End-to-end Graph-constrained Vectorized Floorplan Generation with Panoptic Refinement
End-to-end Graph-constrained Vectorized Floorplan Generation with Panoptic Refinement
复制标题
DOI：
--
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Jiacheng Liu;Yuan Xue;José Duarte;Krishnendra Shekhawat;Zihan Zhou;Xiaolei Huang
中科院分区：
文献类型：
--
作者：
Jiacheng Liu;Yuan Xue;José Duarte;Krishnendra Shekhawat;Zihan Zhou;Xiaolei Huang
我们使用pytorch [4]实现了我们提出的方法，并在NVIDIA RTX 3090 GPU上进行了所有实验。 3×10-4。从top-bos⟩代币中，我们使用了带有Top-K的贪婪解码，然后从概率分布中取样以获得离散的坐标。为128。根据实验结果，我们选择5个改进迭代作为精炼阶段的默认设置。
We implemented our proposed method using PyTorch [4] and ran all experiments on an NVIDIA RTX 3090 GPU. We set up the codebase based on the official implementation of LayoutTransformer [1]. The AdamW [2] optimizer was applied with a constant learning rate 3× 10−4. We trained our model end-to-end for 20 epochs with batch size 128. We applied student forcing to generate the draft floorplan sequence from scratch during testing, starting from the ⟨BoS⟩ token. Specifically, we used greedy decoding with top-k probabilities, then sampled from the probability distribution to obtain discrete coordinates. We set k = 5 in all experiments. All feature dimensions used in the proposed network were set to be 128. Based on experimental results, we chose 5 refinement iterations as the default setting in the refining stage.