Permalink
Fetching contributors…
Cannot retrieve contributors at this time
1083 lines (848 sloc) 44.1 KB

dpnn : deep extensions to nn

This package provides many useful features that aren't part of the main nn package. These include sharedClone, which allows you to clone a module and share parameters or gradParameters with the original module, without incuring any memory overhead. We also redefined type such that the type-cast preserves Tensor sharing within a structure of modules.

The package provides the following Modules:

  • Decorator : abstract class to change the behaviour of an encapsulated module ;
  • DontCast : prevent encapsulated module from being casted by Module:type() ;
  • Serial : decorate a module makes its serialized output more compact ;
  • NaN : decorate a module to detect the source of NaN errors ;
  • Inception : implements the Inception module of the GoogleLeNet article ;
  • Collapse : just like nn.View(-1);
  • Convert : convert between different tensor types or shapes;
  • ZipTable : zip a table of tables into a table of tables;
  • ZipTableOneToMany : zip a table of element el and table of elements into a table of pairs of element el and table elements;
  • CAddTensorTable : adds a tensor to a table of tensors of the same size;
  • ReverseTable : reverse the order of elements in a table;
  • PrintSize : prints the size of inputs and gradOutputs (useful for debugging);
  • Clip : clips the inputs to a min and max value;
  • Constant : outputs a constant value given an input (which is ignored);
  • SpatialUniformCrop : uniformly crops patches from a input;
  • SpatialGlimpse : takes a fovead glimpse of an image at a given location;
  • WhiteNoise : adds isotropic Gaussian noise to the signal when in training mode;
  • OneHot : transforms a tensor of indices into one-hot encoding;
  • Kmeans : Kmeans clustering layer. Forward computes distances with respect to centroids and returns index of closest centroid. Centroids can be updated using gradient descent. Centroids could be initialized randomly or by using kmeans++ algoirthm;
  • SpatialRegionDropout : Randomly dropouts a region (top, bottom, leftmost, rightmost) of the input image. Works with batch and any number of channels;
  • FireModule : FireModule as mentioned in the SqueezeNet;
  • NCEModule : optimized placeholder for a Linear + SoftMax using noise-contrastive estimation.
  • SpatialFeatNormalization : Module for widely used preprocessing step of mean zeroing and standardization for images.
  • SpatialBinaryConvolution : Module for binary spatial convolution (Binary weights) as mentioned in XNOR-Net.
  • SimpleColorTransform : Module for adding independent random noise to input image channels.
  • PCAColorTransform : Module for adding noise to input image using Principal Components Analysis.

The following modules and criterions can be used to implement the REINFORCE algorithm :

Additional differentiable criterions

A lot of the functionality implemented here was pulled from dp, which makes heavy use of this package. However, dpnn can be used without dp (for e.g. you can use it with optim), which is one of the main reasons why we made it.

Tutorials

Sagar Waghmare wrote a nice tutorial on how to use dpnn with nngraph to reproduce the Lateral Connections in Denoising Autoencoders Support Supervised Learning.

A brief (1 hours) overview of Torch7, which includes some details about dpnn, is available via this NVIDIA GTC Webinar video. In any case, this presentation gives a nice overview of Logistic Regression, Multi-Layer Perceptrons, Convolutional Neural Networks and Recurrent Neural Networks using Torch7.

Module

The Module interface has been further extended with methods that facilitate stochastic gradient descent like updateGradParameters (i.e. momentum learning), weightDecay, maxParamNorm (for regularization), and so on.

Module.dpnn_parameters

A table that specifies the name of parameter attributes. Defaults to {'weight', 'bias'}, which is a static variable (i.e. table exists in class namespace). Sub-classes can define their own table statically.

Module.dpnn_gradParameters

A table that specifies the name of gradient w.r.t. parameter attributes. Defaults to {'gradWeight', 'gradBias'}, which is a static variable (i.e. table exists in class namespace). Sub-classes can define their own table statically.

[self] Module:type(type_str)

This function converts all the parameters of a module to the given type_str. The type_str can be one of the types defined for torch.Tensor like torch.DoubleTensor, torch.FloatTensor and torch.CudaTensor. Unlike the type method defined in nn, this one was overriden to maintain the sharing of storage among Tensors. This is especially useful when cloning modules share parameters and gradParameters.

