Model Architectures

FastAI Model Architectures

FastAI Model Version 1 Architecture

1. Data Preparation:

   • DataLoader Creation: Utilizes FastAI methods to create a DataLoader object, encompassing both training and validation data loaders.
   • Batch-Level Transformations:
     • Resize: Each image is resized to 224x224 pixels.
     • Random Scaling: Applies a minimum scale of 75% of the original image size to each image.
     • Data Augmentation (aug_transforms): Incorporates various image augmentations, including:
       • Rotation: Randomly rotates images to simulate different angles.
       • Cropping: Randomly crops images to enhance model robustness.
       • Color Jittering: Adjusts brightness, contrast, saturation, and hue to improve generalization.

2. Model Architecture:

   • Base Network: Uses the ResNet34 model, a pre-defined Convolutional Neural Network (CNN) with 34 layers, known for its deep learning capabilities and residual connections to facilitate training.

FastAI Model Version 2 Architecture

1. Enhanced Data Augmentation:

   • Augmentation Details:
     • Rotation: Allows random rotations up to 30 degrees.
     • Zooming: Applies scaling up to 10% larger than the original size.
     • Lighting Adjustments: Introduces random changes to lighting, including brightness and contrast adjustments.
     • Warping: Applies random distortions to images to simulate various transformations.

2. Model Training:

   • Base Network: Utilizes the same ResNet34 architecture.
   • Learning Rate Finder (lr_find): Employs the learning rate finder to identify the optimal learning rate. This pre-training phase helps in determining the best rate for efficient model training.

PyTorch Model Architecture

PyTorch Model 1 Architecture

1. Network Architecture:

   • Convolutional Layers:
     • Layer 1: 3x3 convolution with 32 filters, stride of 1, and padding of 1.
     • Layer 2: 3x3 convolution with 64 filters, stride of 1, and padding of 1.
     • Layer 3: 3x3 convolution with 128 filters, stride of 1, and padding of 1.
   • Max Pooling Layer: Applies a 2x2 max pooling operation with a stride of 2, reducing the spatial dimensions of the feature maps by half.
   • Fully Connected Layers:
     • Hidden Layer: Linear layer mapping from the flattened output of the final convolutional layer to 512 features.
     • Output Layer: Final linear layer producing class predictions based on the number of classes in the dataset (len(train_dataset.classes)).

2. Forward Propagation:

   • Convolutional Processing: Each convolutional layer is followed by ReLU activation and max pooling to extract features.
   • Flattening: The output from the final convolutional layer is flattened into a 1D tensor.
   • Fully Connected Layers:
     • Hidden Layer Activation: Applies ReLU activation to the output of the first fully connected layer.
     • Output Prediction: The final fully connected layer generates predictions for each class based on the number of classes in the dataset.

PyTorch Model 2 Architecture:

1. Network Architecture:

   • Convolutional Layers:
     • Layer 1: 3x3 convolution with 64 filters, stride of 1, and padding of 1, followed by Batch Normalization.
     • Layer 2: 3x3 convolution with 128 filters, stride of 1, and padding of 1, followed by Batch Normalization.
     • Layer 3: 3x3 convolution with 256 filters, stride of 1, and padding of 1, followed by Batch Normalization.
   • Max Pooling Layer: Applies a 2x2 max pooling operation with a stride of 2, reducing the spatial dimensions of the feature maps by half.
   • Fully Connected Layers:
     • Hidden Layer: Linear layer mapping from the flattened output of the final convolutional layer to 1024 features.
     • Dropout Layer: Applies dropout with a probability of 0.5 to prevent overfitting.
     • Output Layer: Final linear layer producing class predictions based on the number of classes in the dataset (len(train_dataset.classes)).

2. Forward Propagation:

   • Convolutional Processing: Each convolutional layer is followed by Batch Normalization, ReLU activation, and max pooling.
   • Flattening: The output from the final convolutional layer is flattened into a 1D tensor.
   • Fully Connected Layers:
     • Hidden Layer Activation: Applies ReLU activation to the output of the first fully connected layer.
     • Dropout: Regularizes the model by applying dropout before the final output layer.
     • Output Prediction: The final fully connected layer generates predictions for each class based on the number of classes in the dataset.