ipinmi / landmark-classifier-and-tagger

Convolutional Neural Network architectures for classifying landmark images.

Landmark Classification & Tagging for Social Media

Two Convolutional Neural Network architectures for classifying landmark images. One built from the scratch and the other utilizes transfer learning.

Project Steps

1. Creating a CNN architecture from Scratch to Classify Landmarks
   • Data Loading and Preprocessing
     • The data_transforms dictionary contains train, valid and test keys. The values are instances of transforms.Compose.
     • At the minimum, the 3 set of transforms contains a Resize(256) step, a crop step (RandomCrop for train and CenterCrop for valid and test), a ColorJitter step, a ToTensor step and finally a Normalize step (which uses the mean and std of the dataset).
     • The ImageFolder instances for train, valid and test use the appropriate transform from the data_transforms dictionary (using the “transform” keyword of ImageFolder)
     • The data loaders for train, valid and test use the right ImageFolder instance and use the batch_size, sampler, and num_workers that are given in input to the function.
   • Model Architecture
     • The model architecture is a CNN with 5 convolutional layers and 2 linear layers. The output of the model is a logit for each class. The model uses dropout and batch normalization to reduce overfitting.
   • Training and Testing the Model
     • The model is trained first for 70 epochs with a learning rate of 0.001 and achieves a test accuracy of 42%
     • The model is reloaded from its checkpoint and trained for second time for 30 epochs with a learning rate of 0.0005 and achieves a test accuracy of 51%
   • Saving and Loading the Model
     • The trained model is saved as a checkpoint along with associated hyperparameters and the class_to_idx dictionary.
     • There is a function that successfully loads a checkpoint and rebuilds the model.
   • Inference for Classification
     • There is a function that successfully makes a prediction for an image. The function returns the top 𝐾 most likely classes along with the probabilities. It takes a path to an image and a checkpoint, then returns the probabilities and classes.
   • Sanity Checking with matplotlib
     • There is a function that takes a path to an image and a model checkpoint, then plots the image and its predicted classes. The function uses matplotlib to plot the image and its 𝐾 most likely classes with actual flower names.

2. Using Transfer Learning to Classify Landmarks
   • Model Architecture
     • The model architecture uses the ResNet18 pre-trained CNN model.
     • All parameters of the loaded architecture are frozen, and a linear layer at the end has been added using the appropriate input features (as returned by the backbone), and the appropriate output features, as specified by the n_classes parameter
   • Training and Testing the Model
     • The model is trained first for 50 epochs with a learning rate of 0.001 and achieves a test accuracy of 74%.
   • Saving and Loading the Model
     • The trained model is saved as a checkpoint along with associated hyperparameters and the class_to_idx dictionary.
     • There is a function that successfully loads a checkpoint and rebuilds the model.
   • Inference for Classification
     • There is a function that successfully makes a prediction for an image. The function returns the top 𝐾 most likely classes along with the probabilities. It takes a path to an image and a checkpoint, then returns the probabilities and classes.
   • Sanity Checking with matplotlib
     • There is a function that takes a path to an image and a model checkpoint, then plots the image and its predicted classes. The function uses matplotlib to plot the image and its 𝐾 most likely classes with actual flower names.