Master transform architecture for deep learning

Patent metadata

Field	Value
Publication number	US-2021192287-A1
Application number	US-201916719883-A
Country	US
Kind code	A1
Filing date	Dec 18, 2019
Priority date	Dec 18, 2019
Publication date	Jun 24, 2021
Grant date	—

Abstract

Apparatuses, systems, and techniques to transform input data for training neural networks. In at least one embodiment, one or more data transforms are identified in a sequence of data transforms and combined into one or more master data transforms to be performed by one or more parallel processing units in order to prepare data for training an untrained neural network.

First claim

What is claimed is:

1 . A processor, comprising:
one or more circuits to perform a data transform comprising a combination of two or more data transforms, wherein the two or more data transforms are to be combined based, at least in part, on input and output data sizes of the two or more data transforms.

2 . The processor of claim 1 , wherein the data transform is performed by one or more parallel processing units, and the combination of two or more data transforms is based, at least in part, on a profile of resource requirements for each of the two or more data transforms.

3 . The processor of claim 1 , wherein the combination of two or more data transforms results in a sequence of instructions implementing operations on data to train one or more neural networks.

4 . The processor of claim 3 , wherein the sequence of instructions implement operations to be performed by one or more parallel processing units.

5 . The processor of claim 1 , wherein the two or more data transforms are to be combined, based at least in part, on memory requirements of each of the two or more data transforms and memory availability of one or more parallel processing units.

6 . The processor of claim 1 , wherein the two or more data transforms are to be combined, based at least in part, on compute time requirements of each of the two or more data transforms.

7 . The processor of claim 1 , wherein the two or more data transforms are to be combined based, at least in part, on available memory resources of a computing system.

8 . The processor of claim 1 , wherein the two or more data transforms are pre and post transforms, and prepare 3-dimensional image data for use in training a neural network.

9 . A system, comprising:
one or more processors to perform a first set of two or more data transforms and a second set of two or more data transforms, wherein the second set of two or more data transforms are to be combined from individual data transforms from the first set of two or more data transforms based, at least in part, on input and output data sizes of the individual data transforms.

10 . The system of claim 9 , wherein the second set is performed by one or more parallel processing units, and the combination of individual data transforms from the first set is based, at least in part, on resource requirements for each of the two or more data transforms.

11 . The system of claim 9 , wherein the second set performs a sequence of operations on three dimensional (3D) image data.

12 . The system of claim 11 , wherein the second set is accelerated by one or more parallel processing units.

13 . The system of claim 9 , wherein the individual data transforms from the first set are to be combined, based at least in part, on memory requirements of each of the individual data transforms and memory availability of one or more parallel processing units.

14 . The system of claim 9 , wherein the individual data transforms from the first set are to be combined such that a time requirement for applying the first set is reduced.

15 . The system of claim 9 , wherein the individual data transforms are to be combined based on available memory resources of a computing system implementing one or more neural networks.

16 . The system of claim 15 , wherein the one or more neural networks are trained using data transformed by the second set.

17 . The system of claim 9 , wherein the first set of two or more data transforms and the second set of two or more data transforms contain pre and post transforms.

18 . The system of claim 9 , wherein the first set of two or more data transforms and the second set of two or more data transforms prepare three dimensional (3D) image data for use in training one or more neural networks.

19 . The system of claim 9 , wherein:
the second set of two or more data transforms are performed; a third set of data transforms are performed, and the third set of data transforms is comprised of individual data transforms from the first set of two or more data transforms that were not selected to be in the second set of two or more data transforms.

20 . A machine-readable medium having stored thereon a set of instructions, which if performed by one or more processors, cause the one or more processors to at least:
perform a data transform comprising a combination of two or more data transforms, wherein the two or more data transforms are to be combined based, at least in part, on input and output data sizes of the two or more data transforms.

21 . The machine-readable medium of claim 20 , wherein the data transform comprising a combination of two or more data transforms is performed by one or more parallel processing units.

22 . The machine-readable medium of claim 20 , wherein the instructions, when performed, further cause the one or more processors to combine two or more data transforms based, at least in part, on a profile of resource requirements for each of the two or more data transforms.

23 . The machine-readable medium of claim 20 , wherein instructions, when performed, further cause the one or more processors to perform a sequence of data transformation operations on data used to train one or more neural networks, where the sequence of data transformation operations are specified by the combination of two or more data transforms.

24 . The machine-readable medium of claim 23 , wherein the sequence of data transformation operations are accelerated by one or more graphics processing units.

25 . The machine-readable medium of claim 20 , wherein the two or more data transforms are to be combined, based at least in part, on memory requirements of each of the two or more data transforms and memory availability of one or more parallel processing units.

26 . The machine-readable medium of claim 20 , wherein the two or more data transforms are to be combined such that a required computing time to perform each of the two or more data transforms is reduced.

27 . The machine-readable medium of claim 20 , wherein the two or more data transforms are pre and post transforms, and prepare three dimensional (3D) image data for use in training a neural network.

28 . A method, comprising:
performing a first set of two or more data transforms using one or more parallel processing units, wherein the first set of two or more data transforms are based, at least in part, on individual data transforms from a second set of two or more data transforms; and selecting individual data transforms for the first set of two or more data transforms from the second set of two or more data transforms based, at least in part, on input and output data sizes of the individual data transforms.

29 . The method of claim 28 , wherein the second set of two or more data transforms is performed by one or more parallel processing units.

30 . The method of claim 28 , wherein the combination of individual data transforms from the first set is based, at least in part, on resource requirements for each of the two or more data transforms in the first set.

31 . The method of claim 28 , wherein the second set of two or more data transforms performs a sequence of operations on three dimensional (3D) image data.

32 . The method of claim 28 , wherein the first set of two or more data transforms

Assignees

• Nvidia Corp

Inventors

• Dwivedi Shekhar
• Haemel Nicholas Alexander

Classifications

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title

• G06F18/2155

  characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title