Iterative reconstruction network framework construction method based on reversible diffusion model

Abstract

The invention provides an iterative reconstruction network framework construction method based on a reversible diffusion model, and relates to the technical field of wireless communication, and the technical key points are that the construction method comprises the following steps: obtaining a sparse measurement value and environment information of a target area, carrying out the back projection of an adjoint operator passing through a measurement operator, and obtaining a sparse measurement value; generating initial estimation of an original channel gain map, and taking the initial estimation as an initial state of an iteration process; the initial estimation is sent to a cascade network composed of T reconstruction steps for iterative optimization; the framework comprises a core U-Net noise estimator weight, learnable parameters in a two-stage reversible framework and related parameters of a diffusion process, and end-to-end joint optimization is carried out through back propagation. According to the IDM-based channel gain map construction method provided by the invention, a pre-trained diffusion model is used to reconstruct a channel gain map from sparse measurement values. The generated channel gain map is highly matched with a real map visually.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an iterative reconstruction network framework construction method based on a reversible diffusion model. Background Technology

[0002] With the evolution of sixth-generation (6G) mobile communication networks, wireless communication systems are developing from simply providing connectivity services to possessing precise environmental awareness capabilities. Channel gain mapping (CGM), as a key tool for finely describing the propagation characteristics of wireless signals within a specific area, is a core technology for realizing cutting-edge applications such as intelligent resource allocation, network digital twin construction, UAV communication, and vehicle-to-everything (V2X) trajectory optimization.

[0003] Traditionally, constructing high-precision channel gain maps relies primarily on large-scale, high-density physical measurements, but this approach is costly and difficult to implement in real-world scenarios. As an alternative, ray-tracing-based simulation methods can generate maps, but they require enormous computational resources and time to consider complex multipath effects and obtain accurate results. To alleviate the burden of online computation, researchers have attempted methods such as Kriging interpolation, matrix completion, and dictionary learning; however, these methods lack the ability to learn from experience, limiting their generalization performance in dynamically changing network environments.

[0004] In recent years, deep learning models, especially convolutional neural networks (CNNs) and generative adversarial networks (GANs), have been widely used for the prediction and construction of channel gain maps. For example, the U-Net architecture combined with city maps is used to predict signal strength, or conditional generative adversarial networks (cGANs) are used to generate signal strength heatmaps. However, these methods generally have the following drawbacks: first, they are highly dependent on massive and detailed training datasets and environmental information, resulting in high data acquisition costs; second, the GAN framework itself has the problem of training instability, which may lead to artifacts and structural distortions in the generated results that do not conform to physical laws.

[0005] As a cutting-edge generative model, diffusion models have demonstrated outstanding performance in image generation tasks and have been initially applied to channel modeling in the field of wireless communication. The core idea is to progressively transform data into pure noise through a controlled random process, and then train a neural network to execute this process in reverse to reconstruct the original data. With this distributed denoising learning mechanism, diffusion models can generate new samples that are highly similar to the training data. However, despite their promising prospects, applying such large-scale generative models to channel gain map construction tasks still presents the problem of significant GPU memory consumption. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of constructing high-quality channel gain maps in the prior art, and the problem of huge memory consumption during large-scale generative model training.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for constructing an iterative reconstruction network framework based on a reversible diffusion model, the method comprising the following steps:

[0009] S1: Obtain sparse measurements of the target region And environmental information, sparse measurements Through its measurement operator The adjoint operator A back projection is performed to generate a preliminary estimate of the original channel gain map, which is then used as the initial state for the iterative process.

[0010] S2: The initial estimate obtained in S1 Feed into a cascaded network consisting of T reconstruction steps Iterative optimization is performed during this process.

[0011] S3: End-to-end optimization, the framework including the core U-Net noise estimator weights, learnable parameters in the two-level invertible architecture and related parameters of the diffusion process, are all jointly optimized end-to-end through backpropagation.

[0012] Preferably, the environmental information in S1 is a multi-channel environmental feature map, which includes a building distribution map showing the layout and height of buildings within the target area, and a base station location map showing the geographical location of base stations within the target area.

