Step aware quantization
Step-aware quantization in diffusion models addresses resource consumption and accuracy issues by adapting activation encodings based on range estimation, enhancing performance in edge devices.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-09-08
- Publication Date
- 2026-03-12
AI Technical Summary
Conventional diffusion models face challenges with significant compute and memory resource consumption due to attention operations, leading to latency and accuracy degradation when deployed on edge devices, and existing quantization methods accumulate quantization loss over multiple iterations, affecting model performance.
Implement step-aware quantization by determining range estimation for activations at each iteration and switching between multiple activation encodings based on successively increasing activation ranges, reducing latency and improving model accuracy with lower bit widths.
This approach enhances model accuracy and reduces latency in edge device deployments by minimizing quantization loss across iterations, allowing efficient inference without hardware modifications.
Smart Images

Figure CN2024117627_12032026_PF_FP_ABST
Abstract
Description
STEP AWARE QUANTIZATION
[0001] FIELD OF THE DISCLOSURE
[0002] Aspects of the present disclosure generally relate to machine learning models, and more particularly to quantization in diffusion models.BACKGROUND
[0003] Artificial neural networks may comprise interconnected groups of artificial neurons (e.g., neuron models) . The artificial neural network (ANN) may be a computational device or be represented as a method to be performed by a computational device. Convolutional neural networks (CNNs) are a type of feed-forward ANN. Convolutional neural networks may include collections of neurons that each have a receptive field and that collectively tile an input space. Convolutional neural networks, such as deep convolutional neural networks (DCNs) , have numerous applications. In particular, these neural network architectures are used in various technologies, such as image recognition, speech recognition, acoustic scene classification, keyword spotting, autonomous driving, and other classification tasks.
[0004] Diffusion models are generative models designed to transform easily generated data into more complex and realistic data through a sequence of invertible transformations. For example, a diffusion model may generate an image from a text prompt. Diffusion models may employ a U-Net architecture, which is so named because of the u-shape of the network, including a down-sampling path and an up-sampling path. U-Net architectures in current state-of-the-art (SOTA) models adopt a combination of attention layers and convolutional layers at each stage of the U-Net. U-Net architectures adopt global self-attention and cross-attention operations at all spatial resolutions. Attention operations consume significant compute and memory resources making edge device inference and on-device generative artificial intelligence (AI) inference challenging.SUMMARY
[0005] The present disclosure is set forth in the independent claims, respectively. Some aspects of the disclosure are described in the dependent claims.
[0006] In various aspects of the present disclosure, a processor-implemented method includes receiving a diffusion model. The processor-implemented method also includes processing, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations. The processor-implemented method additionally includes determining a range estimation for activations of each iteration of the multiple iterations. The processor-implemented method further includes setting a first activation encoding for at least one iteration based on the range estimation.
[0007] Various aspects of the present disclosure are directed to an apparatus. The apparatus has at least one memory and one or more processors coupled to the at least one memory. The processor (s) is configured to receive a diffusion model. The processor (s) is still further configured to process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations. The processor (s) is also configured to determine a range estimation for activations of each iteration of the multiple iterations. The processor (s) is further configured to set a first activation encoding for at least one iteration based on the range estimation.
[0008] Various aspects of the present disclosure are directed to an apparatus. The apparatus has at least one memory and one or more processors coupled to the at least one memory. The processor (s) is configured to receive a diffusion model. The processor (s) is still further configured to process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations. The processor (s) is also configured to switch between multiple activation encodings for the multiple iterations, wherein the multiple activation encodings are based on groupings of successively increasing activation ranges.
[0009] Additional features and advantages of the disclosure will be described below. It should be appreciated by those skilled in the art that this disclosure may be readily utilized as a basis for modifying or designing other structures for carrying out the same purposes of the present disclosure. It should also be realized by those skilled in the art that such equivalent constructions do not depart from the teachings of the disclosure as set forth in the appended claims. The novel features, which are believed to be characteristic of the disclosure, both as to its organization and method of operation, together with further objects and advantages, will be better understood from the following description when considered in connection with the accompanying figures. It is to be expressly understood, however, that each of the figures is provided for the purpose of illustration and description only and is not intended as a definition of the limits of the present disclosure.BRIEF DESCRIPTION OF THE DRAWINGS
[0010] The features, nature, and advantages of the present disclosure will become more apparent from the detailed description set forth below when taken in conjunction with the drawings in which like reference characters identify correspondingly throughout.
[0011] FIGURE 1 illustrates an example implementation of a neural network using a system-on-a-chip (SOC) , including a general-purpose processor in accordance with certain aspects of the present disclosure.
[0012] FIGURES 2A, 2B, and 2C are diagrams illustrating a neural network in accordance with various aspects of the present disclosure.
[0013] FIGURE 2D is a diagram illustrating an exemplary deep convolutional network (DCN) in accordance with various aspects of the present disclosure.
[0014] FIGURE 3 is a block diagram illustrating an exemplary deep convolutional network (DCN) in accordance with various aspects of the present disclosure.
[0015] FIGURE 4 is a block diagram illustrating an exemplary software architecture that may modularize artificial intelligence (AI) functions, in accordance with various aspects of the present disclosure.
[0016] FIGURE 5 is a block diagram illustrating an example process for step aware quantization, in accordance with various aspects of the present disclosure.
[0017] FIGURES 6A-6B are diagrams illustrating an example range estimation, in accordance with various aspects of the present disclosure.
[0018] FIGURE 7 is a block diagram illustrating an example diffusion model using step aware quantization, in accordance with aspects of the present disclosure.
[0019] FIGURE 8 is a diagram illustrating an example activation encoding assignment based on a range variation pattern, in accordance with various aspects of the present disclosure.
[0020] FIGURE 9 is a block diagram illustrating an example U-Net structure of a diffusion model with mixed bit width encoding quantization, in accordance with various aspects of the present disclosure.
[0021] FIGURE 10 is a flow diagram illustrating a processor-implemented method for step aware quantization in diffusion models, in accordance with aspects of the present disclosure.DETAILED DESCRIPTION
[0022] The detailed description set forth below, in connection with the appended drawings, is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described may be practiced. The detailed description includes specific details for the purpose of providing a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form in order to avoid obscuring such concepts.
[0023] Based on the teachings, one skilled in the art should appreciate that the scope of the disclosure is intended to cover any aspect of the disclosure, whether implemented independently of or combined with any other aspect of the disclosure. For example, an apparatus may be implemented or a method may be practiced using any number of the aspects set forth. In addition, the scope of the disclosure is intended to cover such an apparatus or method practiced using other structure, functionality, or structure and functionality in addition to or other than the various aspects of the disclosure set forth. It should be understood that any aspect of the disclosure disclosed may be embodied by one or more elements of a claim.
[0024] The word “exemplary” is used to mean “serving as an example, instance, or illustration. ” Any aspect described as “exemplary” is not necessarily to be construed as preferred or advantageous over other aspects.
[0025] Although particular aspects are described, many variations and permutations of these aspects fall within the scope of the disclosure. Although some benefits and advantages of the preferred aspects are mentioned, the scope of the disclosure is not intended to be limited to particular benefits, uses or objectives. Rather, aspects of the disclosure are intended to be broadly applicable to different technologies, system configurations, networks, and protocols, some of which are illustrated by way of example in the figures and in the following description of the preferred aspects. The detailed description and drawings are merely illustrative of the disclosure rather than limiting, the scope of the disclosure being defined by the appended claims and equivalents thereof.
