Image-Recognition

Project Structure

Main Script:

• main.py

Module "data_preprocessor":

• DataHandler - responsible for reading images and labels from a certain directory, splitting datasets into training and testing, and creating DataLoader objects:
  • __init__(root: str, size: tuple[int]) - root is a path to a dataset; size is images size after resizing
  • get_data(batch_size: int, test_size: float) - batch_size - size of a single batch; test_size - part of a test data (for instance 0.2 or 20%, so train data's part is 0.8 or 80%); returns train and test DataLoaders respectively as a tuple.
  • class_names@property - a list of class names (folders' names within the root directory)
  • class_indexes@property - a list of class indexes
• old DataHandler (unfinished/inactive) - has the same functionality as an actual DataHandler, but doesn't use ImageFolder, DataLoader, etc; had been left as an example.
• ImagesReader (inactive) - similar to ImageFolder from torchvision, the purpose is the same as well; the class is being used only in the old DataHandel.

Module "models":

• MainModel - allows interacting with the CNN model; doesn't provide a direct access to the model's instance, but has methods for its training and evaluating:
  • __init__(classes_count: int, learning_rate: float, size: tuple[int], load_path: str) - classes_count is a number of classes; learning_rate is an optimizer's learning rate; size is images size after resizing; load_path is a path (including ".pth" file) from which state_dict will be loaded, trains a new model if None.
  • test_loop(dataset: DataLoader) - dataset is testing data; returns an accuracy and a total loss respectively as a tuple.
  • train_loop(epochs: int, dataset: DataLoader, test_dataset: DataLoader, save_path: str) - epochs is epochs count; dataset is training data; test_dataset is testing data; save_path is a path where a model will be saved, or won't if None; performs model's training, fills training_history, testing_history and accuracy_history.
  • history@property - a tuple consisting of training_history, testing_history and accuracy_history; should be taken after performing model's training.
• ConvolutionalLayers - convolutional layers of the CNN (inspired by AlexNet):
  • Conv2d(in=3, out=64, kernel=5x5, padding=2)
  • ReLU()
  • MaxPool2d(kernel=2x2)
  • Conv2d(in=64, out=32, kernel=3x3, padding=1)
  • ReLU()
  • MaxPool2d(kernel=2x2)
  • Conv2d(in=32, out=16, kernel=3x3, padding=1)
  • ReLU()
  • MaxPool2d(kernel=2x2)
• Classifier - classification layer in the CNN with following layers (fully-connected):
  • Flatten()
  • Linear(input_size, classes_count) - input_size can be calculated by analyzing the convolutional part of the network

Tests

Tests were performed including following parameters:

• learning_rate = 0.05,
• batches = 20,
• epochs = 100,
• images_size = 80x80
• testing_dataset_size = 8%
• training_dataset_size = 92%

Additional info

The CNN was learning to recognize weather from different photos. There were 6 weather classes: dew, glaze, hail, rain, sandstorm, snow. You can check them all out in the "./data" folder.

Results