A modular function scheduling and distribution system for image production and product sales

Abstract

The application relates to the technical field of image production and product sales, in particular to a modular function scheduling and distribution system for image production and product sales, which comprises a data acquisition module used for extracting underlying rendering attributes from a scene file of a to-be-produced image task; a feature modeling module used for calling the underlying rendering attributes to determine calculation resistance; a value calculation module used for calculating the instantaneous value density corresponding to each modular task; a scheduling execution module used for calculating a density difference value; when the density difference value is greater than a task switching cost, triggering a resource reorganization operation; otherwise, maintaining the current running queue state; and a closed-loop feedback module used for calculating a prediction deviation between actual sales data and predicted sales data, and correcting the calculation weight parameter in the feature modeling module based on the prediction deviation; the application solves the problem that high calculation power is occupied by low value tasks under a traditional static priority strategy, and significantly improves the calculation resource utilization rate and business return rate.

Description

Technical Field

[0001] This invention relates to the field of image production and product sales technology, specifically a modular function scheduling and allocation system for image production and product sales. Background Technology

[0002] With the rapid development of the digital content industry, the linkage between video production and product sales is becoming increasingly close, and the demand for batch rendering for e-commerce scenarios is showing explosive growth.

[0003] Currently, existing rendering task scheduling typically employs first-in-first-out (FIFO) or static priority-based processing strategies, lacking effective utilization of real-time unstructured behavioral data from sales terminals. Traditional methods struggle to correlate market momentum with underlying rendering computational resistance, failing to quantify the instantaneous value density of each modular task. This resource allocation model often results in high-computing resources being consumed by low-commercial-value tasks, or delays in the production of high-demand product images due to resource competition. Furthermore, the lack of a closed-loop feedback correction mechanism based on actual sales data prevents the scheduling system from adaptively adjusting weight parameters based on prediction deviations, leading to poor overall output efficiency. Therefore, how to comprehensively consider real-time market momentum and computational resource consumption to construct a dynamic value assessment and intelligent scheduling system, thereby achieving a dual improvement in computational resource utilization and commercial return on investment, has become a pressing issue in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a modular function scheduling and allocation system for image production and product sales. Specifically, the technical solution of this invention includes:

[0005] The data acquisition module is used to collect unstructured behavioral data streams from sales terminals and extract underlying rendering attributes from scene files of the image production task.

[0006] The feature modeling module is used to invoke unstructured behavioral data streams to construct market heat potential and invoke underlying rendering attributes to determine computational resistance.

[0007] The value calculation module is used to calculate the instantaneous value density corresponding to each modular task based on market heat potential and computational resistance.

[0008] The scheduling and execution module is used to periodically scan the waiting queue and the running queue, calculate the density difference between the highest value task in the waiting queue and the lowest value task in the running queue; when the density difference is greater than the task switching cost, a resource reorganization operation is triggered; when the density difference is less than or equal to the task switching cost, the current running queue state is maintained.

[0009] The closed-loop feedback module is used to track actual sales data after the images are uploaded, calculate the prediction deviation between the actual sales data and the predicted sales data, and correct the calculated weight parameters in the feature modeling module based on the prediction deviation.

[0010] Preferably, the feature modeling module constructs market heat potential based on unstructured behavioral data streams, including:

[0011] Call the unstructured behavior data stream, which includes the number of clicks on the details page, the frequency of adding items to the shopping cart, and the duration of page dwell.

[0012] Using a decaying memory integral model, discrete details page click counts, shopping cart addition frequency, and page dwell time are mapped to continuous potential energy scalars.

[0013] By introducing a time decay constant as a normalization coefficient and combining it with a potential energy scalar, the weighted relative popularity at the current moment is calculated to generate market popularity potential energy.

[0014] Preferably, the feature modeling module determines the computational resistance based on the underlying rendering properties, including:

[0015] Call the underlying rendering properties, which include the estimated video memory usage and the estimated rendering time of the task;

[0016] Get the total physical video memory of the current rendering node as the normalization denominator, and get the baseline reference duration as the normalization baseline;

[0017] Based on the generalized resource resistance model, the first ratio of the estimated video memory usage to the total physical video memory and the second ratio of the estimated rendering time to the baseline reference time are calculated.

