Patent Application
20230064481
Machine learning (ML) is becoming an increasingly important part of the computing landscape. Machine learning is a branch of artificial intelligence (AI), and ML helps enable a software system to learn to recognize patterns from data without being directly programmed to do so. Neural networks (NN) are a type of ML which utilize a set of linked and layered functions (e.g., nodes, neurons, etc.) which are weighted to evaluate input data. In some NNs, sometimes referred to as convolution NNs (CNNs), convolution operations are performed in NN layers based on inputs received and weights rather than matrix multiplication used in traditional NN. Layers in CNNs may perform many types of functions, including, but not limited to, convolution, deconvolutional, pooling, up-sample, etc. CNNs are often used in a wide array of applications typically for recognition and classification, such as image recognition and classification, prediction and recommendation systems, speech and language recognition and translation, etc.
As ML becomes increasingly useful, there is a desire to execute complex ML techniques, such as NNs and CNNs, efficiently in devices with relatively limited compute and memory resources, such as embedded, or other low-power devices, To help efficiently run a given ML model, the ML model may be analyzed and optimized to tailor how the ML model is run to a target hardware resources to be used.
This disclosure relates to an electronic device, comprising one or more processors. The one or more processors are configured to execute instructions causing the one or more processors to: receive a machine learning (ML) model and execution information associated with the ML model, wherein the execution information including first execution data indicating how to execute the ML model optimized based on a first performance criterion, and second execution data execution data indicating how to execute the ML model optimized based on a second performance criteria, the second performance criterion different from the first performance criteria. The processors are further configured to execute instructions causing the one or more processors to execute the ML model based on the first execution data. The processors are further configured to execute instructions causing the one or more processors to determine to execute the ML model based on the second execution data and execute the ML model based on the second execution data.
Another aspect of the present disclosure relates to a method. The method includes receiving a machine learning (ML) model and execution information associated with the ML model, wherein the execution information including first execution data indicating how to execute the ML model optimized based on a first performance criterion, and second execution data execution data indicating how to execute the ML model optimized based on a second performance criteria, the second performance criterion different from the first performance criteria. The method further includes executing the ML model based on the first execution data. The method further includes determining to execute the ML model based on the second execution data. The method further includes executing the ML model based on the second execution data.
Another aspect of the present disclosure relates to a non-transitory program storage device comprising instructions stored thereon to cause one or more processors to receive a machine learning (ML) model. The instructions further cause the one or more processors to generate first execution data for executing the ML model on target hardware, the first execution data indicating how to execute the ML model on the target hardware optimized based on a first performance criterion. The instructions further cause the one or more processors to generate a second execution data for executing the ML model on target hardware, the second execution data indicating how to execute the ML model on the target hardware optimized based on a second performance criteria, the second performance criteria different from the first performance criterion. The instructions further cause the one or more processors to output the first execution data and second execution data for executing the ML model
For a detailed description of various examples, reference will now be made to the accompanying drawings in which:
As ML has becoming more common and powerful, hardware, such as embedded device, have configured to execute ML models has been introduced. As used herein, an ML model may refer to an implementation of one or more ML algorithms which model a behavior, such as object recognition, behavior of a circuit, behavior of a neuron, etc. In cases where a target hardware for executing ML models is known, the ML models may be optimized for the target hardware configurations to help enhance performance. For example, ML models for object recognition, low-light enhancement, and facial recognition may be optimized to execute on a particular a mobile device, such as a smartphone configured with a certain ML processor. As another example, ML models for object recognition, movement prediction, and behavioral prediction may be optimized to execute on specific hardware (e.g., target hardware) found in some partially or fully self-driving automobiles. In some cases, a ML model may be optimized in multiple times for multiple, different, aspects for the target hardware. During execution of the ML model on the target hardware, the execution of the ML model may be adjusted based on the different optimizations.