[0013] Preferably, in S2, in each iteration step t, the core network module Receive current state Furthermore, by integrating the environmental feature map C and sparse measurements Y as conditional information, a more accurate next state is generated by predicting and removing noise and artifact components from the current state. .

[0014] Preferably, in the single-iteration reconstruction step from time step t to t-1, the internal process deeply integrates the denoising mechanism of the diffusion model with the solution paradigm of the classical inverse problem, specifically including the following sub-steps:

[0015] Step 21: Noise Prediction and Initial Reconstruction Estimation: A neural network noise estimator is used to estimate the noise based on the current channel gain map. And the environmental information C, predict the noise component contained therein. ;

[0016] Step 22: Data consistency correction; based on the current state and predicted noise components The preliminary denoised map was calculated. The initial denoised map is then corrected to ensure consistency with the sparse measurements, resulting in a corrected map. ;

[0017] Step 23: Status Update: Based on the corrected map and predicted noise components Generate the channel gain map estimation for the next state. ;

[0018] The T-iteration reconstruction step is implemented through a first-level reversible architecture, which makes the transformation from state to state mathematically reversible, so that intermediate activation values ​​do not need to be cached when calculating gradients during backpropagation.

[0019] Repeat steps 21-23 T times.

[0020] Preferably, step 21 specifically involves using an adaptively modified U-Net neural network noise estimator. The network accepts channel gain map estimation of the current state. The tensor concatenated with environmental feature C along the channel dimension is used as input to predict... The noise component contained in : ; Calculate the complete channel gain map for noise-free conditions. Preliminary estimate .

[0021] Preferably, step 22 specifically involves: […]. Projecting the data onto the feasible solution space defined by the observed data, this correction step aims to minimize the fidelity loss. This is achieved through a single gradient descent step:

[0022]

[0023] in, It is a measurement operator The adjoint operator (which is the transpose in the real number field). To correct the step size, it is simplified to 1 here.

[0024] Preferably, step 23 specifically involves: based on the deterministic sampling formula of the Denoising Diffusion Implicit Model (DDIM), combined with a clear image that has undergone data consistency correction. and the noise component predicted in the first step Calculate a result that is closer to the real target The state of the next moment :

[0025] .

[0026] Preferably, the framework construction method introduces a multi-step iterative framework, drawing on the design concept of reversible residual networks, and incorporates each iterative step... Reconstruct it into a bijective operation.

[0027] Preferably, the input state at each step is split into a parallel tuple. , Main variable, As auxiliary variables; the forward propagation process follows the following reversible update rules:

[0028]

[0029] in, and It is a learnable scalar parameter that satisfies Used for dynamically balancing from the core computing path Output and auxiliary paths Information flow;

[0030] Due to the second rule The input was retained directly. Therefore, the entire transformation is reversible; its inverse transformation can be directly derived:

[0031]

[0032] During forward propagation, only the final output needs to be buffered. During backpropagation, the inverse transform can be used to... Begin by reconstructing the input from each step, layer by layer. It also calculates the intermediate activation values ​​and gradients in a timely manner; after the calculation is complete, the corresponding memory can be released.

[0033] Preferably, the framework construction method also introduces a noise estimator network, which groups multiple consecutive computational modules with unchanged feature map dimensions in the downsampling and upsampling paths of U-Net, and uses the same wiring technique as the first-level reversibility to encapsulate each group in a reversible unit to form an overall reversible computational flow.

[0034] Compared with the prior art, this application has the following beneficial effects:

[0035] This invention proposes an IDM-based channel gain map construction method that reconstructs a channel gain map from sparse measurements using a pre-trained diffusion model. The generated channel gain map visually closely matches the real map. The signal strength distribution, edge feature clarity, and richness of detail information in the map are all accurately reproduced. The introduction of a two-stage reversible architecture significantly reduces GPU memory overhead, solving the GPU memory bottleneck problem of end-to-end training of large iterative generative models on resource-constrained hardware. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the overall framework of the reversible diffusion model (IDM) in one embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the first-level reversible architecture-multi-step iterative framework in one embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating the application of the second-level reversible architecture within a U-Net network in one embodiment of the present invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to specific embodiments.

