Module dopt.core.ops.nnet

Contains common neural network operations.

These operations are currently only implemented for the CUDA backend.

Functions

NameDescription
convolution(features, filters, padding, stride, mod, line)Creates a convolution operation that performs the computation required to implement a convolutional layer.
convolutionFeaturesGrad(parentGrad, filters, featuresShape, padding, stride, mod, line)Creates an operation representing the derivative of a convolution operation with respect to the feature maps.
convolutionFiltersGrad(parentGrad, features, filtersShape, padding, stride, mod, line)Creates an operation representing the derivative of a convolution operation with respect to the filters.
convolutionTranspose(features, filters, padding, stride, mod, line)Creates a transposed convolution operation (also known, incorrectly, as deconvolution).
maxpool(features, dims, mod, line)Creates a max pool operation that performs the computation required to implement a max pooling layer.
maxpoolGrad(parentGrad, op, mod, line)Creates an operation representing the derivative of a maxpool operation with respect to the feature maps.
relu(inputs, mod, line)Creates an operation representing the computation required for a ReLU layer.
softmax(inputs, mod, line)Creates an operation representing the computation required for a softmax layer.
softmaxGrad(parentGrad, op, mod, line)Creates an operation representing the gradient of the softmax function.