Yowey
Yowey
## Row LSTM/Diagonal BiLSTM ### 结构  ### 计算  > https://www.cnblogs.com/lart/p/10540898.html > Pixel Recurrent Neural Networks
这里的位置敏感得分图应该是什么样的形状呢? (H, W, C), 其中的C有两种想法: 哪一种会更好些呢?
overfeat的框融合的策略,感觉和r-cnn的基于iou的策略有些相似。 > https://www.cnblogs.com/zf-blog/p/6740736.html > 对2000×20维矩阵中每列按从大到小进行排序; > 从每列最大的得分建议框开始,分别与该列后面的得分建议框进行IoU计算,若IoU>阈值,则剔除得分较小的建议框,否则认为图像中存在多个同一类物体; >  > match_score(b1 ,b2)使用两个边界框的中心之间的距离和框的交叉区域之和来计算匹配分数,当它大于某个阈值时算法停止;
+ [ ] Use the `amp` of `pytorch` for AMP. + [ ] Fix some mistakes in the readme.md + [X] Add a simpler version.
The Equ.5:  In my opinion, this equ calculates the sum of features in the neighborhood corresponding to (i,j). But in the code: https://github.com/sail-sg/volo/blob/1f67923404d85cb8012a61b35d7eff782fe90cef/models/volo.py#L94-L95 `F.fold(x, output_size=(H, W), ...)` implements another...
关于缺少的内容
* PyTorch基础知识(这点主要可以看文档基础教程) * PyTorch高级技巧(Hook、梯度累积等等,这我了解不是太多) * PyTorch实战技巧(如何提速(我的这个列表收集了一些)、如何有效的训练、如何降低显存占用等等) * PyTorch辅助工具实践(高层封装库,类似fastai这样的,等等) * PyTorch可视化特征图、感受野、运行过程监控、运行结果显示等等 * 除了PyTorch和Tensorflow这样的工具介绍外,感觉可以逐步引入相关知识点的介绍 再补充一个很重要的: * 如果能够深入浅出的介绍CUDA的编程,那是极好的!
## iHarmony4 - Project: https://github.com/bcmi/Image-Harmonization-Dataset-iHarmony4 > Our dataset iHarmony4 is a synthesized dataset for Image Harmonization. It contains 4 sub-datasets: HCOCO,HAdobe5k, HFlickr, and Hday2night, each of which contains synthesized composite...
## DeepCrack: A Deep Hierarchical Feature Learning Architecture for Crack Segmentation We established a public benchmark dataset with cracks in multiple scales and scenes to evaluate the crack detection systems....