[0040] The core task of channel gain map reconstruction lies in recovering the original high-resolution channel gain map from a finite set of spatially sparsely distributed measurements. This process can be modeled as a constrained optimization problem:

[0041]

[0042] in:

[0043] It is an estimate of the complete channel gain map to be reconstructed, where H and W represent the height and width of the map, respectively.

[0044] It is a known sparse measurement matrix, consisting of M measurements, where .

[0045] It is a linear measurement operator that simulates the sparse sampling process of channel quality in the physical world.

[0046] The environmental information, which serves as a priori condition, is encoded into a multi-channel environmental feature map, which may specifically include: (1) building distribution map. (1) Describe the layout, outline and height of buildings in the area; (2) Base station location map It indicates the geographical location and transmission power of the base stations in the area.

[0047] It is a loss function used to quantify the estimated map. With real map The loss between them, in this embodiment, is adopted loss.

[0048] It is a data consistency constraint that requires the reconstruction results to be consistent with known physical measurements.

[0049] As a priori constraint, the reconstruction result is required. It must conform to the prior distribution of all possible channel maps under the given physical environment C.

[0050] The core of this application is to implicitly learn and capture the prior distribution by utilizing a specially designed generative diffusion model. .

[0051] To effectively solve the above optimization problems, please refer to Figure 1 This application provides an iterative reconstruction network framework based on a reversible diffusion model. The framework abandons the plug-and-play mode of treating the pre-trained noise estimator as an independent and immutable module in traditional methods. Instead, it reconstructs the entire diffusion sampling process, which includes T steps, into an end-to-end deterministic inverse problem solver that can be finely tuned for a specific reconstruction task.

[0052] Specifically, the framework construction method includes the following steps:

[0053] S1: Determine initialization:

[0054] Unlike traditional diffusion models that start sampling from pure Gaussian noise, this framework uses known physical measurement information for initialization.

[0055] Specifically, in one embodiment, sparse measurement values ​​and environmental information of the target area are first acquired, and the acquired sparse measurement values ​​are then... Through its measurement operator The adjoint operator A back projection is performed to generate a preliminary estimate of the original channel gain map, which is then used as the initial state for the iterative process. .

[0056] The environmental information is a multi-channel environmental feature map, including a building distribution map showing the layout and height of buildings within the target area, and a base station location map showing the geographical location of base stations within the target area.

[0057] This strategy injects the prior knowledge contained in the measurement into the model at the beginning of the iteration, providing a higher quality starting point for subsequent refined reconstruction.

[0058] In one embodiment, S1's

[0059] S2: Multi-step iterative optimization:

[0060] The initial estimate obtained in S1 Feed into a cascaded network consisting of T reconstruction steps Iterative optimization is performed. In each iteration step t, the core network module... Receive current state Furthermore, by integrating the environmental feature map C and sparse measurements Y as conditional information, a more accurate next state is generated by predicting and removing noise and artifact components from the current state. .

[0061] In one implementation, the single-iteration reconstruction step from time step t to t-1 internally integrates the denoising mechanism of the diffusion model with the solution paradigm of the classical inverse problem, specifically including the following sub-steps:

[0062] Step 21: Noise Prediction and Initial Reconstruction Estimation. This step aims to remove noise from the current state containing early information using generated priors.

[0063] First, an adaptively modified U-Net neural network noise estimator is used. The network accepts channel gain map estimation of the current state. The tensor concatenated with environmental feature C along the channel dimension is used as input to predict... The noise component contained in : Calculate the complete channel gain map for noise-free conditions. Preliminary estimate This process can be viewed as a single step in solving the inverse diffusion process:

[0064]

[0065] in, For a predefined noise scheduling coefficient, it determines the noise level at time step. The relative ratio of signal to noise. This represents the model at the current step. Based on the most likely sharp image estimated from the generative prior.

[0066] In one embodiment, the network weights of the neural network noise estimator are initialized by loading the weights of a large-scale diffusion model pre-trained on a massive natural image dataset, and are fine-tuned end-to-end for the channel gain map construction task.