Example ML Model
In this example, first layer 106 represents a function based on a set of weights that are applied to the input parameters (e.g., input parameters 102 and 104) to generate output from first layer 106 that is input to the second layer 108. Different weights may be applied for the input received from each node of the previous layer by the subsequent layer. For example, for a node of the second layer 108, the node applies weights to input received from nodes of the first layer 106 and the node may apply a different weight to input received from each node of the first layer 106. Nodes compute one or more functions based on the inputs received and corresponding weights and outputs a number. In some cases, inputs and output to an ML model layer may be referred to as input or output features of the ML model layer. For example, the node may use a linear combination function which multiplies an input values from a node of the previous layer with a corresponding weight and sums across the results of the multiplication, coupled with a non-linear activation function which acts as a floor for the resulting number for output. It may be understood that any known weighted function may be applied by the node within the scope of this disclosure. This output number may be input to subsequent layers, or if the layer is a final layer, such as third layer 110 in this example, the number may be output as a result (e.g., output parameters or ML model outputs 112).
In some cases, the functions applied by nodes of a layer may differ as between layers. In some cases, each layer may have different resource requirements. For example, when the functions of multiple nodes are performed by a processor, the different functions may have different loads on the processor. Additionally, some functions may have different input or output parameters and thus consume more, or less, memory space and bandwidth. These differing processor and memory loads may also influence an amount of energy to power the processor and memory, as well as an amount of heat generated.
After an ML model, such as NN ML model 100, is defined with respect to nodes, layers, etc., the ML model may be trained. In some cases, the ML model 100 may be trained using a labelled data set corresponding to data to be input to ML model 100. For example, an object recognizer may be trained on images of objects. These images may include metadata labelling the object(s) in the image. The ML model 100 may be initiated with initial weights and the images input to the ML model 100 to generate predictions. The weights of the nodes may be adjusted based on how accurate the prediction is as compared to the labels. The weights applied by a node may be adjusted during training based on a loss function, which is a function that describes how accurately the predictions of the NN are as compared to the expected results; an optimization algorithm, which helps determine weight settings adjustments based on the loss function; and/or a backpropagation of error algorithm, which applies the weight adjustments back through the layers of the NN. Any optimization algorithm (e.g., gradient descent, mini-batch gradient descent, stochastic gradient descent, adaptive optimizers, momentum, etc.), loss function (e.g., mean-squared error, cross-entropy, maximum likelihood, etc.), and backpropagation of error algorithm (e.g., static or recurrent backpropagation) may be used within the scope of this disclosure.
In some cases, training the ML model 100 is performed during development of the ML model 100 and may be performed by a system or device separate from the system or device that runs the trained ML model.
Example Hardware for Executing ML Models
The CPU cores 202 may be coupled to a crossbar (e.g., interconnect) 206, which interconnects and routes data between various components of the device. In some cases, the crossbar 206 may be a memory controller or any other circuit that can provide an interconnect between peripherals. Peripherals may include master peripherals (e.g., components that access memory, such as various processors, processor packages, direct memory access (DMA)/input output components, etc.) and slave peripherals (e.g., memory components, such as double data rate (DDR) random access memory, other types of random access memory, DMA/input output components, etc.). In some cases, the processing cores, such as CPU cores 202, ML accelerator 208, and other processing cores 210 and crossbar 206 may be integrated on a single chip, such as a SoC 222 with a separate external memory. In this example, the crossbar 206 couples the CPU cores 202 with other peripherals, such as an ML accelerator 208 and other processing cores 210, such as a graphics processing unit, radio basebands, coprocessors, microcontrollers, etc., and external memory 214, such as DDR memory, dynamic random access memory (DRAM), flash memory, etc., which may be on a separate chip from the SoC. The crossbar 206 may include or provide access to one or more internal memories that may include any type of memory, such as static random-access memory (SRAM), flash memory, etc. The ML accelerator 208 may include one or more ML cores 216. The ML cores 216 may be processor cores configured to accelerate machine learning models and the ML cores 216 may include one or more internal caches (not shown).
In some cases, the device may be an embedded device which is built into another device and may perform a specific function for the other device. Often embedded devices are resource constrained with a relatively limited amount of compute and memory resources. To help improve the performance of a ML model on target hardware, such as for the embedded device, execution of the ML model may be optimized for the target hardware. Various specific optimizations may be applied.