[0026] Diffusion models are a class of generative artificial intelligence (AI) models. Diffusion models may be trained to generate data similar to data on which the model was trained. Diffusion models have grown in popularity because such models may generate high-resolution images of varying quality. The underlying concept of diffusion models was inspired by physical processes, such as the way particles diffuse in a medium over time. Similarly, diffusion models may gradually add noise to input data in a forward diffusion process and may remove the noise in a reverse diffusion process to recover a high quality version of the input.
[0027] Diffusion models may employ a U-Net architecture, which is so named because of the u-shape of the network, including a symmetrical encoder-decoder structure. The encoder has a down-sampling path and the decoder has an up-sampling path. U-Net architectures in current state-of-the-art (SOTA) models adopt a combination of attention layers and convolutional layers at each stage of the U-Net. U-Net architectures adopt global self-attention and cross-attention operations at all spatial resolutions. Attention operations consume significant compute and memory resources making edge device inference challenging and on-device generative artificial intelligence (AI) inference challenging.
[0028] Because of the growing popularity of diffusion models, various enhancements have been implemented. One enhancement is stable diffusion. Stable diffusion refers to a latent text-to-image diffusion model. Stable diffusion models are designed to generate data, such as images, by iteratively denoising a random noise vector through a series of steps. Stable diffusion is specifically an implementation of diffusion models that are optimized for creating high-quality images and have become popular due to efficiency and flexibility.
[0029] That is, as the image size grows, the computational power (e.g., number of computations) to process such images increases. This is especially pronounced in an operation called self-attention, where the amount of operations grows quadratically with the number of inputs. A 128 pixel square image has four times as many pixels as a 64 pixel square image, and thus consumes 16x (e.g., 42) the memory and compute resources in a self-attention layer. Stable diffusion models may mitigate this issue by compressing images into a smaller spatial dimension because images may include a significant amount of redundant information. As such, a stable diffusion model may learn to produce a much smaller representation of an input image and then reconstruct the image based on this smaller latent representation with a high degree of fidelity.
[0030] Conventional forms of stable diffusion (SD) models (e.g., SD 1.5, SD 2.1-low rank adapters (LoRA) and SDXL) may operate over twenty iterative steps (e.g., denoising steps) to generate an image with given text prompts. For higher-quality images, conventional stable diffusion models may perform additional steps (e.g., 50 steps) . Applications for such higher quality images include (but are not limited to) super-resolution image generation, image inpainting generation, or the like and may be widely used in edge devices.
[0031] However, a large number of iteration steps (e.g., 20 or 50 diffusion steps) , may involve an increased latency power cost. As such, some conventional approaches may employ a low bit width model quantization to accelerate the diffusion procedure and the path to compress the SD models for edge device deployment. On the other hand, such a low-bit setting may cause accuracy degradation.
[0032] Quantization may be applied to weights and / or activations of the SD model. That is, the weights of convolution layers may be quantized to a lower precision (e.g., from a 32-bit floating point (also referred to as “full precision” ) value to an 8-bit integer value or a 4-bit integer value) .
[0033] The activations of the SD model may also be quantized. The activations (e.g., intermediate outputs) of a layer may have a wide dynamic range. The large variation in the activation range distribution may result in quantization loss.
[0034] In diffusion procedures, each diffusion step (e.g., a denoising iteration) may generate a noise estimate (e.g., a prediction) . An output of a diffusion step may be simplified and defined as Y (t) =Y (t-1) -α*N (t) , where N (t) is the estimated noise in the t-th step, Y (t) is the output of the t-th step, and α is a scheduler coefficient value.
[0035] The range of data distribution may increase in each iteration step of the diffusion process. As the range of data distribution increases, the quantization errors or loss may increase, and may accumulate with each successive step. That is, in U-Net quantization, each step in the diffusion process may involve a different quantization loss. Each such quantization loss may impact the input of the next step. Accordingly, with each step, the loss is accumulated, ultimately producing a larger loss value in the final U-Net output.
[0036] Conventional approaches attempt to address the issues with respect to quantization loss by buffering all full precision (FP) input samples or output samples as a calibration dataset. Then, conventional approaches use the minimum and maximum for all steps to determine a final shared encoding including a weight encoding and an activation encoding.
[0037] As discussed, when employing the low bit width quantization, there may be large accuracy degradation as the U-Net quantization loss in the diffusion procedure. In diffusion, quantization loss is accumulated into one large value step by step.
[0038] Accordingly, to address these and other issues, aspects of the present disclosure are directed to range estimation for step aware quantization.
[0039] Particular aspects of the subject matter described in this disclosure can be implemented to realize one or more of the following potential advantages. In some examples, the described techniques (e.g., “setting a first activation encoding …based on the range estimation) , may reduce latency and improve the model accuracy with lower bit width (e.g., 4-bit or 8-bit) settings for large vision models. Additionally, the present techniques may provide a unified approach for analyzing the activation range distribution in different iteration steps, and further, for finalizing the activation encoding. Moreover, the described techniques may be applied directly in the legacy hardware setting without hardware modification.
[0040] In some aspects, the described techniques may be implemented and released as an artificial intelligence (AI) model efficiency toolkit (AIMET) feature, to enhance the accuracy for large vision model (LVM) quantization, for instance.
[0041] FIGURE 1 illustrates an example implementation of a system-on-a-chip (SOC) 100, which may include a central processing unit (CPU) 102 or a multi-core CPU configured for range estimation for step aware quantization in diffusion models. Variables (e.g., neural signals and synaptic weights) , system parameters associated with a computational device (e.g., neural network with weights) , delays, frequency bin information, and task information may be stored in a memory block associated with a neural processing unit (NPU) 108, in a memory block associated with a CPU 102, in a memory block associated with a graphics processing unit (GPU) 104, in a memory block associated with a digital signal processor (DSP) 106, in a memory block 118, or may be distributed across multiple blocks. Instructions executed at the CPU 102 may be loaded from a program memory associated with the CPU 102 or may be loaded from a memory block 118.
[0042] The SOC 100 may also include additional processing blocks tailored to specific functions, such as a GPU 104, a DSP 106, a connectivity block 110, which may include fifth generation (5G) connectivity, fourth generation long term evolution (4G LTE) connectivity, Wi-Fi connectivity, USB connectivity, Bluetooth connectivity, and the like, and a multimedia processor 112 that may, for example, detect and recognize gestures. In one implementation, the NPU 108 is implemented in the CPU 102, DSP 106, and / or GPU 104. The SOC 100 may also include a sensor processor 114, image signal processors (ISPs) 116, and / or navigation module 120, which may include a global positioning system.
[0043] The SOC 100 may be based on an ARM, RISC-V (RISC-five) , or any reduced instruction set computing (RISC) architecture. In aspects of the present disclosure, the instructions loaded into the general-purpose processor 102 may include code to receive a diffusion model. The instructions loaded into general-purpose processor 102 may also include code to process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations. The instructions loaded into general-purpose processor 102 may also include code to determine a range estimation for activations of each iteration of the multiple iterations The instructions loaded into general-purpose processor 102 may also include code to set a first activation encoding for at least one iteration based on the range estimation.