[0067] Step 22: Data consistency correction. Although While it may visually possess the characteristics of a high-quality channel gain map, it cannot guarantee complete compliance with known physical measurement constraints. .

[0068] To address this issue, the initial estimate needs to be corrected by projecting it onto a feasible solution space defined by the observed data. This correction step aims to minimize the fidelity loss. This is achieved through a single gradient descent step:

[0069]

[0070] in, It is a measurement operator The adjoint operator (which is the transpose in the real number field). To correct the step size, it is simplified to 1 here. This step forces the current estimate to satisfy physical constraints. This anchors the estimates of the generative model to physical reality. The corrected estimates In keeping While maintaining a good structure, it also maintains a high degree of consistency with the sparse measurement Y. This step forces the current estimate to meet physical constraints, anchoring the generative model's estimate to physical reality.

[0071] Step 23: Status update.

[0072] Again, based on the deterministic sampling formula of the Denoising Diffusion Implicit Model (DDIM), combined with a clear image after data consistency correction... and the noise component predicted in the first step Calculate a result that is closer to the real target The state of the next moment :

[0073]

[0074] This step re-encodes the corrected information back into the diffusion sampling process, providing a higher-quality input with fewer artifacts for the next iteration.

[0075] Repeat steps 21-23 T times, so that the model can gradually start from a coarse initial estimate, iteratively remove noise and artifacts, restore map details, and finally generate a high-quality channel gain map that closely matches the real physical environment.

[0076] This process embodies the synergistic effect of generative prior guidance and physical constraint correction. The generative model provides a strong prior about a reasonable channel gain map, and data consistency correction ensures that this reasonable result must conform to objective measurement data. End-to-end training enables the noise estimator... It can learn a priori knowledge that already predicts subsequent correction steps, thereby making the collaboration between the two more efficient.

[0077] In one embodiment, to overcome the memory bottleneck of end-to-end training of large-scale diffusion models, this application introduces a two-level reversible architecture. Its core lies in making the computation process mathematically reversible through structural modifications. This allows for real-time recovery of intermediate activation values ​​from forward propagation via its reverse computation path during backpropagation, eliminating the need for caching. This reduces the memory complexity of end-to-end training from being linearly related to the number of iterations T. Reduced to constant level .

[0078] Specifically, the first-level invertibility is a multi-step iterative framework; please refer to [link / reference]. Figure 2 Drawing inspiration from the design principles of reversible residual networks, each iteration step... This is refactored into a bijective operation. Specifically, the input state at each step is split into a parallel tuple. , Main variable, These are auxiliary variables. The forward propagation process follows the following reversible update rules:

[0079]

[0080] in, and It is a learnable scalar parameter that satisfies Used for dynamically balancing from the core computing path Output and auxiliary paths Information flow. Due to the second rule. The input was retained directly. Therefore, the entire transformation is reversible. Its inverse transformation can be directly derived:

[0081]

[0082] During forward propagation, only the final output needs to be buffered. During backpropagation, the inverse transform can be used to... Begin by reconstructing the input from each step, layer by layer. It calculates the intermediate activation values ​​and gradients within the graph, and releases the corresponding memory after calculation. In this way, the high-dimensional activation map generated during the iteration steps does not need to be persistently stored in GPU memory, thus achieving constant GPU memory usage relative to the iteration number T.

[0083] The second-level invertibility is achieved through a noise estimator network. Although the first-level invertible framework solves the memory growth problem between iteration steps, the noise estimator network still faces challenges in single-step iterations. (Typically, the U-Net architecture) still generates significant memory usage. To further reduce memory usage, this framework extends the reversible design concept to the interior of U-Net, building a second level of reversibility. For example... Figure 3 As shown.

[0084] Specifically, in one embodiment, the downsampling and upsampling paths of U-Net contain multiple consecutive computational modules (such as residual blocks and self-attention blocks) that maintain the same feature map dimensions. In this application, these equal-dimensional transformation modules are grouped, and each group is encapsulated within a reversible unit using the same wiring technique as the first-level reversibility. These reversible units are chained together to form an overall reversible computational flow. Structural operations that change feature dimensions (such as convolutional downsampling with stride and transposed convolutional upsampling) are retained outside the reversible groups. This design allows for the computation of gradients within U-Net during backpropagation without caching intermediate activation values ​​within these groups; instead, they can be reconstructed on demand through their reverse computational paths.