As discussed above, when a ML model is executed on a processor, the processor typically performs a set of computations based on parameter values input to the ML model along with weights, functions, etc. associated with the ML model. Each node of a layer of the ML model may be associated with a different set of weights, combination functions, activation function, etc., that may be loaded from a memory into the processor for the calculations to be performed. Generally, the closer a memory is to the processor, the faster the processor can retrieve information from the memory. Generally, the faster the processor can retrieve the information from the memory, the faster the processor can perform the calculation. In many cases, cache memory, such as internal cache memory 204 is closer to the processor than an external memory, such as external memory 214 and executing the ML model from the internal cache memory 204 (e.g., retrieving information from the internal cache memory for the processor) is faster than executing from the external memory 214. However, a size of the cache memory is more limited as compared to the size of the external memory.
Conventionally, information associated with nodes of a layer may be loaded, for example, from the external memory into cache memory and then executed from the cache memory. Output from the computations of the layer may also be stored into external memory and reloaded into the cache memory for execution of the next layer. To help avoid loading each layer from external memory, groups of layers, or portions of layers, may be defined for loading into and executing from cache memory. Additionally, storing the output results of the computations for a layer into the cache memory also helps improve performance as compared to storing the output results in the external memory as the output results for a layer are often used as input for the next layer. Techniques to group multiple portions of layers along with output results of these groups are discussed, for example, in discussed in U.S. patent application Ser. No. 17/327,869 and U.S. patent application Ser. No. 16/797,871.
Once an ML model is trained, the ML model 402 may be compiled and/or prepared for a target hardware by an ML model complier 404. It may be understood that the compilation process may include multiple processes, steps, operations, etc., which may be performed separately, and/or in an automated fashion. In this example, the target hardware 406 is shown as a simplified version of the device shown in
It may be understood that the compilation process may include multiple sub-processes. In some cases, in addition to translating the ML model 402 to runtime code, the compilation process may also include one or more sub-processes analyzing execution of the ML model 402 and/or optimizing execution of the ML model 402 on the target hardware 406. In some cases, simulations may be performed after the ML model is trained and as a part of preparing the trained ML model 402 for execution on the target hardware 406. For example, as a part of the compilation and/or translation process, ML model execution on the target hardware 406 may be simulated. In some cases, the simulation of the ML model execution may be performed as a separate process from the compilation/translation process. These simulations may be performed as a part of an optimization process. For example, execution of the ML model using different groupings of the layers of the ML model may be simulated to help identify groupings of layers of the ML model which help improve the performance of the ML model executing on the target hardware 406. In some cases, all possible groupings of the layers of the ML model may be simulated to identify the grouping which optimizes performance of the ML model on the target hardware 406.
Often optimizing for performance on the target hardware 406 is defined by execution times. That is, the optimization process often seeks the lowest execution time for the ML model for the target hardware 406. Of note, the lowest execution time may be described by a variety of metrics, such as frames per second analyzed, minimum amount of computing cycles required to run the ML model against a number of inputs, etc.
In some cases, the optimization process may identify multiple optimization solutions, such as a particular grouping of layers, based on two or more factors. Often it is assumed that optimizing for lowest execution time includes optimizing for certain other factors, such as external memory bandwidth usage. For example, it is often assumed that optimizing for the lowest execution time will automatically result in the lowest amount of external memory bandwidth usage. For some ML models, optimizing for the lowest execution time does minimize the amount of external memory bandwidth needed. For some ML models, there may be a unique solution which minimizes the execution time and another solution which minimizes the amount of external memory bandwidth needed. Additionally, other solutions may offer differing trade-offs between the execution time and the amount of external memory bandwidth needed. These solutions may be considered as forming a pareto-optimal frontier of solutions. In some cases, it may be beneficial to offer different solutions along the pareto-optimal frontier of solutions, for example, to allow more precise tuning of the execution of the ML model.
It should be understood that this example optimization process described herein was chosen for clarity and simplicity purposes and it should be understood that any other optimization technique may be applied to the concepts described herein.
As an example, graphically plotting the solutions that may be obtained, during an optimization process, by simulating the ML model can help illustrate how different solutions along the pareto-optimal frontier of solutions may be selected.