[0018] The first ratio and the second ratio are weighted and summed, and a memory sensitivity index is introduced to perform nonlinear correction on the summation result in order to generate computational resistance.

[0019] Preferably, the value calculation module calculates the instantaneous value density corresponding to each modular task based on market momentum and computational resistance, including:

[0020] Utilize market momentum and calculate resistance;

[0021] Obtain the system's global load adjustment coefficient;

[0022] Calculate the ratio of market heat potential energy to calculated resistance, and multiply the ratio by the system global load adjustment coefficient to generate instantaneous value density;

[0023] The system global load adjustment coefficient is configured to automatically reduce its value when the system load exceeds a preset threshold in order to suppress the submission of low-value tasks.

[0024] Preferably, the scheduling execution module determines the task switching cost, including:

[0025] Obtain the time required to save the current task status and the baseline reference time, and calculate the time cost ratio;

[0026] Obtain the amount of video memory data that needs to be swapped to disk and the total physical video memory of the current rendering node, and calculate the space cost ratio.

[0027] The system introduces a jitter suppression weight to weight the time cost ratio, and then adds the weighted time cost ratio to the space cost ratio to generate the task switching cost.

[0028] Preferably, the scheduling execution module triggers a resource reorganization operation, including:

[0029] Identify the target task with the lowest instantaneous value density in the run queue;

[0030] Implement a service degradation strategy to dynamically adjust the rendering resolution of the target task from high resolution to low resolution;

[0031] Release graphics processor computing units that are idle due to the adjustment of rendering resolution, and allocate graphics processor computing units to the tasks with the highest instantaneous value density in the waiting queue.

[0032] Preferably, the closed-loop feedback module is based on the calculated weight parameters in the prediction bias correction feature modeling module, including:

[0033] Determine the prediction error gradient based on the prediction bias;

[0034] The online machine learning algorithm is invoked to reverse the weights of click behavior and add-to-cart behavior using the prediction error gradient;

[0035] Normalization constraints are applied to the corrected click behavior weights and add-to-cart behavior weights to generate updated calculated weight parameters, which are then fed back to the feature modeling module.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. This system implements dynamic scheduling optimization based on commercial value; it constructs market heat potential by collecting unstructured behavioral data streams from sales terminals and determines computational resistance by combining underlying rendering attributes, thereby calculating the instantaneous value density of tasks; the scheduling execution module decides whether to trigger resource reorganization based on the comparison between density difference and task switching cost, thus ensuring that high commercial value tasks can obtain resources first, solving the problem of high computing power being occupied by low value tasks under the traditional static priority strategy, and significantly improving the utilization rate of computing resources and commercial return rate;

[0038] 2. This system improves market response speed and resource matching efficiency; the value calculation module introduces a global load adjustment coefficient, which automatically reduces the value when the system load is high to suppress the submission of low-value tasks and prevent system overload; at the same time, the scheduling and execution module identifies the lowest-value tasks and performs service degradation during resource reorganization, and allocates the released resources to the highest-value tasks, realizing dynamic and refined resource rollover and allocation, ensuring that the video production of high-demand products can quickly respond to market demands without lag;

[0039] 3. This system provides a quantitative basis for task switching decisions; the scheduling and execution module introduces the concept of task switching cost, which is quantified by the weighted ratio of time cost and space cost; preemption is only executed when the density difference between the highest value task in the waiting queue and the lowest value task in the running queue is greater than the switching cost, thus avoiding system jitter and additional overhead caused by frequent and inefficient task switching, making resource reorganization operations more economical and stable;

[0040] 4. This system possesses sales data-driven adaptive capabilities; after the images are uploaded, the closed-loop feedback module can track the deviation between actual sales data and predicted sales data, and determine the error gradient based on this prediction deviation; online machine learning algorithms are used to perform reverse correction and normalization constraint processing on the calculated weight parameters in the feature modeling module; this mechanism enables the system to adaptively adjust weights according to actual business conversion effects, solving the problem of traditional systems lacking a sales closed-loop feedback correction mechanism, and ensuring the accuracy and long-term benefits of the task value assessment model. Attached Figure Description

[0041] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0042] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0044] Example 1:

[0045] Please see Figure 1 A modular function scheduling and allocation system for image production and product sales, comprising:

[0046] The data acquisition module is used to collect unstructured behavioral data streams from sales terminals and extract underlying rendering attributes from scene files of the image production task.

