Skip to content

Latest commit

 

History

History
executable file
·
342 lines (259 loc) · 10.1 KB

transfer.md

File metadata and controls

executable file
·
342 lines (259 loc) · 10.1 KB

Transfer Function Layers

Transfer functions are normally used to introduce a non-linearity after a parameterized layer like Linear and SpatialConvolution. Non-linearities allows for dividing the problem space into more complex regions than what a simple logistic regressor would permit.

HardTanh

Applies the HardTanh function element-wise to the input Tensor, thus outputting a Tensor of the same dimension.

HardTanh is defined as:

  • f(x) = 1, if x > 1,
  • f(x) = -1, if x < -1,
  • f(x) = x, otherwise.

The range of the linear region [-1 1] can be adjusted by specifying arguments in declaration, for example nn.HardTanh(min_value, max_value). Otherwise, [min_value max_value] is set to [-1 1] by default.

ii=torch.linspace(-2,2)
m=nn.HardTanh()
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

HardShrink

module = nn.HardShrink(lambda)

Applies the hard shrinkage function element-wise to the input Tensor. The output is the same size as the input.

HardShrinkage operator is defined as:

  • f(x) = x, if x > lambda
  • f(x) = x, if x < -lambda
  • f(x) = 0, otherwise
ii=torch.linspace(-2,2)
m=nn.HardShrink(0.85)
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

SoftShrink

module = nn.SoftShrink(lambda)

Applies the soft shrinkage function element-wise to the input Tensor. The output is the same size as the input.

SoftShrinkage operator is defined as:

  • f(x) = x-lambda, if x > lambda
  • f(x) = x+lambda, if x < -lambda
  • f(x) = 0, otherwise
ii=torch.linspace(-2,2)
m=nn.SoftShrink(0.85)
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

SoftMax

Applies the Softmax function to an n-dimensional input Tensor, rescaling them so that the elements of the n-dimensional output Tensor lie in the range (0,1) and sum to 1.

Softmax is defined as f_i(x) = exp(x_i-shift) / sum_j exp(x_j-shift), where shift = max_i x_i.

ii=torch.exp(torch.abs(torch.randn(10)))
m=nn.SoftMax()
oo=m:forward(ii)
gnuplot.plot({'Input',ii,'+-'},{'Output',oo,'+-'})
gnuplot.grid(true)

Note that this module doesn't work directly with ClassNLLCriterion, which expects the nn.Log to be computed between the SoftMax and itself. Use LogSoftMax instead (it's faster).

SoftMin

Applies the Softmin function to an n-dimensional input Tensor, rescaling them so that the elements of the n-dimensional output Tensor lie in the range (0,1) and sum to 1.

Softmin is defined as f_i(x) = exp(-x_i-shift) / sum_j exp(-x_j-shift), where shift = max_i -x_i.

ii=torch.exp(torch.abs(torch.randn(10)))
m=nn.SoftMin()
oo=m:forward(ii)
gnuplot.plot({'Input',ii,'+-'},{'Output',oo,'+-'})
gnuplot.grid(true)

SoftPlus

Applies the SoftPlus function to an n-dimensioanl input Tensor. SoftPlus is a smooth approximation to the ReLU function and can be used to constrain the output of a machine to always be positive. For numerical stability the implementation reverts to the linear function for inputs above a certain value (20 by default).

SoftPlus is defined as f_i(x) = 1/beta * log(1 + exp(beta * x_i)).

ii=torch.linspace(-3,3)
m=nn.SoftPlus()
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

SoftSign

Applies the SoftSign function to an n-dimensioanl input Tensor.

SoftSign is defined as f_i(x) = x_i / (1+|x_i|)

ii=torch.linspace(-5,5)
m=nn.SoftSign()
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

LogSigmoid

Applies the LogSigmoid function to an n-dimensional input Tensor.

LogSigmoid is defined as f_i(x) = log(1/(1+ exp(-x_i))).

ii=torch.randn(10)
m=nn.LogSigmoid()
oo=m:forward(ii)
go=torch.ones(10)
gi=m:backward(ii,go)
gnuplot.plot({'Input',ii,'+-'},{'Output',oo,'+-'},{'gradInput',gi,'+-'})
gnuplot.grid(true)

LogSoftMax

Applies the LogSoftmax function to an n-dimensional input Tensor.

