PyTorch and Transfer Learning

PyTorch and Transfer Learning

• 1.
  Introduction to PyTorch ESE201503120 박준영
• 2.
  Tensor in PyTorch ◦Setting X, Y for input/output ◦ Setting Weights to train. ◦ Dot product. X*W1 ◦ Same with max(0,h) ◦ Calculate with mathematical operators
• 3.
  Autograd ◦ Automate backpropagation. After updating weights, Reset gradient for each training step to 0. <Traditional Back Propagation> <Automated Back Propagation>
• 4.
  Defining a model ◦Define model with functions for layers ◦ Use nn.Sequential() method. ◦ nn.Linear() : FC Layer ◦ nn.Conv2d() : Convolution Layer ◦ nn.MaxPool2d() : Max pooling Layer ◦ Make loss function (Mean Squared Error)
• 5.
  Model Example –AlexNet ◦ https://github.com/pytorch/vision/blob/master/torchvision/models/alexnet.py ◦ Implemented by using nn.Sequential Conv1 Conv2 Conv3 Conv4 Conv5 FC6 FC7
• 7.
  Optimizer < Defining Optimizer(Adam) > ◦ Reset gradients to ‘0’ ◦ Gradient will be updated when .backward() called. (Accumulated in buffers) ◦ Compute gradient of the loss ◦ Optimizer will update its parameters for each step
• 8.
  Transfer Learning ESE 201503120박준영
• 9.
  Transfer Learning ◦ Reference ◦http://cs231n.github.io/transfer-learning/ ◦ http://incredible.ai/artificial-intelligence/2017/05/13/Transfer-Learning/ ◦ http://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html ◦ Learning from scratch takes too much time. ◦ Using pre-built CNN models ◦ Import the weights on another network. ◦ Train again for new network. For 12 years For 4 years Transfer
• 10.
  Difference ◦ Fine tuning ◦Update weights of whole layer for new dataset ◦ ConvNet as fixed feature extractor ◦ Only train final Fully-Connected Layer.
• 11.
  Structure (Inception V3– Finetuning) Update weight of whole model for new data.
• 12.
  Structure (Inception V3– Fixed Extractor) Don’t train ! (Use pre-trained parameter) Only Train here (FC Layer)
• 13.
  Process ◦ Load datafor classification. ◦ Load a pre-trained model. ◦ Re-build a FC Layer (For number of classes). ◦ Train only a FC layer.
• 14.
  Load image data ◦Preprocess images ◦ Crop ◦ Flip ◦ Convert to tensor ◦ Normalize ◦ ImageFolder ◦ Make object for train/val datasets ◦ DataLoader ◦ Load data from ImageFolder object for specific format.
• 15.
  Rebuild FC Layer ◦Freeze all weights ◦ Build new FC Layer ◦ Classes : 2 (Ants, Bees) ◦ Features : Get # from model ◦ Set optimizer (using SGD) ◦ Set LR scheduler ◦ Adjust learning rate per step Learning rate will be decreased by 1/10 per epoch Only optimize for FC parameters
• 16.
  Result < Only Finetuning> < Fixed Feature Extractor >
• 17.
  Assignment ◦ Study moreon image preprocessing. ◦ 파이썬을 이용한 머신러닝, 딥러닝 실전 개발 입문 – 쿠지라 히코우즈쿠에 著 ◦ 웹 크롤링, 스크레이핑 및 이미지 로드 관련 내용 수록 ◦ 인천대학교 학산도서관 소장 ◦ Set inceptionV3 model for transfer learning with my own dataset. ◦ Other insects, faces, cars, places, foods … and so on ◦ Use another models for transfer learning. ◦ ResNet, AlexNet, VGD … and so on
• 18.
  Self Check ◦ Autograd의 역할과 사용 과정을 설명하라. ◦ Transfer Learning 을 사용하는 이유는 ? ◦ Finetuning 과 ConvNet as fixed feature extractor 를 비교하라.