[0047] The feature modeling module is used to invoke unstructured behavioral data streams to construct market heat potential and invoke underlying rendering attributes to determine computational resistance.

[0048] The value calculation module is used to calculate the instantaneous value density corresponding to each modular task based on market heat potential and computational resistance.

[0049] The scheduling and execution module is used to periodically scan the waiting queue and the running queue, calculate the density difference between the highest value task in the waiting queue and the lowest value task in the running queue; when the density difference is greater than the task switching cost, a resource reorganization operation is triggered; when the density difference is less than or equal to the task switching cost, the current running queue state is maintained.

[0050] The closed-loop feedback module is used to track actual sales data after the images are uploaded, calculate the prediction deviation between the actual sales data and the predicted sales data, and correct the calculated weight parameters in the feature modeling module based on the prediction deviation.

[0051] This embodiment introduces a modular function scheduling and allocation system for image production and product sales. The system collects unstructured behavioral data streams from sales terminals via a data acquisition module. These data streams originate from e-commerce front-end tracking logs and are aggregated via a real-time computing engine. Simultaneously, the data acquisition module extracts underlying rendering attributes from the scene files of the image production tasks. These attributes are analyzed by the rendering engine's pre-computation module. A feature modeling module uses the unstructured behavioral data streams to construct market heat potential and uses the underlying rendering attributes to determine computational resistance. A value calculation module calculates the instantaneous value density corresponding to each modular task based on the market heat potential and computational resistance. A scheduling execution module periodically scans the waiting queue and the running queue, calculating the density difference between the highest-value task in the waiting queue and the lowest-value task in the running queue. When the density difference exceeds the task switching cost, a resource reorganization operation is triggered. When the density difference is less than or equal to the task switching cost, the current running queue state is maintained. A closed-loop feedback module tracks actual sales data after the images are uploaded, calculates the prediction deviation between the actual sales data and the predicted sales data, and corrects the calculation weight parameters in the feature modeling module based on the prediction deviation.

[0052] Example 2:

[0053] The feature modeling module constructs market trend momentum based on unstructured behavioral data streams, including:

[0054] Call the unstructured behavior data stream, which includes the number of clicks on the details page, the frequency of adding items to the shopping cart, and the duration of page dwell.

[0055] Using a decaying memory integral model, discrete details page click counts, shopping cart addition frequency, and page dwell time are mapped to continuous potential energy scalars.

[0056] By introducing a time decay constant as a normalization coefficient and combining it with a potential energy scalar, the weighted relative popularity at the current moment is calculated to generate market popularity potential energy.

[0057] The feature modeling module constructs market popularity potential based on unstructured behavioral data streams. This module calls upon unstructured behavioral data streams containing details page click counts, shopping cart addition frequency, and page dwell time. The system utilizes a decaying memory integral model to map discrete details page click counts, shopping cart addition frequency, and page dwell time into continuous potential energy scalars. A time decay constant is introduced as a normalization coefficient, and combined with the potential energy scalar, the weighted relative popularity at the current moment is calculated to generate market popularity potential. The specific calculation logic is as follows:

[0058]

[0059] In this formula, For the first Each product The normalized market potential energy at any given moment is a dimensionless scalar. Let be the time decay constant, with dimensions . It also serves as a normalization coefficient in the formula, used to offset the integral variable. The resulting time dimension To ensure the consistency of the dimensions of the output values; and Representing the integration variable time respectively Click rate and add-to-cart rate at any given time, both in units of 1. ; and These are the maximum click rate and maximum add-to-cart rate, respectively, determined based on historical data. When historical data is empty, a preset non-zero initial baseline value is set to avoid division by zero error and used as the normalization denominator. and These are the click behavior weight coefficient and the add-to-cart behavior weight coefficient, respectively, and they satisfy the following conditions: ; Indicates the current time. Represents the integration variable, time. Indicates the length of the integration window.

