Back to results
US 12,725,029 B2Granted
Information processing apparatus, information processing method, and storage medium
AssigneeCanon Kabushiki Kaisha
InventorTakayuki Komatsu
Life of this patent
Estimating the term…
Abstract
An information processing apparatus designates one or more constraints for constraining a configuration of a neural network, which include information for specifying the configuration of the neural network and the maximum number of computations in the neural network, and executes a computation of a neural network configured based on the designated constraints. When a neural network is trained, the information processing apparatus trains respective neural networks under the one or more designated constraints, and selects, for estimation, a predetermined learned model out of the learned models trained under the one or more designated constraints.
Claims13
An information processing apparatus comprising:
- one or more processors;
- and
- a memory storing instructions which, when the instructions are executed by the one or more processors, cause the information processing apparatus to function as:
- a designation unit configured to designate one or more constraints for constraining a configuration of a neural network, the constraints including information for specifying a neural network architecture and the maximum number of computations in the neural network, wherein the designation unit is further configured to iteratively compare an actually measured number of product-sum computations (PSC) occurring during training with the maximum number of computations and, when the measured PSC exceeds the maximum, automatically adapt the neural network architecture by reducing at least one of a layer count, a node count or a kernel size and restarting a training cycle until the measured PSC becomes equal to or smaller than the maximum number of computations;
- and
- a processing unit configured to execute a computation of a neural network configured based on the designated constraints,
- wherein the processing unit is configured to:
- i) train a plurality of neural networks under the one or more designated constraints;
- ii) during each training cycle, monitor an accumulated count of PSC executed by the neural network and, when the accumulated count exceeds the maximum number of computations, notify the designation unit to trigger the automatic adaptation;
- and
- iii) output, for every trained neural network, an evaluation value indicative of a degree of optimization of a corresponding learned model,
- wherein after completion of training under the one or more designated constraints, the processing unit selects, for estimation, as a predetermined learned model, a learned model satisfying both of (1) the evaluation value being greater than or equal to a preset threshold, and (2) a PSC count that is the smallest among the learned models satisfying (1),
- wherein the maximum number of computations indicates the number of product-sum computations required when a computation of the neural network is executed, and
- wherein the one or more processors uses the selected learned model to perform learning and estimation processes.
11 more elements
An information processing method comprising:
- designating one or more constraints for constraining a configuration of a neural network, the constraints including information for specifying a neural network architecture and the maximum number of computations in the neural network, wherein constraining the configuration includes iteratively comparing an actually measured number of product-sum computations (PSC) occurring during training with the maximum number of computations and, when the measured PSC exceeds the maximum, automatically adapting the neural network architecture by reducing at least one of a layer count, a node count or a kernel size and restarting a training cycle until the measured PSC becomes equal to or smaller than the maximum number of computations;
- and
- processing for executing a computation of a neural network configured based on the one or more designated constraints,
- wherein, the processing comprises;
- i) training a plurality of neural networks under the one or more designated constraints;
- ii) during each training cycle, monitoring an accumulated count of PSC executed by the neural network and, when the accumulated count exceeds the maximum number of computations, notifying the designation unit to trigger the automatic adaptation;
- and
- iii) outputting, for every trained neural network, an evaluation value indicative of a degree of optimization of a corresponding learned model,
- wherein after completion of training under the one or more designated constraints, selecting, for estimation, as a predetermined learned model, a learned model satisfying both of (1) the evaluation value being greater than or equal to a preset threshold, and (2) a PSC count that is minimal among the learned models satisfying (1),
- wherein the maximum number of computations indicates the number of product-sum computations required when a computation of the neural network is executed,
- and
- wherein the selected learned model is used to perform learning and estimation processes.
9 more elements
A non-transitory computer-readable storage medium comprising instructions for performing an information processing method, the information processing method comprising:
- designating one or more constraints for constraining a configuration of a neural network, the constraints including information for specifying a neural network architecture and the maximum number of computations in the neural network, wherein constraining the configuration includes iteratively comparing an actually measured number of product-sum computations (PSC) occurring during training with the maximum number of computations and, when the measured PSC exceeds the maximum, automatically adapting the neural network architecture by reducing at least one of a layer count, a node count or a kernel size and restarting a training cycle until the measured PSC becomes equal to or smaller than the maximum number of computations;
- and
- processing for executing a computation of a neural network configured based on the one or more designated constraints,
- wherein, processing comprises:
- i) training a plurality of neural networks under the one or more designated constraints;
- ii) ii) during each training cycle, monitoring an accumulated count of PSC executed by the neural network and, when the accumulated count exceeds the maximum number of computations, notifying the designation unit to trigger the automatic adaptation;
- and
- iii) outputting, for every trained neural network, an evaluation value indicative of a degree of optimization of a corresponding learned model,
- wherein after completion of training under the one or more designated constraints, selecting, for estimation, as a predetermined learned model, a learned model satisfying both of (1) the evaluation value being greater than or equal to a preset threshold, and (2) a PSC count that is minimal among the learned models satisfying (1),
- wherein the maximum number of computations indicates the number of product-sum computations required when a computation of the neural network is executed,
- and
- wherein the selected learned model is used to perform learning and estimation processes.
9 more elements
Dependencies are inferred from the claim text.