- 博客(24)
- 收藏
- 关注
RSS订阅原创 PathTracing: The implementation of Bidirectional Path Tracing
The reading note for veach chapter8-10
2022-06-15 22:29:32
199
原创 Godot Main::start函数
If have assign a script by “-s”, the script is the game loop. Create script_loop as MainLoopMainLoop is the abstract base class for a Godot project’s game loop. It is inherited by SceneTree, which is the default game loop implementation used in Godot p.
2022-05-06 16:49:18
294
原创 cuda 光追
https://gitee.com/bracelethin/cuda_ray-tracing坑：别用带虚的继承，不然在一个子类调用另一个子类的函数也算递归。结构体的cuda函数：将自身的指针成员转为显存上的（后续要注意一下内存管理）复制一个本类的复制品的指向显存的指针result；返回result。...
2022-04-10 14:23:46
284
原创 疑惑。。。
怀疑某个loss（lossp）造成崩溃（cam全为0），尝试从总loss剔除，结果发现一旦计算lossp就会崩溃（即使没参加backward），不计算就不会，将下述第一个代码段改成第二个就ok了，具体关注cam_little，包括进入计算lossp的部分no——ok def forward(self, x, label,need_inter=False,need_p=True): # print("1=",meminfo.free/1024**3) #已用显存大小 sc
2021-12-29 21:42:56
750
原创 pytorch Tensor
//aten\src\ATen\templates\TensorBody.hTensor: TensorBase//aten\src\ATen\core\TensorBase.hTensorBase{c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> impl_;}//c10\core\TensorImpl.hstruct C10_API TensorImpl : public c10::intrusive_ptr_target{
2021-12-11 23:32:41
168
原创 torch autograd c++ source code 源码主干部分（大部分在csrc里面的Engine.cpp）
Edge{/// The function this `Edge` points to. std::shared_ptr<Node> function; /// The identifier of a particular input to the function. uint32_t input_nr;}Node{ // Sequence number used to correlate backward nodes with forward ops in the
2021-12-09 23:43:58
296
原创 muliti-similar loss
General Pair Weighting (GPW) frame workA unified weighting view of sampling problem of metric learning.multi-similarity (MS) lossα\alphaα and β\betaβ are hyper-parametersthere are two steps,(1)Pair miningA negative pair is compared to the hardest
2021-11-12 13:58:02
277
原创 Class-Specific Affinity based Weakly Supervised Semantic Segmentation with Neutral Region Exploratio
Class specific affinityAdd a new positive pairThe attitude to some pairs except traditional positive pairs
2021-10-06 11:36:04
58
原创 Learning Integral Objects with Intra-Class Discriminator for Weakly-Supervised Semantic Segmentation
When the number of class is n, the number of this method is 2n. As the photo show. Every class separated to foreground and background.Q1: The ‘C’ on this photo means the number of class? If do, why it must be it?Q2: As w is a vector, is f a c dimensio.
2021-10-04 22:30:59
187
原创 Deeplab-v3
Q: Why concat and conv on all feature map to one map?As the effect of “Depthwise separable convolution” and “Atrous Conv”, reduce the cost of calculation or storage.Q: rate?The parameter of “Atrous Conv”
2021-10-02 19:38:21
54
原创 Weakly-Supervised Semantic Segmentation via Sub-category Exploration
Apply clustering on the feature output by Feature Extraction which is also the first part of classification network.Supposing their are K centers of clustering, We expand the labels of classification to c*K kinds labels.
2021-10-02 19:16:21
59
原创 CONTA: Causal Intervention for Weakly-Supervised Semantic Segmentation
As the part of theory is hard to understand, I will start from method part and complete theory part in the futuremethodThe most difficult part is the meaning of C: confounder set. The cic_ici is the average of all XMX_MXM which is the processed CAMs.
2021-10-02 16:51:10
429
原创 Weakly Supervised Semantic Segmentation with Patch-Based Metric Learning Enhancement
Propose a new loss from Metric Learningp (i,j), is the pixels with corporation (i,j) on the k feature map, which form a vectorsampleSample P with number M from all P, P+P_+P+ is the P which is in the same class with PαP_\alphaPαP−P_-P− is the one n
2021-09-17 23:16:16
152
原创 Self-supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation
A new way to generate CAMs by equivalent after affinePCM ( Pixel Correlation Module):less some function:avoid over-fittingno residual:keep the same activation intensity of the original CAMReLU:?We use ReLU activation function with L1 normalizatio
2021-09-17 15:51:26
261
原创 Learning Pixel-level Semantic Affinity with Image-level Supervision
backgroundThe data for segmentation is hard to get as it need pixel-wise label.contributionFind a new way to generate labeled datamethod(1) Use classification network to generate the seed of CAMs:classification network: a fully connected layers appen
2021-09-16 15:34:09
132
原创 General Instance Distillation for Object Detection
BackgroundPrevious distillation methods of detection have weak generalization for different detection frameworks and rely heavily on ground truth(GT), ignoring the valuable relation information between instances.Distillation methods on multi-classificati
2021-08-02 13:55:34
284
原创 pandoc中文
pandoc --pdf-engine=xelatex -V mainfont="SimHei" 2020C-中小微企业的信贷决策.docx -o 2020C-中小微企业的信贷决策.pdf
2021-08-01 21:26:32
218
原创 Structure-Level Knowledge Distillation For Multilingual Sequence Labeling
Background:transfer the knowledge of monolingual models to multilingual model.Structure is hard to distillation because the amount of round.Target:find way to distillate structure-level information。For example, we want to distill the model which used
2021-07-27 23:59:26
103
原创 主成分分析 pca svd
无偏估计x‾\overline xx是采样的平均值μ\muμ 是实际期望值粗略来说：∑(xi−x‾)2<∑(xi−μ)2\sum (x_i-\overline x)^2<\sum (x_i-\mu)^2∑(xi−x)2<∑(xi−μ)2所以要用n-1增大左边推导过程需要E[(X‾−μ)2]=σ2nE[(\overline X-\mu)^2]=\frac{\sigma^2}{n}E[(X−μ)2]=nσ2我还理解不了就不写了。qwq协方差Cov(X,Y)=1n−1∑(xi
2021-04-12 23:39:35
41
原创 python b‘\x00\x00\x08\x03‘ to b‘00000803‘ 奇怪的\x
when process mnist datasetm=b'\x00\x00\x08\x03'can‘t decode directly which make int(m.decode(),16)wrong.or int(m,16)n=base64.b16encode(m) #b'00000803'int(n,16) #2051if you want to add ‘\x’base64.b16encode(n) #b'\x00\x00\x08\x03'
2021-03-29 17:59:42
1858
原创 cpp class property mass
staticstatic data member can be change by all way except const function.static function only use static memberconstruct and de orderclass A{B b;C c;}con:b,c,ade a,c,binitialize listclass A{const a;A(int 1):a=i //can change const{}}oth
2021-03-21 22:37:04
72
转载 redshift 踩坑
参考：https://blog.csdn.net/u014025444/article/details/91488957$ sudo pacman -S redshift安装$ redshift -l geoclue2查询经纬写完一定要关闭终端！！本小白在这被坑了，没意识到这也是在运行。。。多个运行屏幕会闪烁rootmkdir /home/{你的用户名}/.config/redshiftvim /home/{你的用户名}/.config/redshift/redshift.conf;
2020-12-19 14:30:08
615
1
空空如也
空空如也
TA创建的收藏夹 TA关注的收藏夹
TA关注的人