[clone] Module:sharedClone([shareParams, shareGradParams])

Similar to clone. Yet when shareParams = true (the default), the cloned module will share the parameters with the original module. Furthermore, when shareGradParams = true (the default), the clone module will share the gradients w.r.t. parameters with the original module. This is equivalent to :

clone = mlp:clone()
clone:share(mlp, 'weight', 'bias', 'gradWeight', 'gradBias')

yet it is much more efficient, especially for modules with lots of parameters, as these Tensors aren't needlessly copied during the clone. This is particularly useful for Recurrent neural networks which require efficient copies with shared parameters and gradient w.r.t. parameters for each time-step.

Module:maxParamNorm([maxOutNorm, maxInNorm])

This method implements a hard constraint on the upper bound of the norm of output and/or input neuron weights (Hinton et al. 2012, p. 2) . In a weight matrix, this is a contraint on rows (maxOutNorm) and/or columns (maxInNorm), respectively. Has a regularization effect analogous to weightDecay, but with easier to optimize hyper-parameters. Assumes that parameters are arranged (output dim x ... x input dim). Only affects parameters with more than one dimension. The method should normally be called after updateParameters. It uses the C/CUDA optimized torch.renorm function. Hint : maxOutNorm = 2 usually does the trick.

[momGradParams] Module:momentumGradParameters()

Returns a table of Tensors (momGradParams). For each element in the table, a corresponding parameter (params) and gradient w.r.t. parameters (gradParams) is returned by a call to parameters. This method is used internally by updateGradParameters.

Module:updateGradParameters(momFactor [, momDamp, momNesterov])

Applies classic momentum or Nesterov momentum (Sutskever, Martens et al, 2013) to parameter gradients. Each parameter Tensor (params) has a corresponding Tensor of the same size for gradients w.r.t. parameters (gradParams). When using momentum learning, another Tensor is added for each parameter Tensor (momGradParams). This method should be called before updateParameters as it affects the gradients w.r.t. parameters.

Classic momentum is computed as follows :

momGradParams = momFactor*momGradParams + (1-momDamp)*gradParams
gradParams = momGradParams

where momDamp has a default value of momFactor.

Nesterov momentum (momNesterov = true) is computed as follows (the first line is the same as classic momentum):

momGradParams = momFactor*momGradParams + (1-momDamp)*gradParams
gradParams = gradParams + momFactor*momGradParams

The default is to use classic momentum (momNesterov = false).

Module:weightDecay(wdFactor [, wdMinDim])

Decays the weight of the parameterized models. Implements an L2 norm loss on parameters with dimensions greater or equal to wdMinDim (default is 2). The resulting gradients are stored into the corresponding gradients w.r.t. parameters. Such that this method should be called before updateParameters.

Module:gradParamClip(cutoffNorm [, moduleLocal])

Implements a contrainst on the norm of gradients w.r.t. parameters (Pascanu et al. 2012). When moduleLocal = false (the default), the norm is calculated globally to Module for which this is called. So if you call it on an MLP, the norm is computed on the concatenation of all parameter Tensors. When moduleLocal = true, the norm constraint is applied to the norm of all parameters in each component (non-container) module. This method is useful to prevent the exploding gradient in Recurrent neural networks.

Module:reinforce(reward)

This method is used by Criterions that implement the REINFORCE algorithm like VRClassReward. While vanilla backpropagation (gradient descent using the chain rule), REINFORCE Criterions broadcast a reward to all REINFORCE modules between the forward and the backward. In this way, when the following call to backward reaches the REINFORCE modules, these will compute a gradInput using the broadcasted reward. The reward is broadcast to all REINFORCE modules contained within model by calling model:reinforce(reward). Note that the reward should be a 1D tensor of size batchSize, i.e. each example in a batch has its own scalar reward.

Refer to this example for a complete training script making use of the REINFORCE interface.

Decorator

dmodule = nn.Decorator(module)

This module is an abstract class used to decorate a module. This means that method calls to dmodule will call the same method on the encapsulated module, and return its results.

