT, DCT, or DST.
11. The operation method of claim 7 , wherein the performing operations on the transformed input data and parameter matrix includes:
obtaining a linear combination representation of the parameter matrix in the second geometric space according to the obtained transformed parameter matrix, wherein the linear combination representation indicates that the parameter matrix is represented by a linear combination of a space basis vector of the second geometric space, and
coupling and operating the spatially transformed input data and the space basis vector.
12. The operation method of claim 7 , wherein the performing operations on the transformed input data and parameter matrix includes: multiplication, addition, multiply-add operations, inner product operations, non-linear operations, or convolution operations.
13. The operation method of claim 7 , wherein the performing operations on the transformed input data and parameter matrix further includes: activation function operations or pooling operations.
14. An electronic device, comprising:
an operation device that includes:
a transformation circuit configured to perform spatial transformation on input data and/or a parameter matrix from first geometric space into second geometric space; and
an operation circuit connected to the transformation circuit and configured to receive transformed input data and parameter matrix, and then perform operations,
wherein the operation device further includes an optimization circuit configured to perform optimization operation on transformed input data and/or a parameter matrix, wherein the optimization operation includes component selection, quantifying, sparsification, or compressing,
wherein the optimization circuit performs component selection according to a component selection condition through a principal component analysis, and
wherein the component selection condition includes:
a threshold condition that with a given threshold, if an absolute value of a coefficient of a component is greater than the threshold, the component is retained, otherwise, if an absolute value of a coefficient of a component is less than the threshold, the component is deleted, and
a quantity condition that with a given quantity for selection, components that satisfy the quantity are selected according to a descending order of coefficients of the components.
15. The electronic device of claim 14 , wherein the operation device further includes:
a multiplier configured to multiply data that are input into the multiplier to obtain output after multiplication, and
one or a plurality of adders configured to add data that are input into the adders to obtain output data.
16. The electronic device of claim 15 , wherein the operation circuit further includes an activation function circuit configured to perform activation function operations on data that are input into the activation function circuit to obtain activated output data, wherein an activation function includes sigmoid, tanh, ReLU, or softmax.