[0044] In aspects of the present disclosure, the instructions loaded into the general-purpose processor 102 may include code to receive a diffusion model. The instructions loaded into general-purpose processor 102 may also include code to process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations. The instructions loaded into general-purpose processor 102 may also include code to switch between multiple activation encodings for the multiple iterations, wherein the multiple activation encodings are based on groupings of successively increasing activation ranges.
[0045] Deep learning architectures may perform an object recognition task by learning to represent inputs at successively higher levels of abstraction in each layer, thereby building up a useful feature representation of the input data. In this way, deep learning addresses a major bottleneck of traditional machine learning. Prior to the advent of deep learning, a machine learning approach to an object recognition problem may have relied heavily on human engineered features, perhaps in combination with a shallow classifier. A shallow classifier may be a two-class linear classifier, for example, in which a weighted sum of the feature vector components may be compared with a threshold to predict to which class the input belongs. Human engineered features may be templates or kernels tailored to a specific problem domain by engineers with domain expertise. Deep learning architectures, in contrast, may learn to represent features that are similar to what a human engineer might design, but through training. Furthermore, a deep network may learn to represent and recognize new types of features that a human might not have considered.
[0046] A deep learning architecture may learn a hierarchy of features. If presented with visual data, for example, the first layer may learn to recognize relatively simple features, such as edges, in the input stream. In another example, if presented with auditory data, the first layer may learn to recognize spectral power in specific frequencies. The second layer, taking the output of the first layer as input, may learn to recognize combinations of features, such as simple shapes for visual data or combinations of sounds for auditory data. For instance, higher layers may learn to represent complex shapes in visual data or words in auditory data. Still higher layers may learn to recognize common visual objects or spoken phrases.
[0047] Deep learning architectures may perform especially well when applied to problems that have a natural hierarchical structure. For example, the classification of motorized vehicles may benefit from first learning to recognize wheels, windshields, and other features. These features may be combined at higher layers in different ways to recognize cars, trucks, and airplanes.
[0048] Neural networks may be designed with a variety of connectivity patterns. In feed-forward networks, information is passed from lower to higher layers, with each neuron in a given layer communicating to neurons in higher layers. A hierarchical representation may be built up in successive layers of a feed-forward network, as described above. Neural networks may also have recurrent or feedback (also called top-down) connections. In a recurrent connection, the output from a neuron in a given layer may be communicated to another neuron in the same layer. A recurrent architecture may be helpful in recognizing patterns that span more than one of the input data chunks that are delivered to the neural network in a sequence. A connection from a neuron in a given layer to a neuron in a lower layer is called a feedback (or top-down) connection. A network with many feedback connections may be helpful when the recognition of a high-level concept may aid in discriminating the particular low-level features of an input.
[0049] The connections between layers of a neural network may be fully connected or locally connected. FIGURE 2A illustrates an example of a fully connected neural network 202. In a fully connected neural network 202, a neuron in a first layer may communicate its output to every neuron in a second layer, so that each neuron in the second layer will receive input from every neuron in the first layer. FIGURE 2B illustrates an example of a locally connected neural network 204. In a locally connected neural network 204, a neuron in a first layer may be connected to a limited number of neurons in the second layer. More generally, a locally connected layer of the locally connected neural network 204 may be configured so that each neuron in a layer will have the same or a similar connectivity pattern, but with connections strengths that may have different values (e.g., 210, 212, 214, and 216) . The locally connected connectivity pattern may give rise to spatially distinct receptive fields in a higher layer because the higher layer neurons in a given region may receive inputs that are tuned through training to the properties of a restricted portion of the total input to the network.
[0050] One example of a locally connected neural network is a convolutional neural network. FIGURE 2C illustrates an example of a convolutional neural network 206. The convolutional neural network 206 may be configured such that the connection strengths associated with the inputs for each neuron in the second layer are shared (e.g., 208) . Convolutional neural networks may be well suited to problems in which the spatial location of inputs is meaningful.
[0051] One type of convolutional neural network is a deep convolutional network (DCN) . FIGURE 2D illustrates a detailed example of a DCN 200 designed to recognize visual features from an image 226 input from an image capturing device 230, such as a car-mounted camera. The DCN 200 of the current example may be trained to identify traffic signs and a number provided on the traffic sign. Of course, the DCN 200 may be trained for other tasks, such as identifying lane markings or identifying traffic lights.
[0052] The DCN 200 may be trained with supervised learning. During training, the DCN 200 may be presented with an image, such as the image 226 of a speed limit sign, and a forward pass may then be computed to produce an output 222. The DCN 200 may include a feature extraction section and a classification section. Upon receiving the image 226, a convolutional layer 232 may apply convolutional kernels (not shown) to the image 226 to generate a first set of feature maps 218. As an example, the convolutional kernel for the convolutional layer 232 may be a 5x5 kernel that generates 28x28 feature maps. In the present example, because four different feature maps are generated in the first set of feature maps 218, four different convolutional kernels were applied to the image 226 at the convolutional layer 232. The convolutional kernels may also be referred to as filters or convolutional filters.
[0053] The first set of feature maps 218 may be subsampled by a max pooling layer (not shown) to generate a second set of feature maps 220. The max pooling layer reduces the size of the first set of feature maps 218. That is, a size of the second set of feature maps 220, such as 14x14, is less than the size of the first set of feature maps 218, such as 28x28. The reduced size provides similar information to a subsequent layer while reducing memory consumption. The second set of feature maps 220 may be further convolved via one or more subsequent convolutional layers (not shown) to generate one or more subsequent sets of feature maps (not shown) .
[0054] In the example of FIGURE 2D, the second set of feature maps 220 is convolved to generate a first feature vector 224. Furthermore, the first feature vector 224 is further convolved to generate a second feature vector 228. Each feature of the second feature vector 228 may include a number that corresponds to a possible feature of the image 226, such as “sign, ” “60, ” and “100. ” A softmax function (not shown) may convert the numbers in the second feature vector 228 to a probability. As such, an output 222 of the DCN 200 may be a probability of the image 226 including one or more features.
[0055] In the present example, the probabilities in the output 222 for “sign” and “60” are higher than the probabilities of the others of the output 222, such as “30, ” “40, ” “50, ” “70, ” “80, ” “90, ” and “100” . Before training, the output 222 produced by the DCN 200 may likely be incorrect. Thus, an error may be calculated between the output 222 and a target output. The target output is the ground truth of the image 226 (e.g., “sign” and “60” ) . The weights of the DCN 200 may then be adjusted so the output 222 of the DCN 200 is more closely aligned with the target output.
[0056] To adjust the weights, a learning algorithm may compute a gradient vector for the weights. The gradient may indicate an amount that an error would increase or decrease if the weight were adjusted. At the top layer, the gradient may correspond directly to the value of a weight connecting an activated neuron in the penultimate layer and a neuron in the output layer. In lower layers, the gradient may depend on the value of the weights and on the computed error gradients of the higher layers. The weights may then be adjusted to reduce the error. This manner of adjusting the weights may be referred to as “back propagation” as it involves a “backward pass” through the neural network.