DontCast

dmodule = nn.DontCast(module)

This module is a decorator. Use it to decorate a module that you don't want to be cast when the type() method is called.

module = nn.DontCast(nn.Linear(3,4):float())
module:double()
th> print(module:forward(torch.FloatTensor{1,2,3}))
 1.0927
-1.9380
-1.8158
-0.0805
[torch.FloatTensor of size 4]

Serial

dmodule = nn.Serial(module, [tensortype])
dmodule:[light,medium,heavy]Serial()

This module is a decorator that can be used to control the serialization/deserialization behavior of the encapsulated module. Basically, making the resulting string or file heavy (the default), medium or light in terms of size.

Furthermore, when specified, the tensortype attribute (e.g torch.FloatTensor, torch.DoubleTensor and so on.), determines what type the module will be cast to during serialization. Note that this will also be the type of the deserialized object. The default serialization tensortype is nil, i.e. the module is serialized as is.

The heavySerial() has the serialization process serialize every attribute in the module graph, which is the default behavior of nn.

The mediumSerial() has the serialization process serialize everything except the attributes specified in each module's dpnn_mediumEmpty table, which has a default value of {'output', 'gradInput', 'momGradParams', 'dpnn_input'}. During serialization, whether they be tables or Tensors, these attributes are emptied (no storage). Some modules overwrite the default Module.dpnn_mediumEmpty static attribute with their own.

The lightSerial() has the serialization process empty
everything a call to mediumSerial(type) would (so it uses dpnn_mediumEmpty). But also empties all the parameter gradients specified by the attribute dpnn_gradParameters, which defaults to {gradWeight, gradBias}.

We recomment using mediumSerial() for training, and lightSerial() for production (feed-forward-only models).

NaN

dmodule = nn.NaN(module, [id])

The NaN module asserts that the output and gradInput of the decorated module do not contain NaNs. This is useful for locating the source of those pesky NaN errors. The id defaults to automatically incremented values of 1,2,3,....

For example :

linear = nn.Linear(3,4)
mlp = nn.Sequential()
mlp:add(nn.NaN(nn.Identity()))
mlp:add(nn.NaN(linear))
mlp:add(nn.NaN(nn.Linear(4,2)))
print(mlp)

As you can see the NaN layers are have unique ids :

nn.Sequential {
  [input -> (1) -> (2) -> (3) -> output]
  (1): nn.NaN(1) @ nn.Identity
  (2): nn.NaN(2) @ nn.Linear(3 -> 4)
  (3): nn.NaN(3) @ nn.Linear(4 -> 2)
}

And if we fill the bias of the linear module with NaNs and call forward:

nan = math.log(math.log(0)) -- this is a nan value
linear.bias:fill(nan)
mlp:forward(torch.randn(2,3))

We get a nice error message:

/usr/local/share/lua/5.1/dpnn/NaN.lua:39: NaN found in parameters of module :
nn.NaN(2) @ nn.Linear(3 -> 4)

Inception

References :

module = nn.Inception(config)

This module uses n+2 parallel "columns". The original paper uses 2+2 where the first two are (but there could be more than two):

  • 1x1 conv (reduce) -> relu -> 5x5 conv -> relu
  • 1x1 conv (reduce) -> relu -> 3x3 conv -> relu

and where the other two are :

  • 3x3 maxpool -> 1x1 conv (reduce/project) -> relu
  • 1x1 conv (reduce) -> relu.

This module allows the first group of columns to be of any number while the last group consist of exactly two columns. The 1x1 convoluations are used to reduce the number of input channels (or filters) such that the capacity of the network doesn't explode. We refer to these here has reduce. Since each column seems to have one and only one reduce, their initial configuration options are specified in lists of n+2 elements.

