modules

Repository of NN modules, metrics, losses and transforms by AdeptusN

conv_modules.py

Contains custom convolutional neural net modules:

• Conv3x3 - convolution module with optional batch_norm, dropout layers and activation function;
• Conv5x5 - convolution module that replaces conv layer with kernel_size=5 by 2 conv layers with kernel_size=3. Have optional Dropout and BatchNorm layers after every conv layer and optional residual connection;
• Conv7x7 - convolution module that replaces conv layer with kernel_size=7 by 3 conv layers with kernel_size=3. Have optional Dropout and BatchNorm layers after every conv layer and optional residual connection;

losses.py

Contains custom losses:

• FocalLossBin - focal loss for binary classification;
• FocalLossMulti - [IN TEST] focal loss for multiclass classification [IN TEST];
• L1Loss - based on manhattan distance loss;
• VAELoss - a loss for variational autoencoder;
• PerceptualLoss - perceptual loss calculate MSE loss between vgg16 activation maps of model output image and target image;
• WassersteinLoss - a loss for discriminator of generative adversarial network;
• GradientPenalty - gradient penalty (for Wasserstein loss);

metrics.py

Contains custom metrics:

• IoUMetricBin - intersection over union metric for binary segmentation;
• DiceScoreBin - metric for binary segmentation based on dice-coefficient;

transforms.py

Contains custom transforms:

• MinMaxScale - Min-Max scale to transform image pixels to [0, 1];
• ThresholdTransform - threshold image transformation;
• DividerScaler - scalar division of image tensor (can be added to torchvision.transforms.Compose);