[0057] In practice, the error gradient of weights may be calculated over a small number of examples, so that the calculated gradient approximates the true error gradient. This approximation method may be referred to as stochastic gradient descent. Stochastic gradient descent may be repeated until the achievable error rate of the entire system has stopped decreasing or until the error rate has reached a target level. After learning, the DCN 200 may be presented with new images (e.g., the speed limit sign of the image 226) and a forward pass through the DCN 200 may yield an output 222 that may be considered an inference or a prediction of the DCN 200.
[0058] Deep belief networks (DBNs) are probabilistic models comprising multiple layers of hidden nodes. DBNs may be used to extract a hierarchical representation of training data sets. A DBN may be obtained by stacking up layers of Restricted Boltzmann Machines (RBMs) . An RBM is a type of artificial neural network that can learn a probability distribution over a set of inputs. Because RBMs can learn a probability distribution in the absence of information about the class to which each input should be categorized, RBMs are often used in unsupervised learning. Using a hybrid unsupervised and supervised paradigm, the bottom RBMs of a DBN may be trained in an unsupervised manner and may serve as feature extractors, and the top RBM may be trained in a supervised manner (on a joint distribution of inputs from the previous layer and target classes) and may serve as a classifier.
[0059] DCNs are networks of convolutional networks, configured with additional pooling and normalization layers. DCNs have achieved state-of-the-art performance on many tasks. DCNs can be trained using supervised learning in which both the input and output targets are known for many exemplars and are used to modify the weights of the network by use of gradient descent methods.
[0060] DCNs may be feed-forward networks. In addition, as described above, the connections from a neuron in a first layer of a DCN to a group of neurons in the next higher layer are shared across the neurons in the first layer. The feed-forward and shared connections of DCNs may be exploited for fast processing. The computational burden of a DCN may be much less, for example, than that of a similarly sized neural network that comprises recurrent or feedback connections.
[0061] The processing of each layer of a convolutional network may be considered a spatially invariant template or basis projection. If the input is first decomposed into multiple channels, such as the red, green, and blue channels of a color image, then the convolutional network trained on that input may be considered three-dimensional, with two spatial dimensions along the axes of the image and a third dimension capturing color information. The outputs of the convolutional connections may be considered to form a feature map in the subsequent layer, with each element of the feature map (e.g., 220) receiving input from a range of neurons in the previous layer (e.g., feature maps 218) and from each of the multiple channels. The values in the feature map may be further processed with a non-linearity, such as a rectification, max (0, x) . Values from adjacent neurons may be further pooled, which corresponds to down-sampling, and may provide additional local invariance and dimensionality reduction. Normalization, which corresponds to whitening, may also be applied through lateral inhibition between neurons in the feature map.
[0062] FIGURE 3 is a block diagram illustrating a DCN 350. The DCN 350 may include multiple different types of layers based on connectivity and weight sharing. As shown in FIGURE 3, the DCN 350 includes the convolution blocks 354A, 354B. Each of the convolution blocks 354A, 354B may be configured with a convolution layer (CONV) 356, a normalization layer (LNorm) 358, and a max pooling layer (MAX POOL) 360.
[0063] Although only two of the convolution blocks 354A, 354B are shown, the present disclosure is not so limiting, and instead, any number of the convolution blocks 354A, 354B may be included in the DCN 350 according to design preference.
[0064] The convolution layers 356 may include one or more convolutional filters, which may be applied to the input data to generate a feature map. The normalization layer 358 may normalize the output of the convolution filters. For example, the normalization layer 358 may provide whitening or lateral inhibition. The max pooling layer 360 may provide down-sampling aggregation over space for local invariance and dimensionality reduction.
[0065] The parallel filter banks, for example, of a deep convolutional network may be loaded on a CPU 102 or GPU 104 of an SOC 100 (e.g., FIGURE 1) to achieve high performance and low power consumption. In alternative embodiments, the parallel filter banks may be loaded on the DSP 106 or an ISP 116 of an SOC 100. In addition, the DCN 350 may access other processing blocks that may be present on the SOC 100, such as sensor processor 114 and navigation module 120, dedicated, respectively, to sensors and navigation.
[0066] The DCN 350 may also include one or more fully connected layers 362 (FC1 and FC2) . The DCN 350 may further include a logistic regression (LR) layer 364. Between each layer 356, 358, 360, 362, 364 of the DCN 350 are weights (not shown) that are to be updated. The output of each of the layers (e.g., 356, 358, 360, 362, 364) may serve as an input of a succeeding one of the layers (e.g., 356, 358, 360, 362, 364) in the DCN 350 to learn hierarchical feature representations from input data 352 (e.g., images, audio, video, sensor data and / or other input data) supplied at the first of the convolution blocks 354A. The output of the DCN 350 is a classification score 366 for the input data 352. The classification score 366 may be a set of probabilities, where each probability is the probability of the input data including a feature from a set of features.
[0067] FIGURE 4 is a block diagram illustrating an exemplary software architecture 400 that may modularize artificial intelligence (AI) functions. Using the architecture 400, applications may be designed that may cause various processing blocks of an SOC 420 (for example a CPU 422, a DSP 424, a GPU 426 and / or an NPU 428) (which may be similar to SOC 100 of FIGURE 1) to support range estimation for step aware quantization in diffusion models for an AI application 402, according to aspects of the present disclosure. The architecture 400 may, for example, be included in a computational device, such as a smartphone.
[0068] The AI application 402 may be configured to call functions defined in a user space 404 that may, for example, provide for the detection and recognition of a scene indicative of the location at which the computational device including the architecture 400 currently operates. The AI application 402 may, for example, configure a microphone and a camera differently depending on whether the recognized scene is an office, a lecture hall, a restaurant, or an outdoor setting such as a lake. The AI application 402 may make a request to compiled program code associated with a library defined in an AI function application programming interface (API) 406. This request may ultimately rely on the output of a deep neural network configured to provide an inference response based on video and positioning data, for example.
[0069] The run-time engine 408, which may be compiled code of a runtime framework, may be further accessible to the AI application 402. The AI application 402 may cause the run-time engine 408, for example, to request an inference at a particular time interval or triggered by an event detected by the user interface of the AI application 402. When caused to provide an inference response, the run-time engine 408 may in turn send a signal to an operating system in an operating system (OS) space 410, such as a Kernel 412, running on the SOC 420. In some examples, the Kernel 412 may be a LINUX Kernel. The operating system, in turn, may cause a continuous relaxation of quantization to be performed on the CPU 422, the DSP 424, the GPU 426, the NPU 428, or some combination thereof. The CPU 422 may be accessed directly by the operating system, and other processing blocks may be accessed through a driver, such as a driver 414, 416, or 418 for, respectively, the DSP 424, the GPU 426, or the NPU 428. In the exemplary example, the deep neural network may be configured to run on a combination of processing blocks, such as the CPU 422, the DSP 424, and the GPU 426, or may be run on the NPU 428.
[0070] As described, when employing a low bit width quantization in diffusion models, there may be significant accuracy degradation due to U-Net quantization loss in the diffusion procedure.
[0071] In diffusion, U-Net quantization loss is cumulative, and accumulates with each successive step into one large loss value. Aspects of the present disclosure discuss techniques for avoiding the quantization loss in the beginning steps, and further to reduce the accumulated loss in the final step.
[0072] Accordingly, aspects of the present disclosure are directed to step aware quantization. In various aspects, specific activation encodings may be enabled for each of the different diffusion steps. Further, in some aspects, a tool or method for range estimation may be employed for analyzing the difference of the range distribution in different steps. Furthermore, techniques for determining activation encodings for inference given different hardware constraints may be provided.
