Method of structured network pruning and sparsity speed-up

1. A computer implemented method for pruning a convolutional neural network during training for image recognition, the method comprising:

storing the convolutional neural network in a memory device, wherein the convolutional neural network is used for recognizing an object in an image;

determining, by a processor; a number of channels (N) between an input and an output of a respective residual block of the convolutional neural network, wherein a respective channel corresponds to a filter in the convolutional neural network, wherein at least one channel is a color filter which represents intensity of a single color in the image, and wherein the filter is stored as a matrix in the memory device;

constructing at least one lookup table, wherein the lookup table indicates whether a channel is needed for a path between the input and output of the residual block of the convolutional neural network;

storing the at least one lookup table in the memory device; and

pruning filters in the convolutional neural network to create a shortcut between the input and the output based on the at least one lookup table; thereby facilitating reduction of size of the convolutional neural network and corresponding reduction of required storage space and computation cost associated with deployment of the convolutional neural network.

2. The method of pruning the convolutional neural network of claim 1 further comprising:

merging the channels between the input and the output;

categorizing N channel data flows; and

pruning filters by categorized N channel data flows.

3. The method of pruning the convolutional neural network of claim 2 , wherein the categorization of N channel data flows is one of “only use by res,” “add to res,” and “bypass;” and wherein if the channel data flow is categorized as “only use by res,” the method further comprises removing the channel data flow after calculation based on the N lookup tables.

4. The method of pruning the convolutional neural network of claim 1 further comprising:

determining a plurality of groupings of filter elements within a set of related filters;

determining a weight of each of the plurality of groupings of filter elements;

determining a mask based on the weight; and

pruning the plurality of groupings of filter elements based on the mask in each of the set of related filters.

5. The method of pruning the convolutional neural network of claim 1 further comprising:

determining a number of parallel process channels (PPC);

determining the number of channels (N) between the input and the output; and

pruning the number of channels (N) to be an integer multiple of the number of parallel process channels (PPC).

6. The method of pruning the convolutional neural network of claim 5 wherein the determining of the number of parallel process channels (PPC) is based on a hardware specific number of mac arrays which may be processed in parallel.

7. The method of pruning the convolutional neural network of claim 6 wherein the pruning of the number of channels is performed during a training process.

8. A computer implemented method for pruning a convolutional neural network during training for image recognition, the method comprising:

storing the convolutional neural network in a memory device, wherein the convolutional neural network is used for recognizing an object in an image;

determining, by a processor, a number of channels (N) between an input and an output of a respective residual block of the convolutional neural network, wherein a respective channel corresponds to a filter in the convolutional neural network, wherein at least one channel is a color filter which represents intensity of a single color in the image, and wherein the filter is stored as a matrix in the memory device;

constructing N lookup tables, each lookup table matched to a respective channel;

storing the N lookup tables in the memory device;

merging the channels between the input and the output;

categorizing N channel data flows; and

pruning filters by categorized N channel data flows; thereby facilitating reduction of size of the convolutional neural network and corresponding reduction of required storage space and computation cost associated with deployment of the convolutional neural network.

9. The method of pruning the convolutional neural network of claim 8 , wherein the categorization of the N channel data flows is one of “only use by res,” “add to res” and “bypass;” and if the channel data flow is categorized as “only use by res,” the method further comprises removing the channel data flow after calculation based on the N lookup tables.

10. The method of pruning the convolutional neural network of claim 8 further comprising pruning filters in the convolutional neural network to create a shortcut between the input and the output based on the N lookup tables.

11. The method of pruning the convolutional neural network of claim 10 further comprising:

determining a plurality of groupings of filter elements within a set of related filters;

determining a weight of each of the plurality of groupings of filter elements;

determining a mask based on the weight; and

pruning the plurality of groupings of filter elements based on the mask in each of the set of related filters.

12. The method of pruning the convolutional neural network of claim 11 further comprising:

determining a number of parallel process channels (PPC); and

pruning the number of channels (N) to be an integer multiple of the number of parallel process channels (PPC).

13. A computer implemented method for pruning a convolutional neural network during training for image recognition, the method comprising:

storing the convolutional neural network in a memory device, wherein the convolutional neural network is used for recognizing an object in an image;

determining a plurality of groupings of filter elements within a set of related filters, wherein at least one filter is a color filter which represents intensity of a single color in the image, and wherein the filter is stored as a matrix in the memory device;

determining a weight of each of the plurality of groupings of filter elements;

determining a mask based on the weight; and

pruning the plurality of groupings of filter elements based on the mask in each of the set of related filters;

determining a number of channels (N) between an input and an output of a respective residual block of the convolutional neural network, wherein a respective channel corresponds to a filter in the convolutional neural network, wherein at least one channel is a color filter which represents intensity of a single color in the image, and wherein the filter is stored as a matrix in the memory device;

constructing N lookup tables, each lookup table matched to a respective channel, and wherein a respective lookup table indicates whether a channel is needed for a path between the input and output of the residual block of the convolutional neural network;

storing the lookup tables in the memory device; and

pruning filters in the convolutional neural network to create a shortcut between the input and the output based on the N lookup tables; thereby facilitating reduction of size of the convolutional neural network and corresponding reduction of required storage space and computation cost associated with deployment of the convolutional neural network.

14. The method of pruning the convolutional neural network of claim 13 , further comprising:

merging the channels between the input and the output;

categorizing N channel data flows; and

pruning filters by categorized N channel data flows.

15. A computer implemented method for pruning a convolutional neural network during training for image recognition, the method comprising:

storing the convolutional neural network in a memory device, wherein the convolutional neural network is used for recognizing an object in an image;

determining, by a processer, a number of parallel process channels (PPC);

determining a number of channels (N) between an input and an output of the residential block of the convolutional neural network wherein a respective channel corresponds to a filter in the convolutional neural network, wherein at least one channel is a color filter which represents intensity of a single color in the image, and wherein the filter is stored as a matrix in the memory device; and

pruning the number of channels (N) to be an integer multiple of the number of parallel process channels (PPC);

wherein the determining of the number of parallel process elements (PPC) is based on a hardware specific number of parallel process channels (PPC) is based on hardware specific number of mac arrays which may be processed in parallel, thereby facilitating reduction of size of the convolutional neural network and corresponding reduction of required storage space and computation cost associated with deployment of the convolutional neural network.

16. The method of pruning the convolutional neural network of claim 15 further comprising:

determining a plurality of groupings of filter elements within a set of related filters;

determining a weight of each of the plurality of groups o filter elements;

determining a mask based on the weight; and

pruning the plurality of groupings of filter elements based on the mask in each of the set of related filters.

17. The method of pruning the convolutional neural network of claim 16 further comprising:

constructing N lookup tables, each lookup table matched to a respective channel; and

pruning filters in the convolutional neural network to create a shortcut between the network input and the network output based on the N lookup tables.