《科研实验室-如何做实验》由会员分享，可在线阅读，更多相关《科研实验室-如何做实验（28页珍藏版）》请在金锄头文库上搜索。
1、1 title Some Tips for Experiments 2 titleContents Why When to Conduct Experiments 1 Experiment Design2 Coding Tips Tricks3 Result Management 4 What If It Does Not Work 5 3 titleContents Why When to Conduct Experiments 1 Experiment Design2 Coding Tips Tricks3 Result Management 4 What If It Does Not W
2、ork 5 4 titleWhy When to Conduct Experiments experiment design coding training testing idea Why experiment Validate the idea 5 titleWhy When to Conduct Experiments experiment design coding training testing idea draft When experiment Think about it deeply 6 titleContents Why When to Conduct Experimen
3、ts 1 Experiment Design2 Coding Tips Tricks3 Result Management 4 What If It Does Not Work 5 7 titleExperiment Design Something to Consider 1 What kind of task 2 Any existing baselines 3 What kind of model 4 How many layers and channels 5 Which data set 6 Value of hyperparameters 7 Ablation study 8 ti
4、tleExperiment Design 1 What kind of task do we consider Image classification Dog Bird Bear Cat Object detection Semantic segmentationImage generation 9 titleExperiment Design 2 Is there any existing baseline Download the code Run and read it 10 titleExperiment Design 3 What kind of model should we u
5、se Input image Feature maps AlexNet VGG ResNet MobileNet etc Is there any constraint e g Params FLOPs Feature maps 11 titleExperiment Design 4 How many layers and channels should be considered 1 32 layers 128 filters 2 56 layers 64 filters 3 101 layers 48 filters different settings Try more settings
6、 12 titleExperiment Design 5 Which data set should be considered Data sets CIFAR SVHN ImageNet CelebA etc Data Set Training Samples Testing Samples CIFAR50 00010 000 SVHN604 36826 032 ImageNet1 28 million50 000 Varying sizes Varying Resolutions 28x28 32x32 256x256 1024x1024 etc 13 titleExperiment De
7、sign 6 How to choose the value of hyperparameters 1 0 1 0 01 0 001 etc 14 titleExperiment Design 7 How to design ablation study xxx model with vs without xxx module 15 titleContents 15 Why When to Conduct Experiments 1 Experiment Design2 Coding Tips Tricks3 Result Management 4 What If It Does Not Wo
8、rk 5 16 titleCoding Tips Tricks Clear Coding Structure Experiment Option Checkpoint Data loader Model Optimizer Logger 17 titleCoding Tips Tricks Write Useful Comments 18 titleCoding Tips Tricks Use Meaningful Names Use meaningful names for different functional coding files e g train gan py training
9、 with adversarial loss train L1 py training with L1 loss train mse py training with mean squared loss 19 titleCoding Tips Tricks Use Version Control Software Branch managementCommit history 20 titleCoding Tips Tricks Summary More Coding Tips 1 Clear coding structure 2 Write useful comments 3 Use mea
10、ningful names 4 Use version control software 5 Write code documents or readme file 6 Write test code 7 Use pdb log to debug 21 titleContents 21 Why When to Conduct Experiments 1 Experiment Design2 Coding Tips Tricks3 Result Management 4 What If It Does Not Work 5 22 titleResult Management Data SetTr
11、aining Time CIFAR5 10 h SVHN15 20 h ImageNet3d Save Checkpoints High training cost necessary to save checkpoints Name the file according to the settings save log training error testing error loss save model models at each epoch the best model 23 titleResult Management Visualization Tools Visdom Tens
12、orBoard 24 titleResult Management Result Management Save and manage your results different settings 25 titleResult Management Result Management Get something from your results 26 titleContents 26 Why When to Conduct Experiments 1 Experiment Design2 Coding Tips Tricks3 Result Management 4 What If It Does Not Work 5 27 titleWhat If It Does Not Work Check your code and training settings 90 cases Data processing Flip crop normalize etc Optimizer SGD vs Adam momentum weight decay etc Loss eps differentiable some coefficient etc Use the debug tool pdb 28 title Thanks