[0073] FIGURE 5 is a block diagram illustrating an example process 500 for step aware quantization, in accordance with various aspects of the present disclosure. Referring to FIGURE 5, in a buffering stage 502, a diffusion model 504 may be operated to determine a calibration dataset. The diffusion model 504 may be configured to process full precision (FP) data (e.g., 32-bit floating point values) . That is, the diffusion model 504 may receive input samples 506. The input samples 506 may comprise (but are not limited to) image data, for instance. The diffusion model 504 may process the input samples 506 without quantization over a complete set of iteration steps (e.g., 20 iteration steps) to generate a set of outputs. Because quantization is not performed, the accuracy of the diffusion model 504 may be higher.
[0074] The input samples 506 and corresponding outputs may be collected and saved as a calibration dataset 508 on a per-step basis. The calibration dataset 508 may be represented as { (Input0, Output0) 0, (Input0, Output0) 1, …, (Inputk, Outputk) s, …, } , where k is the kth sample, and s is the step index. For the kth sample, the Outputk (s) is the Inputk (s+1) stored in a buffer such as a location in memory (e.g., 118 of FIGURE 1) , for instance.
[0075] In a calibration stage 510, the diffusion model 504 may be subjected to a post-training quantization process. The calibration dataset 508 may be loaded and provided to the diffusion model 504 on a per-step basis to perform a calibration to determine a separate encoding for each iteration step. That is, all samples in the sth step: (Input0, Output0) s, (Input1, Output1) s, (Input2, Output2) s, …, may be processed and used to calculate an activation range for each of the layers in the sth step.
[0076] In addition, an activation encoding 516 may be derived for the sth step. The calibration stage 510 may be repeated to determine the corresponding encodings for each step. The activation encodings 516 for each step may be saved in memory.
[0077] Thereafter, in an inference stage 512, the activation encodings 516 may be loaded into a quantized diffusion model 520 according to the iteration step. Then, the diffusion model 520 may process a set of input samples 514 to generate an inference at each step. That is, the diffusion model 520 may generate inferences with the specific activation encoding 516 for each of the steps. The weight encodings (not shown) of the diffusion model 520 may be static and may remain unchanged during inference. On the other hand, the activation encoding 516 may be switched in each of the different iteration steps.
[0078] FIGURES 6A-6B are diagrams 600, 650 illustrating an example range estimation, in accordance with various aspects of the present disclosure. In FIGURE 6A, a range set of activation outputs for each step are shown in the diagram 600. For a layer l and step s, a range distribution may be normalized by rangnormalized (l, s) =rang (l, 0) -rang (l, s) . Then, the range set for each step s may also be generated as {rangnormalized (0, s) , rangnormalized (1, s) , rangnormalized (l, s) , .... , } for all layers. Rather than using the minimum and maximum activation values to determine the activation range, one or more outliers (maximum or minimum value) may be removed in each step. For example, the extreme outliers in a step s may be removed by checking whether the values > (4Q3-3Q1) , or < (-3Q3-2Q1) , where Q3 represents the upper quartile of the range value distributions and Q1 is the lower quartile of the range value distributions.
[0079] In FIGURE 6B, a determined activation range at each step is shown in the diagram 650. In the example of FIGURE 6B, at each step, an effective activation range may be determined according to the maximum / minimum values of the remaining activation encoding values in each step. In some examples, the effective activation range may be expressed as (4Q3-3Q1) - (-3Q3-2Q1) =7Q3-Q1.
[0080] FIGURE 7 is a block diagram illustrating an example diffusion model 700 using step aware quantization, in accordance with aspects of the present disclosure. The example diffusion model 700 may include a U-Net structure 702 and a buffer 704. One U-Net structure 702 is shown (at different iteration steps) in the example diffusion model 700 for ease of illustration but the present disclosure is not so limited. Rather, it should be understood that the example diffusion model 700 may include numerous U-Net structures.
[0081] The buffer 704 may comprise a location in a memory (e.g., memory 118 of FIGURE 1) . The buffer 704 may store activation encodings 708a-708i for each of multiple iteration steps. The iteration steps may correspond to diffusion iterations conducted by the U-Net structure 702 in processing an input 710. A set of weight encodings 706 may be loaded to the U-Net structure 702. In some aspects, the weight encodings 706 may be fixed or static during inference.
[0082] The first activation encoding 708a for a first step may be loaded to the U-Net structure 702. The U-Net structure 702 may then receive and process an input 710 and generate an inference for a first step. Then in a second iteration, a second activation encoding 708b for a second step may be loaded to the U-Net structure 702. The U-Net structure 702 may receive and process the input 710 using the second activation encoding 708b to generate an inference for the second step. Thereafter, the process may be repeated for a number of iteration steps (e.g., 20 iterations or 50 iterations) or according to design preference.
[0083] In some aspects, only a limited number of encodings may be supported. In such cases, an activation encoding may be employed for multiple iteration steps. To determine an arrangement for sharing the activation encodings, the data distribution of all steps may be considered. A threshold may be set based on the range of the data distribution with one activation encoding being shared for the steps within the threshold. Then, the range distribution to select encodings may be determined based on the range estimation. In one example, where the number of activation encodings that may be set is limited to three, a threshold (T) may be determined based on the distribution range.
[0084] A range variation may be set from zero to max (7Q3-Q1) for all steps in a step index s. The threshold T1 may be set as follows: T1 = max (7Q3-Q1) *1 / 3, T2 = max (7Q3-Q1) *2 / 3, T0 = 0.
[0085] The step index s may be given by:
[0086] S0 is a step with range = 0.
[0087] The activation encoding for each step may be determined as follows:
[0088] when {s∣ (7Q3-Q1) s<T1} , sth step uses the S0 encoding,
[0089] when {s∣T1≤ (7Q3-Q1) s≤ T2} , sth step uses the S1 encoding,
[0090] when {s∣ (7Q3-Q1) s>T3} , sth step uses the S2 encoding.
[0091] Accordingly, the determined activation encodings may be used for the respective iteration steps to improve quantization performance (e.g., accuracy) .
[0092] In some aspects, when a limited number of activation encodings are supported, the arrangement for sharing activation encodings may be simplified by considering a range variation pattern. For example, if the activation range increases over the sequence of iteration steps (e.g., iteration 0-19) and the number of different activation encodings is limited to three, the step index may be defined given three thresholds as follows:
[0093] S0 is the step with range = 0.
[0094] As such, the activation encodings for each step may be determined as follows:
[0095] when {s∣s<S1} , sth step uses the S0 encoding,
[0096] when {s∣S1≤s≤S2} , sth step uses the S1 encoding,
[0097] when {s∣s>S2} , sth step uses the S2 encoding.
[0098] FIGURE 8 is a diagram illustrating an example activation encoding assignment 800 based on a range variation pattern, in accordance with various aspects of the present disclosure. Referring to FIGURE 8, a simplified assignment of activation encodings to iteration steps may be determined using a range variation pattern. In the example of FIGURE 8, an activation range may essentially increase with each successive iteration step. The range may be grouped based on the applied thresholds and corresponding activation encodings may be applied for such steps. For instance, an S0 activation encoding may be applied for iteration steps 0-4, an S1 activation encoding may be applied for iteration steps 5-11, and an S2 activation encoding may be applied for iteration steps 12-19. The activation encoding assignment approach of FIGURE 8 may beneficially reduce repeated loading of the same activation encoding from memory.