[0060] Example 3:

[0061] The feature modeling module determines computational resistance based on underlying rendering properties, including:

[0062] Call the underlying rendering properties, which include the estimated video memory usage and the estimated rendering time of the task;

[0063] Get the total physical video memory of the current rendering node as the normalization denominator, and get the baseline reference duration as the normalization baseline;

[0064] Based on the generalized resource resistance model, the first ratio of the estimated video memory usage to the total physical video memory and the second ratio of the estimated rendering time to the baseline reference time are calculated.

[0065] The first ratio and the second ratio are weighted and summed, and a memory sensitivity index is introduced to perform nonlinear correction on the summation result in order to generate computational resistance.

[0066] The feature modeling module determines computational resistance based on underlying rendering attributes. It calls upon underlying rendering attributes containing estimated task memory usage and estimated task rendering time; it obtains the total physical memory of the current rendering node as the normalization denominator and the baseline reference time as the normalization baseline; based on the generalized resource resistance model, it calculates a first ratio of estimated task memory usage to total physical memory and a second ratio of estimated task rendering time to the baseline reference time; it then performs a weighted summation of the first and second ratios and introduces a memory sensitivity index to non-linearly correct the summation result, thereby generating computational resistance. The specific calculation logic is as follows:

[0067]

[0068] In this formula, For the first The normalized computational resistance of a rendering task is a dimensionless scalar. Estimated memory usage for tasks derived from scene analysis; The total physical video memory of the current rendering node; This is the estimated rendering time for a task based on the average of similar historical tasks. The baseline reference duration, such as the rendering time per standard frame; and These are the weights for memory resource scarcity and time resource scarcity, respectively, and satisfy the following conditions: ; This is the memory sensitivity index, and its value is greater than or equal to... This is used to reflect the rigid constraint characteristics of video memory resources, that is, when the video memory usage approaches the physical limit, the resistance value increases sharply and non-linearly; when the task estimates the video memory usage... Exceeding the total physical video memory At that time, resistance value It will be forcibly set to infinity or trigger the high task switching cost calculation logic in Implementation 5.

[0069] Example 4:

[0070] The value calculation module, based on market momentum and computational resistance, calculates the instantaneous value density for each modular task, including:

[0071] Utilize market momentum and calculate resistance;

[0072] Obtain the system's global load adjustment coefficient;

[0073] Calculate the ratio of market heat potential energy to calculated resistance, and multiply the ratio by the system global load adjustment coefficient to generate instantaneous value density;

[0074] The system global load adjustment coefficient is configured to automatically reduce its value when the system load exceeds a preset threshold in order to suppress the submission of low-value tasks.

[0075] The value calculation module calculates the instantaneous value density for each modular task based on market heat potential energy and computational resistance. The module then calls upon market heat potential energy and computational resistance data; obtains the system's global load adjustment coefficient; calculates the ratio of market heat potential energy to computational resistance; and multiplies this ratio by the system's global load adjustment coefficient to generate the instantaneous value density. The specific calculation logic is as follows:

[0076]

[0077] In this formula, For corresponding products Task The instantaneous value density is a dimensionless scalar; This refers to the market momentum energy calculated above; The calculated resistance is the resistance obtained from the aforementioned calculation; This refers to the system's global load adjustment factor; the system's global load adjustment factor. It is configured to automatically reduce the value when the system load exceeds a preset threshold in order to suppress the submission of low-value tasks; the preset threshold is set to the historical system crash load value. .