[0085] S3: End-to-end optimization.

[0086] The entire framework, including the core U-Net noise estimator weights, the learnable parameters in the two-stage invertible architecture, and the relevant parameters of the diffusion process, is jointly optimized end-to-end through backpropagation. The optimization objective is to minimize the reconstruction result of the network's final output. Compared with the actual channel gain map Between Loss. In this way, all components of the model are directly adjusted to the final reconstruction goal, eliminating the task deviation problem between the generation prior and the reconstruction task in traditional methods.

[0087] Through this two-stage reversible system, this invention enables end-to-end training of complex models. Table 1 below shows a comparison of peak GPU memory usage between the reversible and non-reversible architectures at different iteration steps T.

[0088] Table 1 Comparison of peak memory usage for reversible and irreversible architectures at different iteration steps.

[0089] T (number of steps) Peak video memory (GB) for reversible models Peak video memory (GB) for the irreversible model Video memory saving rate 1 5.22 16.67 68.69% 2 6.9 33.33 79.30% 3 6.89 49.0251 85.59% 4 6.9 67.7128 89.34%

[0090] Experimental data clearly shows that the reversible architecture can significantly reduce memory consumption during training. In the standard irreversible training mode, when T=3, the peak memory usage of the irreversible model reaches 49.02GB, exceeding the memory capacity of the NVIDIA RTX 4090 GPU used in the experiment. However, after introducing the two-stage reversible architecture, the memory usage drops sharply to 6.89GB, a saving of 85.95%. This is not only a performance optimization but also a key enabling technology. It fundamentally solves the bottleneck problem of end-to-end training of large iterative generative models on resource-constrained hardware.

[0091] Exemplary embodiments and performance results:

[0092] To verify the effectiveness of this invention, in a specific embodiment, all experiments were conducted based on the publicly released RadioMap3DSeer dataset, which contains 701 maps with independent geographic information for multiple cities, with a carrier frequency of 3.5 GHz. The model was implemented based on the PyTorch framework, using the Adam optimizer for end-to-end fine-tuning, with the iteration step T set to 3.

[0093] As shown in Table 2, the reversible diffusion model (IDM) proposed in this invention achieves the best performance across all quantitative metrics, comprehensively outperforming methods including U-Net-based supervised learning (RadioUNet), GAN-based generative methods (REM-GAN), and other advanced diffusion model methods (RadioDiff). In particular, it outperforms by nearly 3 dB in PSNR. This indicates that the reversible diffusion model is superior not only in the accuracy of predicting signal strength values.

[0094] Table 2 Quantitative Comparison

[0095] method PSNR SSIM NMSE RMSE RadioUNet 28.79 0.8671 0.01308 0.0318 RME-GAN 23.68 0.7289 0.12487 0.0982 RadioDiff 35.32 0.9435 0.0041 0.0191 This article's method 38.02 0.9634 0.00196 0.0123

[0096] The channel gain map generated by this invention visually closely matches the real map. The signal strength distribution, edge feature clarity, and richness of detail in the map are all accurately reproduced. In contrast, the reconstruction results of RadioUNet exhibit significant blurring effects in areas with abrupt signal changes, such as building edges. Furthermore, the results of RME-GAN introduce numerous artifacts and structural distortions that do not conform to physical laws. This demonstrates that this invention, through powerful generative priors, can implicitly learn and follow complex physical propagation laws, thereby generating an accurate and realistic channel gain map.

[0097] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the channel gain map construction method based on the reversible diffusion model.

[0098] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the channel gain map construction method based on the reversible diffusion model.

