Method and electronic apparatus for adjusting a neural network
AssigneeHTC Corporation
InventorsYu-Hsun Lin, Chun-Nan Chou, Edward Chang
Life of this patent
Estimating the term…
Abstract
A method for adjusting a convolutional neural network includes following operations. The convolutional neural network includes convolution layers in a sequential order. Receptive field widths of the convolution layers in a first model of the convolutional neural network are determined. Channel widths of the convolution layers in the first model are reduced into reduced channel widths according to the receptive field widths of the convolution layers and an input image width. A structure of a second model of the convolutional neural network is formed according to the reduced channel widths. The second model of the convolutional neural network is trained according the structure of the second model.
Claims20
A method for adjusting a convolutional neural network, a first model of the convolutional neural network comprising a plurality of convolution layers in a sequential order, the adjustment method comprising:
- determining a plurality of receptive field widths of the convolution layers in the first model of the convolutional neural network, wherein the receptive field widths of the convolution layers are determined by calculating widths of projective regions on an input image affecting one feature point in convolution output tensors of the convolution layers;
- reducing a plurality of channel widths of the convolution layers in the first model into a plurality of reduced channel widths according to the receptive field widths of the convolution layers and an input image width;
- forming a structure of a second model of the convolutional neural network according to the reduced channel widths;
- and
- training the second model of the convolutional neural network.
An electronic apparatus, suitable for adjusting a convolution neural network, the electronic apparatus comprising:
- a data storage, configured to store a first model of the convolution neural network, the first model of the convolution neural network comprising a plurality of convolution layers;
- and
- a processor, coupled with the data storage, the processor being configured to:
- determine a plurality of receptive field widths of the convolution layers in the first model of the convolutional neural network, wherein the receptive field widths of the convolution layers are determined by calculating widths of projective regions on an input image affecting one feature point in convolution output tensors of the convolution layers;
- reducing a plurality of channel widths of the convolution layers in the first model into a plurality of reduced channel widths according to the receptive field widths of the convolution layers and an input image width;
- form a structure of a second model of the convolutional neural network according to the reduced channel widths;
- and
- train the second model of the convolutional neural network.
Dependencies are inferred from the claim text.