Each solution may be evaluated, for example as a part of the simulation, based on a set of factors. In this example, the factors include a number of processor cycles used per frame 504 to execute the ML model on the target hardware, as plotted on the X-axis, and an amount of external memory bandwidth used per frame 506, as plotted on the Y-axis. For some ML models, a best performing solution with respect to one factor may not be the best performing solution with respect to another factor. Here, a first solution 508 (e.g., grouping of the ML layers) may use less computing cycles than other solutions of the set of solutions 502. The first solution 508 may be identified by determining the solution, of a set of solutions, having the lowest number of used processor cycles per frame. In this example, the first solution 508 may also use more external memory bandwidth per frame 506 than a second solution 510. In this example, the first solution 508 represents the optimized solution for the first factor (e.g., lowest execution time/minimum processor cycles) on the target hardware and the second solution 510 represents the optimized solution for the second factor (e.g., external memory bandwidth) on the target hardware. The second solution 510 may be identified by determining the solution, of a set of solutions, having the lowest external memory bandwidth usage. Thus, when optimizing based on the first and second factors, the optimization process may identify these two optimized solutions.
In some cases, certain solutions of the set of solutions 502 may form a pareto frontier 512 such that no one factor may be improved upon without making another factor worse off. Thus, in this example, for solutions on the pareto frontier 512, any reductions in an amount of processor cycles used per frame 504 results in an increase in an amount of external memory bandwidth user per frame 506. In some cases, solutions on a pareto optimal frontier 512 may be determined. Solutions on the pareto optimal frontier 512 may be determined by comparing an evaluated solution to the other solutions to determine whether any other solution improves on the evaluated solution in every dimension (e.g., factors). For example, both the first solution 508 and the second solution 510 are on the pareto optimal frontier 512 as no other solution has a lower number of used processor cycles for the first solution 508, and no other solution has a lower amount of external memory bandwidth usage for the second solution 510. As another example, when evaluating evaluated solution 514 and 516, these evaluated solutions 514 and 516 may be compared with other solutions including solution 518. Solution 518 has both a lower number of used processor cycles and a lower amount of external memory bandwidth usage than both evaluated solution 514 and 516. Thus, solution 518 strictly dominates both evaluated solution 514 and 516 and evaluated solution 514 and 516 are not solutions on the pareto optimal frontier 512. In contrast, solution 518 is not strictly dominated by any other solution. While the second solution 510 may have a lower amount of external memory bandwidth usage as compared to solution 518, the second solution 510 uses a higher number of processor cycles. Similarly, the first solution 508 may have a lower number of used processor cycles per frame as compared to solution 518, the first solution 508 uses a larger amount of external memory bandwidth.
Additionally, one or more compromise solutions may also be identified by the optimization process. These compromise solutions may represent solutions that trade-off between two or more factors. For example, some solutions of the set of solutions 502 may use relatively more computing cycles, but less external memory bandwidth, than the first solution 508 and less computing cycles, but more external memory bandwidth than the second solution 510. In some cases, these compromise solutions may fall between the solutions most optimized for a particular factor (e.g., the first solution 508 and second solution 510). One or more compromise solutions may be selected from solutions on the pareto frontier 512. In some cases, these compromise solutions may be selected from solutions on the pareto frontier 512 which minimize the distance between the solution, as plotted against the factors, and a hypothetical perfect solution (e.g., 1 processor cycle, no external memory bandwidth used). When optimizing based on the first and second factors, the optimization process may identify, in addition to the two optimized solutions, one or more of the selected compromise solutions.
In some cases, compromise solutions may be identified based on a range. Ranges may be defined based on the best performing solutions for the factors. As an example, a first range from the first solution 508 may be defined and a second range from the second solution 510 may also be defined. The first range and the second range may be predefined or defined at run time, such as by a user. The first range and the second range may be the same or different. As an example, the first range may be a value range, such as within 10%, for the factor that the first solution 508 is the best performing solution for. Thus, pareto frontier solutions having several used processor cycles per frame within 10% of the number of used processor cycles per frame of first solution 508 may be selected as a first set of candidate compromise solutions. Of this first set of candidate compromise solutions, the solution with the best performance with respect to the other factor (here, the least amount of external memory bandwidth used per frame) may be selected as a compromise solution 518A. Similarly, the second range may be a value range, such as 10%, for the factor that the second solution 510 is the best performing solution for. Thus, pareto frontier solutions having an external memory bandwidth usage per frame within 10% of the external memory bandwidth usage per frame of the second solution 510 may be selected as a second set of candidate compromise solutions. Of these second set of candidate compromise solutions, the solution with the best performance with respect to the other factor (here, the least number of used processor cycles per frame) may be selected as a compromise solution 518B.