[0078] Example 5:

[0079] The scheduling and execution module determines the task switching cost, including:

[0080] Obtain the time required to save the current task status and the baseline reference time, and calculate the time cost ratio;

[0081] Obtain the amount of video memory data that needs to be swapped to disk and the total physical video memory of the current rendering node, and calculate the space cost ratio.

[0082] The system introduces a jitter suppression weight to weight the time cost ratio, and then adds the weighted time cost ratio to the space cost ratio to generate the task switching cost.

[0083] The scheduling and execution module determines the task switching cost; the scheduling and execution module obtains the time required to save the current task state. Compared with the benchmark reference duration Calculate the time-cost ratio; obtain the amount of video memory data that needs to be swapped to disk. Total physical video memory of the current rendering node The space cost ratio is calculated; a jitter suppression weight set by the system is introduced to weight the time cost ratio, and the weighted time cost ratio is added to the space cost ratio to generate the task switching cost. The specific calculation logic is as follows:

[0084]

[0085] In this formula, The task switching cost is a dimensionless scalar. The jitter suppression weights set for the system; The time required to save the current task context; The amount of video memory data that needs to be swapped to disk during the switching process; and The definition remains consistent with the aforementioned embodiments; in the scheduling judgment logic, only when the instantaneous value density of the highest value task in the waiting queue is... Instantaneous value density of the lowest value task in the running queue The difference is greater than Only when the time is right will the preemption operation be triggered.

[0086] Example 6:

[0087] The scheduling execution module triggers resource reorganization operations, including:

[0088] Identify the target task with the lowest instantaneous value density in the run queue;

[0089] Implement a service degradation strategy to dynamically adjust the rendering resolution of the target task from high resolution to low resolution;

[0090] Release the video memory and computing resources freed up due to the adjustment of rendering resolution, and allocate the freed-up resources to the tasks with the highest instantaneous value density in the waiting queue.

[0091] The scheduling and execution module triggers a resource reorganization operation; the scheduling and execution module identifies the target task with the lowest instantaneous value density in the running queue; it executes a service degradation strategy, dynamically adjusting the rendering resolution of the target task from high resolution to low resolution; it releases the video memory space and computing power resources freed up due to the adjustment of the rendering resolution, and allocates the freed-up resources to the task with the highest instantaneous value density in the waiting queue; high resolution refers to the standard resolution required for the final delivery of the product, and low resolution refers to the resolution after reducing the sampling rate for previewing or proxy production; for example, setting it to 50% of the original resolution in length and width or the 720p standard.

[0092] Example 7:

[0093] The closed-loop feedback module calculates weight parameters based on the prediction bias correction feature modeling module, including:

[0094] Determine the prediction error gradient based on the prediction bias;

[0095] The online machine learning algorithm is invoked to reverse the weights of click behavior and add-to-cart behavior using the prediction error gradient;

[0096] Normalization constraints are applied to the corrected click behavior weights and add-to-cart behavior weights to generate updated calculated weight parameters, which are then fed back to the feature modeling module.

[0097] The closed-loop feedback module corrects the calculated weight parameters in the feature modeling module based on the prediction bias; it determines the prediction error gradient based on the prediction bias; it calls an online machine learning algorithm to use the prediction error gradient to reverse-correct the click behavior weight and add-to-cart behavior weight; it performs normalization constraint processing on the corrected click behavior weight and add-to-cart behavior weight, generates updated calculated weight parameters, and feeds them back to the feature modeling module. The logic for calculating the weight correction value is as follows:

[0098]

[0099] In this formula, This is the adjustment amount for the click behavior weight; As the learning rate, it possesses the inverse square property of monetary units to ensure the correction amount. It is a dimensionless scalar; This refers to actual sales revenue; To predict sales revenue, the system calculates the result based on the product of historical conversion rates and market momentum. ,in This is the historical conversion coefficient; To predict the partial derivative of sales revenue with respect to the weight of click behavior, one can use the formula... In The updated click behavior weights are obtained by differentiating the terms. Original weights With correction amount The sum and passing through the interval Truncation processing; updated weighting of add-to-cart behavior Set as .