[0099] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

Claims

1. A method for constructing an iterative reconstruction network framework based on a reversible diffusion model, characterized in that: The construction method includes the following steps: S1: Obtain sparse measurements of the target region And environmental information, sparse measurements Through its measurement operator The adjoint operator A back projection is performed to generate a preliminary estimate of the original channel gain map, which is then used as the initial state for the iterative process. ; S2: The initial estimate obtained in S1 Feed into a cascaded network consisting of T reconstruction steps Iterative optimization is performed during this process. S3: End-to-end optimization, the framework including the core U-Net noise estimator weights, learnable parameters in the two-level invertible architecture, and relevant parameters of the diffusion process, are all jointly optimized end-to-end through backpropagation.

2. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 1, characterized in that: The environmental information in S1 is a multi-channel environmental feature map, which includes a building distribution map showing the layout and height of buildings within the target area, and a base station location map showing the geographical location of base stations within the target area.

3. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 2, characterized in that: In S2, at each iteration step t, the core network module Receive current state Furthermore, by integrating the environmental feature map C and sparse measurements Y as conditional information, a more accurate next state is generated by predicting and removing noise and artifact components from the current state. .

4. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 1, characterized in that: The single-iteration reconstruction step from time step t to t-1 internally integrates the denoising mechanism of the diffusion model with the solution paradigm of the classical inverse problem, specifically including the following sub-steps: Step 21: Noise Prediction and Initial Reconstruction Estimation: A neural network noise estimator is used to estimate the noise based on the current channel gain map. And the environmental information C, predict the noise component contained therein. ; Step 22: Data consistency correction; based on the current state and predicted noise components The preliminary denoised map was calculated. The initial denoised map is then corrected to ensure consistency with the sparse measurements, resulting in a corrected map. ; Step 23: Status Update: Based on the corrected map and predicted noise components Generate the channel gain map estimation for the next state. ; The T-iteration reconstruction step is implemented through a first-level reversible architecture, which makes the transformation from state to state mathematically reversible, so that intermediate activation values ​​do not need to be cached when calculating gradients during backpropagation. Repeat steps 21-23 T times.

5. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 1, characterized in that: Step 21 specifically involves using an adaptively modified U-Net neural network noise estimator. The network accepts channel gain map estimation of the current state. The tensor concatenated with environmental feature C along the channel dimension is used as input to predict... The noise component contained in : ; Calculate the complete channel gain map for noise-free conditions. Preliminary estimate .

6. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 4, characterized in that: The specific steps of step 22 are as follows: ... Projecting the data onto the feasible solution space defined by the observed data, this correction step aims to minimize the fidelity loss. This is achieved through a single gradient descent step: in, It is a measurement operator The adjoint operator (which is the transpose in the real number field). To correct the step size, it is simplified to 1 here.

7. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 4, characterized in that: Step 23 specifically involves: based on the deterministic sampling formula of the Denoising Diffusion Implicit Model (DDIM), combined with a clear image that has undergone data consistency correction. and the noise component predicted in the first step Calculate a result that is closer to the real target The state of the next moment : 。 8. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 4, characterized in that: The framework construction method introduced a multi-step iterative framework, drawing on the design concept of invertible residual networks, and defining each iteration step... Reconstruct it into a bijection operation.

9. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 8, characterized in that: The input state at each step is split into a parallel tuple. , Main variable, As auxiliary variables; the forward propagation process follows the following reversible update rules: in, and It is a learnable scalar parameter that satisfies Used for dynamically balancing from the core computing path Output and auxiliary paths Information flow; Due to the second rule The input was retained directly. Therefore, the entire transformation is reversible; its inverse transformation can be directly derived: During forward propagation, only the final output needs to be buffered. During backpropagation, the inverse transform can be used to... Begin by reconstructing the input from each step, layer by layer. It also calculates the intermediate activation values ​​and gradients in a timely manner; after the calculation is complete, the corresponding memory can be released.

10. The method for constructing an iterative reconstruction network framework based on a reversible diffusion model according to claim 5, characterized in that: The framework construction method also introduces a noise estimator network, which groups multiple consecutive computational modules with unchanged feature map dimensions in the downsampling and upsampling paths of U-Net, and uses the same wiring technique as the first-level reversibility to encapsulate each group in a reversible unit to form an overall reversible computational flow.

Images