[0099] FIGURE 9 is a block diagram illustrating an example U-Net structure 900 of a diffusion model with mixed bit width encoding quantization, in accordance with various aspects of the present disclosure. In some aspects, mixed bit width encoding quantization may be enabled. Mixed bit width encoding quantization refers to using a different bit width for quantization for different iteration steps. As shown in the example of FIGURE 9, at iteration step 0, the U-Net structure 900 may operate on full precision (FP) data without employing quantization. In iteration steps 1 and 3, the U-Net structure 900 may employ an 8-bit encoding quantization setting –which enables 8-bit weight encodings and 8-bit activation encodings (W8A8) ) . In iteration step 2, the bit width encoding quantization setting enables 8-bit weight encodings and 16-bit activations encodings (W8A16) .
[0100] When mixed encoding is enabled, the bit width encoding for an iteration may be determined by analyzing the sensitivity of a per-step encoding. The iteration steps with greater sensitivity may be set to use higher bit width, for example.
[0101] Further, if FP and mixed bit width are supported, then the iteration steps with greater sensitivity may be set to use FP.
[0102] The sensitivity may be analyzed by selecting one step as quantized, and the other steps are set with FP and evaluating the accuracy. The iteration step with lowest accuracy may be considered the most sensitive step and may be set to use FP. A next most sensitive step may be set to use higher bit width (e.g., 16-bit) settings, while other iteration steps may use lower bit widths (e.g., 8-bit) settings.
[0103] In some aspects, only one activation encoding may be supported. For example, the limitation may be due to an activation encoding being embedded into the model static graph, (e.g., compressed as one static binary file (e.g., an . ELF file) in which the activation encoding cannot be configured / changed after model deployment in the end device. In other cases, the limitation may be due to hardware constraints (e.g., not enough memory to store more than one activation encoding) .
[0104] When only one activation encoding is supported, the activation encoding may be set by selecting the activation encoding for the first iteration step or the last iteration step based on the range estimation. The selected activation encoding may be used for all iteration steps.
[0105] The activation encoding may also be selected based on the iteration step with the shortest range. Then the selected activation encoding may be employed for all iteration steps.
[0106] Furthermore, the activation encoding may be selected based on the performance for each of the iteration steps. The activation encoding for the iteration having the highest accuracy may be selected and used for all iteration steps.
[0107] FIGURE 10 is a flow diagram illustrating a processor-implemented method 1000 for step aware quantization in diffusion models, in accordance with various aspects of the present disclosure. The processor-implemented method 1000 may be performed by one or more processors such as the CPU (e.g., 102, 422) , GPU (e.g., 104, 426) , and / or other processing units (e.g., DSP 424, NPU 428) , for example.
[0108] As shown in FIGURE 10, at block 1002, the at least one processor receives a diffusion model. The diffusion model may comprise (but is not limited to) a stable diffusion model such as SD 1.5, SD 2.1-low rank adapters (LoRA) and SDXL, for instance.
[0109] At block 1004, the at least one processor processes, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations. As described, for instance with reference to FIGURE 5, in a buffering stage 502, a diffusion model 504 may be operated to determine a calibration dataset. The diffusion model 504 may be configured to process full precision (FP) data (e.g., 32-bit floating point values) . That is, the diffusion model 504 may receive input samples 506. The input samples 506 may comprise (but are not limited to) image data, for instance. The diffusion model 504 may process the input samples 506 without quantization over a complete set of iteration steps (e.g., 20 iteration steps) to generate a set of outputs.
[0110] The input samples 506 and corresponding outputs may be collected and saved as a calibration dataset 508 on a per-step basis. The calibration dataset 508 may be represented as { (Input0, Output0) 0, (Input0, Output0) 1, …, (Inputk, Outputk) s, …, } , where k is the kth sample, and s is the step index. For the kth sample, the Qutputk (s) is the Inputk (s+1) stored in a buffer such as a location in memory (e.g., 118 of FIGURE 1) , for instance.
[0111] At block 1006, the at least one processor determines a range estimation for activations of each iteration of the multiple iterations. For example, as described with reference to FIGURES 6A-6B, for a layer l and step s, a range distribution may be normalized by rangnormalized (l, s) =rang (l, 0) -rang (l, s) . Then, the range set for each step s may also be generated as {rangnormalized (0, s) , rangnormalized (1, s) , rangnormalized (l, s) , .... , } for all layers. Rather than using the minimum and maximum activation values to determine the activation range, one or more outliers (maximum or minimum value) may be removed in each step. For example, the extreme outliers in a step s may be removed by checking whether the values > (4Q3-3Q1) , or <(-3Q3-2Q1) , where Q3 represents the upper quartile of the range value distributions and Q1 is the lower quartile of the range value distributions.
[0112] At block 1008, the at least one processor sets a first activation encoding for at least one iteration based on the range estimation. As described, for instance, with reference to FIGURE 7, the first activation encoding 708a for a first step may be loaded to the U-Net structure 702. The U-Net structure 702 may then receive and process an input 710 and generate an inference for a first step. Then in a second iteration, a second activation encoding 708b for a second step may be loaded to the U-Net structure 702. The U-Net structure 702 may receive and process the input 710 using the second activation encoding 708b to generate an inference for the second step. Thereafter, the process may be repeated for a number of iteration steps (e.g., 20 iterations or 50 iterations) or according to design preference.
[0113] Example Aspects
[0114] Aspect 1: An apparatus, comprising:
[0115] at least one memory; and
[0116] at least one processor coupled to the at least one memory, the at least one processor configured to:
[0117] receive a diffusion model;
[0118] process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations;
[0119] determine a range estimation for activations of each iteration of the multiple iterations; and
[0120] set a first activation encoding for at least one iteration based on the range estimation.
[0121] Aspect 2: The apparatus of Aspect 1, in which the at least one processor is further configured to set at least a second activation encoding for at least a second iteration, the second activation encoding being different than the first activation encoding.
[0122] Aspect 3: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to selectively apply the first activation encoding or the at least the second activation encoding for each iteration of the multiple iterations based on a range distribution.
[0123] Aspect 4: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to set a different activation encoding for each iteration of the multiple iterations.
[0124] Aspect 5: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to determine the range estimation for the activations of each iteration by removing at least one outlier value and computing the range estimation based on a minimum activation value and a maximum activation value of remaining activations of the iteration.
[0125] Aspect 6: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to identify the at least one outlier value based on one or more thresholds.
[0126] Aspect 7: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to set the first activation encoding based on a sensitivity metric.
[0127] Aspect 8: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to:
[0128] store input samples and corresponding activations for each iteration as a calibration dataset of the iteration;
[0129] derive a set of activation encodings for all of the multiple iterations; and
[0130] determine the range estimation based on the set of activation encodings.
[0131] Aspect 9: The apparatus of any of the preceding Aspects, in which the at least one processor is further configured to set the first activation encoding based on a hardware constraint.
[0132] Aspect 10: The apparatus of any of the preceding Aspects, in which the first activation encoding is set based on an activation of a first iteration, a last iteration, an iteration having a shortest range or an iteration having a highest accuracy.
