深度学习在各个领域百花齐放的今天，如何使其更好地服务于建筑设计和实践是亟待解决的问题。本文试图建立建筑设计知识经验和建筑相关深度学习模型的双向连通：以建筑设计知识经验为先验指导深度模型框架的构建，以及以深度模型的解释性分析来验证、获取更多的经验知识。建筑平面图在信息表达和生成流程方面都与计算机视觉、图形学关注的传统领域有较大不同。因此，直接套用成熟的深度模型并不是建筑领域智能设计的最优方案。我们在理解建筑设计问题的基础上，构建了与矢量住宅平面图生成任务相适配的深度模型FloorplanGAN。这是一个生成式对抗模型框架，其结构包含了矢量生成器、可微渲染器和栅格判别器。生成器中嵌入了自注意力机制以提高建模房间相互关系的能力。完成模型训练后，我们以多角度的评价体系对FloorplanGAN的住宅平面图生成效果进行了评估。FloorplanGAN的实现代码已上传至在线代码仓库：https://github.com/luozn15/FloorplanGAN。知其然也知其所以然。在构建FloorplanGAN来应对住宅平面图生成任务之后，我们也希望从中获取住宅设计相关的信息和解释。可解释性方法能够从特定角度解释分析深度模型的运行机制。除了本研究提出的FloorplanGAN之外，我们另外引入了两个同样关注住宅平面图生成的深度模型Graph2Plan和House-GAN作为解释对象，以减小主观释读的偏差。我们使用了卷积层可视化、线性分类器探针、显著性分析和注意力可视化这四种可解释性方法，针对上述三个深度模型分别开展分析。对比总结后，我们发现了深度模型对住宅平面图生成任务的有趣理解，包括：平面图识读的主次关系和层级顺序；平面图中的信息量与相应所需的模型规模；对图底关系的反常认识；对平面图属性重要性的不同理解等。住宅平面图模型可解释性分析的实现代码已上传至在线代码仓库：https://gitee.com/luozn19/interpret_floorplannets。综上，本研究使用建筑学知识和经验指导深度模型构建、训练和性能评价，使得深度学习方法能更好地满足建筑设计需求，融入建筑设计流程。同时探索了使用可解释性方法分析建筑相关深度模型的新路径，获得机器对建筑设计任务的“认知”和“理解”，或将有助于推动人机协同设计研究、提供设计认知研究新视角。

With the blossoming of deep learning in various fields today, how to make it better serve architectural design and practice is an urgent issue. This study attempts to establish a two-way connection between architectural design knowledge and architecture-related deep learning models: using architectural design knowledge and experience as a priori to guide the construction of the deep model framework, and using the interpretability analysis of the deep model to verify and obtain more empirical knowledge.Architectural floorplans are quite different from the traditional domain of computer vision and graphics concerns in terms of both information representation and generation process. Therefore, a direct application of a mature depth model is not the optimal solution for intelligent design in the architectural domain. Based on our understanding of the architectural design problem, we construct FloorplanGAN, a depth model adapted to the task of vector residential floorplan generation. This is a generative adversarial model framework that contains a vector generator, a differentiable renderer, and a raster discriminator. A self-attentive mechanism is embedded in the generator to improve the ability to model room interrelationships. After completing the model training, we evaluated the effectiveness of FloorplanGAN for residential floorplan generation with a multi-perspective evaluation system. The code for the FloorplanGAN implementation has been uploaded to the online code repository: https://github.com/luozn15/FloorplanGAN.Knowing what it is and also why it is. After constructing the FloorplanGAN to cope with the residential floorplan generation task, we also want to obtain the information and interpretation related to residential design from it. Interpretable methods can explain the operational mechanism of deep models from specific perspectives. In addition to the FloorplanGAN proposed in this study, we introduced two other deep models Graph2Plan and House-GAN, which also focus on residential floorplan generation task, as interpretable objects in order to reduce the bias of subjective interpretations. We use four interpretable methods, namely convolutional layer visualization, linear classifier probe, saliency analysis and attention visualization, to analyze each of these three deep models. The comparative summary reveals interesting understandings of the deep models for the residential floorplan generation task, including: the primary and secondary relationships and hierarchical order of floorplan understanding; the amount of information in the floorplan and the corresponding required model size; the perverse understanding of the figure-ground relationships; and the different understanding of the importance of the attributes of the floorplan. The implementation code for the interpretability analysis of the floorplan models has been uploaded to the online code repository: https://gitee.com/luozn19/interpret_floorplannets.In summary, this study uses architectural knowledge and experience to guide deep model construction, model training and performance evaluation, so that deep learning methods can better meet architectural design needs and integrate into the architectural design process. This study explores a new path to analyze architecture-related deep models using interpretability methods. It obtains machine "knowledge" and "understanding" of architectural design tasks, which may help promote human-machine collaborative design research and provide new perspectives on design cognition research.