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A modular function scheduling and allocation system for image production and product sales, characterized in that, include: The data acquisition module is used to collect unstructured behavioral data streams from sales terminals and extract underlying rendering attributes from scene files of the image production task. The feature modeling module is used to invoke unstructured behavioral data streams to construct market heat potential and invoke underlying rendering attributes to determine computational resistance. The value calculation module is used to calculate the instantaneous value density corresponding to each modular task based on market heat potential and computational resistance. The scheduling and execution module is used to periodically scan the waiting queue and the running queue, calculate the density difference between the highest value task in the waiting queue and the lowest value task in the running queue, and trigger a resource reorganization operation when the density difference is greater than the task switching cost. When the density difference is less than or equal to the task switching cost, maintain the current running queue state; The closed-loop feedback module is used to track actual sales data after the images are uploaded, calculate the prediction deviation between the actual sales data and the predicted sales data, and correct the calculation weight parameters in the feature modeling module based on the prediction deviation. The feature modeling module constructs market heat potential based on unstructured behavioral data streams, including: Call the unstructured behavior data stream, which includes the number of clicks on the details page, the frequency of adding items to the shopping cart, and the duration of page dwell. Using a decaying memory integral model, discrete details page click counts, shopping cart addition frequency, and page dwell time are mapped to continuous potential energy scalars. By introducing a time decay constant as a normalization coefficient and combining it with a potential energy scalar, the weighted relative popularity at the current moment is calculated to generate market popularity potential energy. The feature modeling module determines computational resistance based on underlying rendering attributes, including: Call the underlying rendering properties, which include the estimated video memory usage and the estimated rendering time of the task; Get the total physical video memory of the current rendering node as the normalization denominator, and get the baseline reference duration as the normalization baseline; Based on the generalized resource resistance model, the first ratio of the estimated video memory usage to the total physical video memory and the second ratio of the estimated rendering time to the baseline reference time are calculated. The first ratio and the second ratio are weighted and summed, and a memory sensitivity index is introduced to perform nonlinear correction on the summation result in order to generate computational resistance. The value calculation module calculates the instantaneous value density corresponding to each modular task based on market momentum and computational resistance, including: Utilize market momentum and calculate resistance; Obtain the system's global load adjustment coefficient; Calculate the ratio of market heat potential energy to calculated resistance, and multiply the ratio by the system global load adjustment coefficient to generate instantaneous value density; The system global load adjustment coefficient is configured to automatically reduce its value when the system load exceeds a preset threshold in order to suppress the submission of low-value tasks.

2. The modular function scheduling and allocation system for image production and product sales according to claim 1, characterized in that, The scheduling execution module determines the task switching cost, including: Obtain the time required to save the current task status and the baseline reference time, and calculate the time cost ratio; Obtain the amount of video memory data that needs to be swapped to disk and the total physical video memory of the current rendering node, and calculate the space cost ratio. The system introduces a jitter suppression weight to weight the time cost ratio, and then adds the weighted time cost ratio to the space cost ratio to generate the task switching cost.

3. A modular function scheduling and allocation system for image production and product sales according to claim 1, characterized in that, The scheduling execution module triggers a resource reorganization operation, including: Identify the target task with the lowest instantaneous value density in the run queue; Implement a service degradation strategy to dynamically adjust the rendering resolution of the target task from high resolution to low resolution; Release the video memory and computing resources freed up due to the adjustment of rendering resolution, and allocate the freed-up resources to the tasks with the highest instantaneous value density in the waiting queue.

4. A modular function scheduling and allocation system for image production and product sales according to claim 1, characterized in that, The closed-loop feedback module is based on the calculated weight parameters in the prediction bias correction feature modeling module, including: Determine the prediction error gradient based on the prediction bias; The online machine learning algorithm is invoked to reverse the weights of click behavior and add-to-cart behavior using the prediction error gradient; Normalization constraints are applied to the corrected click behavior weights and add-to-cart behavior weights to generate updated calculated weight parameters, which are then fed back to the feature modeling module.

Images