The sole argument config is a table taking the following key-values :

  • Required Arguments :
    • inputSize : number of input channels or colors, e.g. 3;
    • outputSize : numbers of filters in the non-1x1 convolution kernel sizes, e.g. {32,48}
    • reduceSize : numbers of filters in the 1x1 convolutions (reduction) used in each column, e.g. {48,64,32,32}. The last 2 are used respectively for the max pooling (projection) column (the last column in the paper) and the column that has nothing but a 1x1 conv (the first column in the paper). This table should have two elements more than the outputSize
  • Optional Arguments :
    • reduceStride : strides of the 1x1 (reduction) convolutions. Defaults to {1,1,...}.
    • transfer : transfer function like nn.Tanh,nn.Sigmoid, nn.ReLU, nn.Identity, etc. It is used after each reduction (1x1 convolution) and convolution. Defaults to nn.ReLU.
    • batchNorm : set this to true to use batch normalization. Defaults to false. Note that batch normalization can be awesome
    • padding : set this to true to add padding to the input of the convolutions such that output width and height are same as that of the original non-padded input. Defaults to true.
    • kernelSize : size (height = width) of the non-1x1 convolution kernels. Defaults to {5,3}.
    • kernelStride : stride of the kernels (height = width) of the convolution. Defaults to {1,1}
    • poolSize: size (height = width) of the spatial max pooling used in the next-to-last column. Defaults to 3.
    • poolStride : stride (height = width) of the spatial max pooling. Defaults to 1.

For a complete example using this module, refer to the following :

Collapse

module = nn.Collapse(nInputDim)

This module is the equivalent of:

view = nn.View(-1)
view:setNumInputDim(nInputDim)

It collapses all non-batch dimensions. This is useful for converting a spatial feature map to the single dimension required by a dense hidden layer like Linear.

Convert

module = nn.Convert([inputShape, outputShape])

Module to convert between different data formats. For example, we can flatten images by using :

module = nn.Convert('bchw', 'bf')

or equivalently

module = nn.Convert('chw', 'f')

Lets try it with an input:

print(module:forward(torch.randn(3,2,3,1)))
 0.5692 -0.0190  0.5243  0.7530  0.4230  1.2483
-0.9142  0.6013  0.5608 -1.0417 -1.4014  1.0177
-1.5207 -0.1641 -0.4166  1.4810 -1.1725 -1.0037
[torch.DoubleTensor of size 3x6]

You could also try:

module = nn.Convert('chw', 'hwc')
input = torch.randn(1,2,3,2)
input:select(2,1):fill(1)
input:select(2,2):fill(2)
print(input)
(1,1,.,.) = 
  1  1
  1  1
  1  1
(1,2,.,.) = 
  2  2
  2  2
  2  2
[torch.DoubleTensor of size 1x2x3x2]
print(module:forward(input))
(1,1,.,.) = 
  1  2
  1  2

(1,2,.,.) = 
  1  2
  1  2

(1,3,.,.) = 
  1  2
  1  2
[torch.DoubleTensor of size 1x3x2x2]

Furthermore, it automatically converts the input to have the same type as self.output (i.e. the type of the module). So you can also just use is for automatic input type converions:

module = nn.Convert()
print(module.output) -- type of module
[torch.DoubleTensor with no dimension]
input = torch.FloatTensor{1,2,3}
print(module:forward(input))
 1
 2
 3
[torch.DoubleTensor of size 3]

ZipTable

module = nn.ZipTable()

Zips a table of tables into a table of tables.

Example:

print(module:forward{ {'a1','a2'}, {'b1','b2'}, {'c1','c2'} })
{ {'a1','b1','c1'}, {'a2','b2','c2'} }

ZipTableOneToMany

module = nn.ZipTableOneToMany()

Zips a table of element el and table of elements tab into a table of tables, where the i-th table contains the element el and the i-th element in table tab

Example:

print(module:forward{ 'el', {'a','b','c'} })
{ {'el','a'}, {'el','b'}, {'el','c'} }

CAddTensorTable

module = nn.CAddTensorTable()

Adds the first element el of the input table tab to each tensor contained in the second element of tab, which is itself a table

Example:

print(module:forward{ (0,1,1), {(0,0,0),(1,1,1)} })
{ (0,1,1), (1,2,2) }

ReverseTable

module = nn.ReverseTable()

Reverses the order of elements in a table.

Example:

print(module:forward{1,2,3,4})
{4,3,2,1}

PrintSize

module = nn.PrintSize(name)

This module is useful for debugging complicated module composites. It prints the size of the input and gradOutput during forward and backward propagation respectively. The name is a string used to identify the module along side the printed size.

Clip

module = nn.Clip(minval, maxval)

This module clips input values such that the output is between minval and maxval.

