Pytorch & front-end work combing
Data Set & Classic Model
Image classification data (mnist, cifar10, stl10, svhn)-VGG16, ResNet, AlexNet, LeNet, GoogleNet, DenseNet, Inception
Image segmentation data (PortraitDataset)-Unet
Target detection data (PennFudanPed)-Faster RCNN
Image generation data (img_align_celeba_2k)-GCGAN
Name classification data (names)-RNN
The front end needs to be able to select a column of the data set
Network layer
Pool-MaxPool2d, AvgPool2d, MaxUnpool2d (see Pytorch_Part3_model module )
Convolution layer-Convxd, ConvTranspose2d
Activation function-Relu, Sigmoid, tanh, RRelu, Leaky Relu
softmax
Dropout (see Pytorch_Part6_ regularization )
Standardization-BatchNormxd, LayerNorm, InstanceNorm2d, GroupNorm
The front end adds the above network layer structure
Loss function & optimizer & learning rate drop
Loss function-CrossEntropyLoss, NLLLoss, BCELoss, BCEWithLogitsLoss (see Pytorch_Part4_loss function )
Optimizer-SGD, RMSprop, Adam (see PyTorch study notes (7): PyTorch's ten optimizers )
Decreased learning rate-StepLR, MultiStepLR, ExponentialLR, ReduceLRonPlateau (see Pytorch_Part5_iterative training )
Add a self-selected column to the front end
Image enhancement
Crop-CenterCrop, RandomCrop, RandomResizeCrop (see Pytorch_Part2_data module )
Flip and rotate-RandomHorizontalFlip, RandomVerticalFlip, RandomRotation
Image transformation-pad, ColorJitter, GrayScale, RandomGrayScale, RandomAffine, RandomErasing
Randomly select the above methods-RandomChoice, RandomApply, RandomOrder
Out of order, Batch-size-parameters in DataLoader
Add an image enhancement method column to the front end
Network layer encapsulation
Unified use of Sequential structure for packaging (see Pytorch_Part3_model module )
Wrap the same network layer-ModuleList
Network layer multiplexer-ModuleDict
Unified front and back interfaces, implementation of the encapsulation layer
other
Epoch
CPU / GPU-model.to ('cuda'), tensor.to ('cuda') (see Pytorch_Part7_model usage )
Model saving and loading (in the \ (\ beta \) stage)