LogSoftmax is defined as f_i(x) = log(1/a exp(x_i)), where a = sum_j exp(x_j).

ii=torch.randn(10)
m=nn.LogSoftMax()
oo=m:forward(ii)
go=torch.ones(10)
gi=m:backward(ii,go)
gnuplot.plot({'Input',ii,'+-'},{'Output',oo,'+-'},{'gradInput',gi,'+-'})
gnuplot.grid(true)

Sigmoid

Applies the Sigmoid function element-wise to the input Tensor, thus outputting a Tensor of the same dimension.

Sigmoid is defined as f(x) = 1/(1+exp(-x)).

ii=torch.linspace(-5,5)
m=nn.Sigmoid()
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

Tanh

Applies the Tanh function element-wise to the input Tensor, thus outputting a Tensor of the same dimension.

Tanh is defined as f(x) = (exp(x)-exp(-x))/(exp(x)+exp(-x)).

ii=torch.linspace(-3,3)
m=nn.Tanh()
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

ReLU

Applies the rectified linear unit (ReLU) function element-wise to the input Tensor, thus outputting a Tensor of the same dimension.

ReLU is defined as f(x) = max(0,x)

Can optionally do its operation in-place without using extra state memory:

m=nn.ReLU(true) -- true = in-place, false = keeping separate state.
ii=torch.linspace(-3,3)
m=nn.ReLU()
oo=m:forward(ii)
go=torch.ones(100)
gi=m:backward(ii,go)
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

PReLU

Applies parametric ReLU, which parameter varies the slope of the negative part:

PReLU is defined as f(x) = max(0,x) + a * min(0,x)

When called without a number on input as nn.PReLU() uses shared version, meaning has only one parameter. Otherwise if called nn.PReLU(nOutputPlane) has nOutputPlane parameters, one for each input map. The output dimension is always equal to input dimension. Note that weight decay should not be used on it. For reference see Delving Deep into Rectifiers.

RReLU

Applies the randomized leaky rectified linear unit (RReLU) element-wise to the input tensor, thus outputting a tensor of the same dimension. Informally the RReLU is also known as 'insanity' layer.

RReLU is defined as f(x) = max(0,x) + a * min(0,x), where a ~ U(l,u).

In training mode negative inputs are multiplied by a factor a drawn from a uniform random distribution U(l, u). In evaluation mode a RReLU behaves like a LeakyReLU with a constant mean factor a = (l+u)/2.

Syntax:

m=nn.ReLU(
   l,       -- minimum factor for negative inputs, default: 1/8;
   u,       -- maximum factor for negative inputs, default: 1/3;
   inplace  -- if true the result will be written to the input tensor, default: false;
)

If l == u a RReLU effectively becomes a LeakyReLU. Regardless of operating in in-place mode a RReLU will internally allocate an input-sized noise tensor to store random factors for negative inputs. The backward() operation assumes that forward() has been called before.

For reference see Empirical Evaluation of Rectified Activations in Convolutional Network.

ii=torch.linspace(-3, 3)
m=nn.RReLU()
oo=m:forward(ii):clone()
gi=m:backward(ii,torch.ones(100))
gnuplot.plot({'f(x)',ii,oo,'+-'},{'df/dx',ii,gi,'+-'})
gnuplot.grid(true)

SpatialSoftMax

Applies SoftMax over features to each spatial location (height x width of planes). The module accepts 1D (vector), 2D (batch of vectors), 3D (vectors in space) or 4D (batch of vectors in space) tensor as input. Functionally it is equivalent to SoftMax when 1D or 2D input is used. The output dimension is always the same as input dimension.

ii=torch.randn(4,8,16,16)  -- batchSize x features x height x width
m=nn.SpatialSoftMax()
oo = m:forward(ii)

AddConstant

Adds a (non-learnable) scalar constant. This module is sometimes useful for debugging purposes: f(x) = x + k, where k is a scalar.

Can optionally do its operation in-place without using extra state memory:

m=nn.AddConstant(k,true) -- true = in-place, false = keeping separate state.

In-place mode restores the original input value after the backward pass, allowing its use after other in-place modules, like MulConstant.

MulConstant

Multiplies input tensor by a (non-learnable) scalar constant. This module is sometimes useful for debugging purposes: f(x) = k * x, where k is a scalar.

Can optionally do its operation in-place without using extra state memory:

m=nn.MulConstant(k,true) -- true = in-place, false = keeping separate state.

In-place mode restores the original input value after the backward pass, allowing its use after other in-place modules, like AddConstant.