[0133] Aspect 11: A processor-implemented method performed by one or more processors, the processor-implemented method comprising:
[0134] receiving a diffusion model;
[0135] processing, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations;
[0136] determining a range estimation for activations of each iteration of the multiple iterations; and
[0137] setting a first activation encoding for at least one iteration based on the range estimation.
[0138] Aspect 12: The processor-implemented method of Aspect 11, further comprising setting at least a second activation encoding for at least a second iteration, the second activation encoding being different than the first activation encoding.
[0139] Aspect 13: The processor-implemented method of any of Aspects 11-12, further comprising selectively applying the first activation encoding or the at least the second activation encoding for each iteration of the multiple iterations based on a range distribution.
[0140] Aspect 14: The processor-implemented method of any of Aspects 11-13, further comprising setting a different activation encoding for each iteration of the multiple iterations.
[0141] Aspect 15: The processor-implemented method of any of Aspects 11-14, further comprising determining the range estimation for the activations of each iteration by removing at least one outlier value and computing the range estimation based on a minimum activation value and a maximum activation value of remaining activations of the iteration.
[0142] Aspect 16: The processor-implemented method of any of Aspects 11-15, further comprising identifying the at least one outlier value based on one or more thresholds.
[0143] Aspect 17: The processor-implemented method of any of Aspects 11-16, further comprising setting the first activation encoding based on a sensitivity metric.
[0144] Aspect 18: The processor-implemented method of any of Aspects 11-17, further comprising:
[0145] storing input samples and corresponding activations for each iteration as a calibration dataset of the iteration;
[0146] deriving a set of activation encodings for all of the multiple iterations; and
[0147] determining the range estimation based on the set of activation encodings.
[0148] Aspect 19: The processor-implemented method of any of Aspects 11-18, further comprising setting the first activation encoding based on a hardware constraint.
[0149] Aspect 20: The processor-implemented method of any of Aspects 11-19, in which the first activation encoding is set based on an activation of a first iteration, a last iteration, an iteration having a shortest range or an iteration having a highest accuracy.
[0150] Aspect 21: An apparatus, comprising:
[0151] at least one memory; and
[0152] at least one processor coupled to the at least one memory, the at least one processor configured to:
[0153] receive a diffusion model;
[0154] process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations; and
[0155] switching between multiple activation encodings for the multiple iterations, wherein the multiple activation encodings are based on groupings of successively increasing activation ranges.
[0156] Aspect 22: The apparatus of Aspect 21, further comprising determining range estimation for each of the activation ranges based on the set of activations.
[0157] Aspect 23: The apparatus of any of Aspects 21-22, wherein the grouping are determined based on one or more thresholds.
[0158] Aspect 24: The apparatus of any of Aspects 21-23, wherein a number of grouping is based on a hardware constraint.
[0159] In one aspect, the receiving means, processing means, determining means, and / or setting means may be the CPU 102 GPU 104, program memory associated with the CPU 102 or GPU 104, fully connected layers 362, NPU 428, and / or the routing connection processing unit 216 configured to perform the functions recited. In another configuration, the aforementioned means may be any module or any apparatus configured to perform the functions recited by the aforementioned means.
[0160] The various operations of methods described above may be performed by any suitable means capable of performing the corresponding functions. The means may include various hardware and / or software component (s) and / or module (s) , including, but not limited to, a circuit, an application specific integrated circuit (ASIC) , or processor. Generally, where there are operations illustrated in the figures, those operations may have corresponding counterpart means-plus-function components with similar numbering.
[0161] As used, the term “determining” encompasses a wide variety of actions. For example, “determining” may include calculating, computing, processing, deriving, investigating, looking up (e.g., looking up in a table, a database, or another data structure) , ascertaining and the like. Additionally, “determining” may include receiving (e.g., receiving information) , accessing (e.g., accessing data in a memory) and the like. Furthermore, “determining” may include resolving, selecting, choosing, establishing, and the like.
[0162] As used, a phrase referring to “at least one of” a list of items refers to any combination of those items, including single members. As an example, “at least one of: a, b, or c” is intended to cover: a, b, c, a-b, a-c, b-c, and a-b-c.
[0163] The various illustrative logical blocks, modules and circuits described in connection with the present disclosure may be implemented or performed with a general-purpose processor, a digital signal processor (DSP) , an application specific integrated circuit (ASIC) , a field programmable gate array signal (FPGA) or other programmable logic device (PLD) , discrete gate or transistor logic, discrete hardware components or any combination thereof designed to perform the functions described. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any commercially available processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.
[0164] The steps of a method or algorithm described in connection with the present disclosure may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in any form of storage medium that is known in the art. Some examples of storage media that may be used include random access memory (RAM) , read only memory (ROM) , flash memory, erasable programmable read-only memory (EPROM) , electrically erasable programmable read-only memory (EEPROM) , registers, a hard disk, a removable disk, a CD-ROM and so forth. A software module may comprise a single instruction, or many instructions, and may be distributed over several different code segments, among different programs, and across multiple storage media. A storage medium may be coupled to a processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor.
[0165] The methods disclosed comprise one or more steps or actions for achieving the described method. The method steps and / or actions may be interchanged with one another without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of specific steps and / or actions may be modified without departing from the scope of the claims.
[0166] The functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in hardware, an example hardware configuration may comprise a processing system in a device. The processing system may be implemented with a bus architecture. The bus may include any number of interconnecting buses and bridges depending on the specific application of the processing system and the overall design constraints. The bus may link together various circuits including a processor, machine-readable media, and a bus interface. The bus interface may be used to connect a network adapter, among other things, to the processing system via the bus. The network adapter may be used to implement signal processing functions. For certain aspects, a user interface (e.g., keypad, display, mouse, joystick, etc. ) may also be connected to the bus. The bus may also link various other circuits such as timing sources, peripherals, voltage regulators, power management circuits, and the like, which are well known in the art, and therefore, will not be described any further.
[0167] The processor may be responsible for managing the bus and general processing, including the execution of software stored on the machine-readable media. The processor may be implemented with one or more general-purpose and / or special-purpose processors. Examples include microprocessors, microcontrollers, DSP processors, and other circuitry that can execute software. Software shall be construed broadly to mean instructions, data, or any combination thereof, whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise. Machine-readable media may include, by way of example, random access memory (RAM) , flash memory, read only memory (ROM) , programmable read-only memory (PROM) , erasable programmable read-only memory (EPROM) , electrically erasable programmable Read-only memory (EEPROM) , registers, magnetic disks, optical disks, hard drives, or any other suitable storage medium, or any combination thereof. The machine-readable media may be embodied in a computer-program product. The computer-program product may comprise packaging materials.
[0168] In a hardware implementation, the machine-readable media may be part of the processing system separate from the processor. However, as those skilled in the art will readily appreciate, the machine-readable media, or any portion thereof, may be external to the processing system. By way of example, the machine-readable media may include a transmission line, a carrier wave modulated by data, and / or a computer product separate from the device, all which may be accessed by the processor through the bus interface. Alternatively, or in addition, the machine-readable media, or any portion thereof, may be integrated into the processor, such as the case may be with cache and / or general register files. Although the various components discussed may be described as having a specific location, such as a local component, they may also be configured in various ways, such as certain components being configured as part of a distributed computing system.
