Contrib

Modules

Layers

class Residual : public fl::Container

A module for creating a generic Residual block as given by He et al (2015) and Kumar et al (2015).

Example: 

auto res = Residual();
// Add multiple layers
res.add(Conv2D(30, 50, 9, 7, 2, 3, 3, 2));
res.add(BatchNorm(2, 50));
res.add(ReLU());
// Add a shortcut from the input to the block to the third layer
res.addShortcut(0, 3);
// Add a shortcut from the second layer to the output
res.addShortcut(2, 4);
// Scale the inputs to the third layer by some constant
res.addScale(3, 0.5);

// Create a model
Sequential model;
// ...
// Add our residual block as needed
model.add(res);
model.add(Pool2D(2, 3, 1, 1, 1, 1, PoolingMode::MAX));
model.add(res);
// ...

Public Functions

void addScale(int beforeLayer, float scale)

Adds a scaling factor to all residual connections connecting to a layer given by some index index.

Given some scale \( \alpha \), the input to beforeLayer becomes \( (x + f(x)) * \alpha \).

Parameters:

  • beforeLayer – [in] the index of the layer to which to scale the input and residual connection output.

  • scale – [in] the value by which to scale the sum of the previous layer and output of the residual connection.

void addShortcut(int fromLayer, int toLayer)

Adds a shortcut between two layers.

Parameters:

  • fromLayer – [in] the layer index from which the skip connection will originate; must be in the range \( [0, N_{layers} - 1] \). If the index 0 is used, the input to the shortcut will be equal to the input to the residual block.

  • toLayer – [in] the layer index to which the skip connection outputs a tensor; must be in the range \( [1, N_{layers} + 1] \). If the index \( N_{layers} + 1 \) is used, the output of the shortcut will be added to the output of the entire residual block.

template<typename T>
inline void addShortcut(int fromLayer, int toLayer, const T &module)

See Residual::addShortcut.

template<typename T>
inline void addShortcut(int fromLayer, int toLayer, std::shared_ptr<T> module)

Adds a shortcut connection between two layers such that tensors passed through the connection are forwarded through a passed module before being added to the resultant module’s input.

Can be used to reshape the output of input module to match the input dimensions for the output module.

Parameters:

  • fromLayer – [in] the layer index from which the shortcut connection will originate; must be in the range \( [0, N_{layers} - 1] \). If the index 0 is used, the input to the shortcut will be equal to the input to the residual block.

  • toLayer – [in] the layer index to which the shortcut connection outputs a tensor; must be in the range \( [1, N_{layers} + 1] \). If the index \( N_{layers} + 1 \) is used, the output of the shortcut will be added to the output of the entire residual block.

  • module – [in] a specified module through which the input to the shortcut connection will be forwarded before being added to the input to the destination module.

virtual std::vector<Variable> forward(const std::vector<Variable> &inputs) override

Performs forward computation for the module, given some inputs.

Parameters:

inputs – the values to compute forward computation for the module.

Returns:

a vector of Variable tensors containing the result of the forward computation

virtual std::string prettyString() const override

Generates a stringified representation of the module.

Returns:

a string containing the module label