Constant

module = nn.Constant(value, nInputDim)

This module outputs a constant value given an input. If nInputDim is specified, it uses the input to determine the size of the batch. The value is then replicated over the batch. Otherwise, the value Tensor is output as is. During backward, the returned gradInput is a zero Tensor of the same size as the input. This module has no trainable parameters.

You can use this with nn.ConcatTable() to append constant inputs to an input :

nn.ConcatTable():add(nn.Constant(v)):add(nn.Identity())

This is useful when you want to output a value that is independent of the input to the neural network (see this example).

SpatialUniformCrop

module = nn.SpatialUniformCrop(oheight, owidth)

During training, this module will output a cropped patch of size oheight, owidth within the boundaries of the input image. For each example, a location is sampled from a uniform distribution such that each possible patch has an equal probability of being sampled.

During evaluation, the center patch is cropped and output.

This module is commonly used at the input layer to artificially augment the size of the dataset to prevent overfitting.

SpatialGlimpse

Ref. A. Recurrent Model for Visual Attention

module = nn.SpatialGlimpse(size, depth, scale)

A glimpse is the concatenation of down-scaled cropped images of increasing scale around a given location in a given image. The input is a pair of Tensors: {image, location} location are (y,x) coordinates of the center of the different scales of patches to be cropped from image image. Coordinates are between (-1,-1) (top-left) and (1,1) (bottom-right). The output is a batch of glimpses taken in image at location (y,x).

size can be either a scalar which specifies the width = height of glimpses, or a table of {height, width} to support a rectangular shape of glimpses. depth is number of patches to crop per glimpse (one patch per depth). scale determines the size(t) = scale * size(t-1) of successive cropped patches.

So basically, this module can be used to focus the attention of the model on a region of the input image. It is commonly used with the RecurrentAttention module (see this example).

WhiteNoise

module = nn.WhiteNoise([mean, stdev])

Useful in training Denoising Autoencoders. Takes mean and stdev of the normal distribution as input. Default values for mean and standard deviation are 0 and 0.1 respectively. With module:training(), noise is added during forward. During backward gradients are passed as it is. With module:evaluate() the mean is added to the input.

SpatialRegionDropout

module = nn.SpatialRegionDropout(p)

Following is an example of SpatialRegionDropout outputs on the famous lena image.

Input

Lena

Outputs

Lena Lena

FireModule

Ref: http://arxiv.org/pdf/1602.07360v1.pdf

module = nn.FireModule(nInputPlane, s1x1, e1x1, e3x3, activation)

FireModule is comprised of two submodules 1) A squeeze convolution module comprised of 1x1 filters followed by 2) an expand module that is comprised of a mix of 1x1 and 3x3 convolution filters. Arguments: s1x1: number of 1x1 filters in the squeeze submodule, e1x1: number of 1x1 filters in the expand submodule, e3x3: number of 3x3 filters in the expand submodule. It is recommended that s1x1 be less than (e1x1+e3x3) if you want to limit the number of input channels to the 3x3 filters in the expand submodule. FireModule works only with batches, for single sample convert the sample to a batch of size 1.

SpatialFeatNormalization

module = nn.SpatialFeatNormalization(mean, std)

This module normalizies each feature channel of input image based on its corresponding mean and standard deviation scalar values. This module does not learn the mean and std, they are provided as arguments.

SpatialBinaryConvolution

module = nn.SpatialBinaryConvolution(nInputPlane, nOutputPlane, kW, kH)

Functioning of SpatialBinaryConvolution is similar to nn/SpatialConvolution. Only difference is that Binary weights are used for forward/backward and floating point weights are used for weight updates. Check Binary-Weight-Network section of XNOR-net.