Once the optimized/compromise solutions are identified, the solutions may be provided to the target hardware for use when executing the ML model, for example, as a part of the execution information. Returning to the example discussed above, the execution information may include an indication of a plurality of groupings of the layers and an indication of a factor associated with the grouping. This indication of a factor may be implicit or explicit. For example, the indication of the factor may be implicit based on an ordering of the groupings of the layers where the first layer grouping may be the layer grouping optimizing for a first factor, a second layer grouping, in the order, may be the layer grouping optimizing for the second factor, and so forth. In some cases, implicit indication of a factor may be predefined. As an example of explicitly indicating a factor, the execution information may include a metadata and/or a tag indicating which grouping is associated with which factor, or when to use a particular grouping. The execution information may be provided to the target hardware along with runtime code associated with the ML model.
As shown, the ML model may be associated with execution information indicating two ML model optimization configurations. In this example, a first ML model configuration may be optimized for a lower number of processor cycles (or lower execution time) as compared to a second ML model configuration, which may be optimized for a lower external memory bandwidth usage. The target hardware may be configured to determine whether a particular runtime condition has been met 602 prior to a next execution of the ML model. For example, the target hardware may be configured to determine whether a certain percentage or amount of external memory bandwidth is already in use, whether there are external memory access stalls occurring, memory bus temperatures out of a certain range, etc. Other runtime conditions may include a temperature of one or more portions of the target hardware, an amount of power used (or available for use) by the target hardware, an amount of processor compute cycles used (or available for use), etc. If such a runtime condition exits, the target hardware may execute the ML model, for example during the next execution of the ML model, based on the second ML model configuration 604. For example, if the certain percentage of external memory is already in use, the target hardware may execute the ML model using the second ML model configuration by setting a flag/parameter, passing the ML model configuration information associated with the second ML model configuration, etc. As a more specific example, external memory bandwidth of the target device may be limited and if another process is already using sufficient external memory bandwidth such that execution of the ML model using the first ML model configuration would be impacted by the external memory bandwidth available, then the target hardware may execute the ML model using the second ML model configuration. In some cases, multiple runtime conditions may be defined. For example, if the certain percentage of external memory is already in use or if there are external memory access stalls, then the target hardware may execute the ML model based on the second ML model configuration 604.
If such a runtime condition does not exit, then the target hardware may execute the ML model based on the first ML model configuration 604. For example, if the certain percentage of external memory is not already in use, the target hardware may execute the ML model using the first ML model configuration. In some cases, the ML model may default to the first ML model configuration unless an indication that the ML model should be executed using the second ML model configuration is received.
In some cases, the execution information may include information for configuring the target hardware to adjust the execution of the ML model based on the execution information. For example, the execution information may include instructions (e.g., software) for execution on the target hardware to configure the target hardware to determine whether a particular runtime condition has been met and/or instructions for how to execute the ML model using the various ML model configurations.
At block 704, the ML model is executed based on the first execution data. For example, the ML model may execute using a particular set of execution information. As a more specific example, the ML model may execute using execution information optimizing the ML model to execute in the fewest number of processor cycles. In some cases, a particular set of execution information may be a default configuration. At block 706, a determination is made to execute the ML model based on the second execution data. For example, the target hardware may monitor one or more runtime conditions. If one or more of these runtime conditions are met, the target hardware may determine to execute the ML model based on the second execution data. For example, if the target hardware determines that a certain amount of memory bandwidth is in use, then the target hardware may determine to alter the way the ML model is executing to help reduce the amount of memory bandwidth used by the ML model. As a more detailed example, the target hardware may determine to execute the ML model using a grouping of the ML model layers which executes using a lowest amount of memory bandwidth. At block 708, the ML model is executed based on the second execution data. For example, the target hardware may then execute the ML model again using the grouping of the ML model layers which executes using a lowest amount of memory bandwidth.
In this description, the term “couple” may cover connections, communications, or signal paths that enable a functional relationship consistent with this description. For example, if device A generates a signal to control device B to perform an action: (a) in a first example, device A is coupled to device B by direct connection; or (b) in a second example, device A is coupled to device B through intervening component C if intervening component C does not alter the functional relationship between device A and device B, such that device B is controlled by device A via the control signal generated by device A.
Modifications are possible in the described embodiments, and other embodiments are possible, within the scope of the claims.