[0169] The processing system may be configured as a general-purpose processing system with one or more microprocessors providing the processor functionality and external memory providing at least a portion of the machine-readable media, all linked together with other supporting circuitry through an external bus architecture. Alternatively, the processing system may comprise one or more neuromorphic processors for implementing the neuron models and models of neural systems described. As another alternative, the processing system may be implemented with an application specific integrated circuit (ASIC) with the processor, the bus interface, the user interface, supporting circuitry, and at least a portion of the machine-readable media integrated into a single chip, or with one or more field programmable gate arrays (FPGAs) , programmable logic devices (PLDs) , controllers, state machines, gated logic, discrete hardware components, or any other suitable circuitry, or any combination of circuits that can perform the various functionality described throughout this disclosure. Those skilled in the art will recognize how best to implement the described functionality for the processing system depending on the particular application and the overall design constraints imposed on the overall system.
[0170] The machine-readable media may comprise a number of software modules. The software modules include instructions that, when executed by the processor, cause the processing system to perform various functions. The software modules may include a transmission module and a receiving module. Each software module may reside in a single storage device or be distributed across multiple storage devices. By way of example, a software module may be loaded into RAM from a hard drive when a triggering event occurs. During execution of the software module, the processor may load some of the instructions into cache to increase access speed. One or more cache lines may then be loaded into a general register file for execution by the processor. When referring to the functionality of a software module below, it will be understood that such functionality is implemented by the processor when executing instructions from that software module. Furthermore, it should be appreciated that aspects of the present disclosure result in improvements to the functioning of the processor, computer, machine, or other system implementing such aspects.
[0171] If implemented in software, the functions may be stored or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage medium may be any available medium that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Additionally, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL) , or wireless technologies such as infrared (IR) , radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk and disc, as used, include compact disc (CD) , laser disc, optical disc, digital versatile disc (DVD) , floppy disk, and disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Thus, in some aspects, computer-readable media may comprise non-transitory computer-readable media (e.g., tangible media) . In addition, for other aspects computer-readable media may comprise transitory computer-readable media (e.g., a signal) . Combinations of the above should also be included within the scope of computer-readable media.
[0172] Thus, certain aspects may comprise a computer program product for performing the operations presented. For example, such a computer program product may comprise a computer-readable medium having instructions stored (and / or encoded) thereon, the instructions being executable by one or more processors to perform the operations described. For certain aspects, the computer program product may include packaging material.
[0173] Further, it should be appreciated that modules and / or other appropriate means for performing the methods and techniques described can be downloaded and / or otherwise obtained by a user terminal and / or base station as applicable. For example, such a device can be coupled to a server to facilitate the transfer of means for performing the methods described. Alternatively, various methods described can be provided via storage means (e.g., RAM, ROM, a physical storage medium such as a compact disc (CD) or floppy disk, etc. ) , such that a user terminal and / or base station can obtain the various methods upon coupling or providing the storage means to the device. Moreover, any other suitable technique for providing the methods and techniques described to a device can be utilized.
[0174] It is to be understood that the claims are not limited to the precise configuration and components illustrated above. Various modifications, changes, and variations may be made in the arrangement, operation, and details of the methods and apparatus described above without departing from the scope of the claims.
Claims
1.An apparatus, comprising:at least one memory; andat least one processor coupled to the at least one memory, the at least one processor configured to:receive a diffusion model;process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations;determine a range estimation for activations of each iteration of the multiple iterations; andset a first activation encoding for at least one iteration based on the range estimation.2.The apparatus of claim 1, in which the at least one processor is further configured to set at least a second activation encoding for at least a second iteration, the second activation encoding being different than the first activation encoding.3.The apparatus of claim 1, in which the at least one processor is further configured to selectively apply the first activation encoding or the at least the second activation encoding for each iteration of the multiple iterations based on a range distribution.4.The apparatus of claim 1, in which the at least one processor is further configured to set a different activation encoding for each iteration of the multiple iterations.5.The apparatus of claim 1, in which the at least one processor is further configured to determine the range estimation for the activations of each iteration by removing at least one outlier value and computing the range estimation based on a minimum activation value and a maximum activation value of remaining activations of the iteration.6.The apparatus of claim 5, in which the at least one processor is further configured to identify the at least one outlier value based on one or more thresholds.7.The apparatus of claim 1, in which the at least one processor is further configured to set the first activation encoding based on a sensitivity metric.8.The apparatus of claim 1, in which the at least one processor is further configured to:store input samples and corresponding activations for each iteration as a calibration dataset of the iteration;derive a set of activation encodings for all of the multiple iterations; anddetermine the range estimation based on the set of activation encodings.9.The apparatus of claim 1, in which the at least one processor is further configured to set the first activation encoding based on a hardware constraint.10.The apparatus of claim 9, in which the first activation encoding is set based on an activation of a first iteration, a last iteration, an iteration having a shortest range or an iteration having a highest accuracy.11.A processor-implemented method performed by one or more processors, the processor-implemented method comprising:receiving a diffusion model;processing, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations;determining a range estimation for activations of each iteration of the multiple iterations; andsetting a first activation encoding for at least one iteration based on the range estimation.12.The processor-implemented method of claim 11, further comprising setting at least a second activation encoding for at least a second iteration, the second activation encoding being different than the first activation encoding.13.The processor-implemented method of claim 11, further comprising selectively applying the first activation encoding or the at least the second activation encoding for each iteration of the multiple iterations based on a range distribution.14.The processor-implemented method of claim 11, further comprising setting a different activation encoding for each iteration of the multiple iterations.15.The processor-implemented method of claim 11, further comprising determining the range estimation for the activations of each iteration by removing at least one outlier value and computing the range estimation based on a minimum activation value and a maximum activation value of remaining activations of the iteration.16.The processor-implemented method of claim 15, further comprising identifying the at least one outlier value based on one or more thresholds.17.The processor-implemented method of claim 11, further comprising setting the first activation encoding based on a sensitivity metric.18.The processor-implemented method of claim 11, further comprising:storing input samples and corresponding activations for each iteration as a calibration dataset of the iteration;deriving a set of activation encodings for all of the multiple iterations; anddetermining the range estimation based on the set of activation encodings.19.The processor-implemented method of claim 11, further comprising setting the first activation encoding based on a hardware constraint.20.The processor-implemented method of claim 19, in which the first activation encoding is set based on an activation of a first iteration, a last iteration, an iteration having a shortest range or an iteration having a highest accuracy.21.An apparatus, comprising:at least one memory; andat least one processor coupled to the at least one memory, the at least one processor configured to:receive a diffusion model;process, by the diffusion model, a set of input samples over multiple iterations to generate a set of activations for each of the multiple iterations; andswitching between multiple activation encodings for the multiple iterations, wherein the multiple activation encodings are based on groupings of successively increasing activation ranges.22.The apparatus of claim 21, further comprising determining range estimation for each of the activation ranges based on the set of activations.23.The apparatus of claim 21, wherein the grouping are determined based on one or more thresholds.24.The apparatus of claim 21, wherein a number of grouping is based on a hardware constraint.
Citation Information
Patent Citations
Deep reinforcement learning for recursive segmentation
US20190287292A1
Image quantization using machine learning
US20230186429A1
Quantization range estimation for quantized training
US20240144017A1
Systems and methods for sports tracking using diffusion models
US20240221178A1