Traffic Sign Classification

This project uses the German Traffic Sign Recognition Benchmarks Dataset(GTSRB). This repository is aimed at exploring the utilization of various torchvision.models for transfer learning.

Data Loader and Directory Structure

I included a data loader for the GTSRB data set with the help of Pytorch's data loading tutorial. I wrote two different data loaders. get_train_valid_loader for the training and validation sets (courtesy kevinzakka) and TrafficSignDataset for the test data set. get_train_valid_loader uses torchvision.datasets.ImageFolder that leverages the following directory structure) and makes our life extremely easy to load data. The test data loader requires a .csv file with annotations, and is explained in the ipython notebook here

The directory structure is assumed to look as

• GTSRB/

  • Final_Training/

    • Images/

      • 00000/

      • 00001/

        ...

  • Final_Test/

    • Images/

      • 00000.pm

        ...

      • GT-final_test.csv

The folder names 00000/, 00001/ correspond to the class labels, and there are 43 such labels.

Fine tuning torchvision.models

• Resnet takes 224x224 images as input, so we need to apply an appropriate transform (torchvision.transforms.Resize, torchvision.transforms.RandomResizedCrop) while calling the data loader. To tune the model for a custom number of classes, we modify the final fully connected layer with the number of classes (here 43).

  model = models.resnet34(pretrained=True)
  num_ftrs = model.fc.in_features
  model.fc = nn.Linear(num_ftrs, 43)
  
  

• Inceptionv3 takes in 299x299 images, so we need to apply an appropriate transform while calling the data loader. Inceptionv3 enables Auxiliary Classifiers while training. To modify the model for a custom number of classes (in the output layer)

  def get_pretrained_inception(num_classes, pretrained=True):
    inception = torchvision.models.inception_v3(pretrained=pretrained)
  
    fc_in_features = inception.fc.in_features
    inception.fc = nn.Linear(in_features=fc_in_features, out_features=num_classes)
    inception.AuxLogits = InceptionAux(in_channels=768, num_classes=num_classes)
  
    return inception
  
  

  inception(inputs) returns a tuple corresponding to the outputs of the main classifier and the auxiliary classifier. If required, the losses corresponding to both the classifiers can be weighted and passed onto the optimizer. Although, only the predictions of the main classifier are used during test time.

• Squeezenet takes 255x255 images as input. It is mentioned in some threads in Pytorch forums that it can adaptively take different input shapes, but I haven't tried it so far. Training squeezenet is extremely fast. As compared to inception, modifying squeezenet is relatively straightforward. We need to modify the conv2d layer in the Classifier

  model = models.squeezenet1_1(pretrained=True)
  num_ftrs = model.classifier._modules['1'].in_channels
  
  nn.Linear(num_ftrs, 42)
  model.classifier._modules['1'] = nn.Conv2d(num_ftrs, 43, 3)
  model.num_classes = 43
  

  We just modify the Conv2d layer with the required number of classes.

• VGG We only need to modify the Classifier with our required number of classes.

  model = models.vgg11(pretrained=True)
  num_ftrs = model.classifier._modules['6'].in_features
  model.classifier._modules['6'] = nn.Linear(num_ftrs, 43)
  
  

Image Augmentation

• Todo

Hyperparameter optimization

• Todo