fuhao7i的博客 | fuhao7i Blog

Detector🎯5—— one-stage VS two-stage

"FPN, RPN"

1. main differences one-stage network is faster. two-stage network is more accurate. 2. one-stage Directly regress the category probability and position coordinate value of the object w...

Posted by fuhao7i on April 13, 2021

Neural Network🦖8——model.eval() VS with torch.no_grad()

"BatchNorm, Dropout, grad_backprop"

In short model.eval() will notify all your layers that you are in eval mode, that way, batchnorm or dropout layers will work in eval mode instead of training mode. torch.no_grad()...

Posted by fuhao7i on April 13, 2021

Dali杂货铺🐰19——Rules for defining the names of hyperparameters in torch

"model.state_dict(), model.named_parameters()"

Sometimes, we need to train models jointly, or use one model to help another model training. So it’s critically for us to konw the name of parameters in the model. 1 2 3 4 5 6 7 8 9 10 11 12 1...

Posted by fuhao7i on April 10, 2021

Neural Network🦖7——How to clip gradient?

"clip grad, Gradient clipping, Gradient Scaling"

Due to the improper selection of the learning rate, the weight update is larger. There is a lot of noise in the prepared data, resulting in large differences in target variables. Th...

Posted by fuhao7i on April 7, 2021

Dali杂货铺🐰18——The role of hook in mmdetection

"HOOK, hook"

Reference 目标检测(MMdetection)-HOOK机制

Posted by fuhao7i on April 7, 2021

Dali杂货铺🐰17——torch 训练过程查看某一层的梯度

"Drawing neural network"

在loss.backward()之后查看某一层的梯度 1 2 print('=== after loss.backward() ===') print(self.model.conv0.weight.grad)

Posted by fuhao7i on April 5, 2021

Dali杂货铺🐰16——Summary of drawing neural network tools

"Drawing neural network"

1. plotNeuralNet 2. Bokeh Bokeh官方文档 Bokeh是一个交互式可视化库，面向现代Web浏览器进行演示。它的目标是提供优雅，简洁的多功能图形构造，并通过非常大或流数据集的高性能交互来扩展此功能。 Bokeh可以帮助任何想要快速轻松地创建交互式图表，仪表板和数据应用程序的人。

Posted by fuhao7i on April 5, 2021

Neural Network🦖6——How dose the convolutional layer achieve higher or lower dimensionality?

"Convolutional principle"

The convolution kernel not only has height and width but also has depth And it has the same depth as the feature map being convolved. Therefore, each convolution kernel can traverse all the feat...

Posted by fuhao7i on April 5, 2021

Dali杂货铺🐰15——loss不能正常下降 or 下降到一定程度便不下降了

"loss problems"

1. 分析:如何解决神经网络训练时loss不下降的问题 1. 学习率没有灵魂学习率是个神奇的东西, 你是不是也对learning rate的选取而苦恼过, 总是感觉一成不变的学习率缺少灵魂: 太小？ loss降低的太慢💦 太大？ loss可能达不到最优, 而可能在最优值范围震动解决: torch.optim.lr_scheduler 学习率下降机制 How to...

Posted by fuhao7i on April 2, 2021

Dali杂货铺🐰14——GPU显存占满，但利用率却很低？

"Tensorflow, 渴望GPU"

1. float32 查看各个张量(即每一层的输入输出，一般输入类型对了，输出类型也对)是否为float32，Tensorflow在GPU上不能处理float64的tensor，因此我们需要确保每个tensor为float32位。注意 np.ones 和 np.zeros 都默认为float64位. astype: 转换np.array的数据类型 a.dtype: 输出a的数据类型...

Posted by fuhao7i on April 2, 2021

fuhao7i Blog

Detector🎯5—— one-stage VS two-stage

"FPN, RPN"

Neural Network🦖8——model.eval() VS with torch.no_grad()

"BatchNorm, Dropout, grad_backprop"

Dali杂货铺🐰19——Rules for defining the names of hyperparameters in torch

"model.state_dict(), model.named_parameters()"

Neural Network🦖7——How to clip gradient?

"clip grad, Gradient clipping, Gradient Scaling"

Dali杂货铺🐰18——The role of hook in mmdetection

"HOOK, hook"

Dali杂货铺🐰17——torch 训练过程查看某一层的梯度

"Drawing neural network"

Dali杂货铺🐰16——Summary of drawing neural network tools

"Drawing neural network"

Neural Network🦖6——How dose the convolutional layer achieve higher or lower dimensionality?

"Convolutional principle"

Dali杂货铺🐰15——loss不能正常下降 or 下降到一定程度便不下降了

"loss problems"

Dali杂货铺🐰14——GPU显存占满，但利用率却很低？

"Tensorflow, 渴望GPU"

FEATURED TAGS

ABOUT ME

FRIENDS