SimpleColorTransform

range = torch.rand(inputChannels) -- Typically range is specified by user.
module = nn.SimpleColorTransform(inputChannels, range)

This module performs a simple data augmentation technique. SimpleColorTransform module adds random noise to each color channel independently. In more advanced data augmentation technique noise is added using principal components of color channels. For that please check PCAColorTransform

PCAColorTransform

eigenVectors = torch.rand(inputChannels, inputChannels) -- Eigen Vectors
eigenValues = torch.rand(inputChannels) -- Eigen
std = 0.1 -- Std deviation of normal distribution with mean zero for noise.
module = nn.PCAColorTransform(inputChannels, eigenVectors, eigenValues, std)

This module performs a data augmentation using Principal Component analysis of pixel values. When in training mode, mulitples of principal components are added to input image pixels. Magnitude of value added (noise) is dependent upon the corresponding eigen value and a random value sampled from a Gaussian distribution with mean zero and std (default 0.1) standard deviation. This technique was used in the famous AlexNet paper.

OneHot

module = nn.OneHot(outputSize)

Transforms a tensor of input indices having integer values between 1 and outputSize into a tensor of one-hot vectors of size outputSize.

Forward an index to get a one-hot vector :

> module = nn.OneHot(5) -- 5 classes
> module:forward(torch.LongTensor{3})
 0  0  1  0  0
[torch.DoubleTensor of size 1x5]

Forward a batch of 3 indices. Notice that these need not be stored as torch.LongTensor :

> module:forward(torch.Tensor{3,2,1})
 0  0  1  0  0
 0  1  0  0  0
 1  0  0  0  0
[torch.DoubleTensor of size 3x5]

Forward batch of 2 x 3 indices :

oh:forward(torch.Tensor{{3,2,1},{1,2,3}})
(1,.,.) = 
  0  0  1  0  0
  0  1  0  0  0
  1  0  0  0  0

(2,.,.) = 
  1  0  0  0  0
  0  1  0  0  0
  0  0  1  0  0
[torch.DoubleTensor of size 2x3x5]

Kmeans

km = nn.Kmeans(k, dim)

k is the number of centroids and dim is the dimensionality of samples. You can either initialize centroids randomly from input samples or by using kmeans++ algorithm.

km:initRandom(samples) -- Randomly initialize centroids from input samples.
km:initKmeansPlus(samples) -- Use Kmeans++ to initialize centroids.

Example showing how to use Kmeans module to do standard Kmeans clustering.

attempts = 10
iter = 100 -- Number of iterations
bestKm = nil
bestLoss = math.huge
learningRate = 1
for j=1, attempts do
   local km = nn.Kmeans(k, dim)
   km:initKmeansPlus(samples)
   for i=1, iter do
      km:zeroGradParameters()
      km:forward(samples) -- sets km.loss
      km:backward(samples, gradOutput) -- gradOutput is ignored

      -- Gradient Descent weight/centroids update
      km:updateParameters(learningRate)
   end

   if km.loss < bestLoss then
      bestLoss = km.loss
      bestKm = km:clone()
   end
end

nn.Kmeans() module maintains loss only for the latest forward. If you want to maintain loss over the whole dataset then you who would need do it my adding the module loss for every forward.

You can also use nn.Kmeans() as an auxillary layer in your network. A call to forward will generate an output containing the index of the nearest cluster for each sample in the batch. The gradInput generated by updateGradInput will be zero.

ModuleCriterion

criterion = nn.ModuleCriterion(criterion [, inputModule, targetModule, castTarget])

This criterion decorates a criterion by allowing the input and target to be fed through an optional inputModule and targetModule before being passed to the criterion. The inputModule must not contain parameters as these would not be updated.

When castTarget = true (the default), the targetModule is cast along with the inputModule and criterion. Otherwise, the targetModule isn't.

NCEModule

Ref. A RNNLM training with NCE for Speech Recognition

ncem = nn.NCEModule(inputSize, outputSize, k, unigrams, [Z])

When used in conjunction with NCECriterion, the NCEModule implements noise-contrastive estimation.

The point of the NCE is to speedup computation for large Linear + SoftMax layers. Computing a forward/backward for Linear(inputSize, outputSize) for a large outputSize can be very expensive. This is common when implementing language models having with large vocabularies of a million words. In such cases, NCE can be an efficient alternative to computing the full Linear + SoftMax during training and cross-validation.

The inputSize and outputSize are the same as for the Linear module. The number of noise samples to be drawn per example is k. A value of 25 should work well. Increasing it will yield better results, while a smaller value will be more efficient to process. The unigrams is a tensor of size outputSize that contains the frequencies or probability distribution over classes. It is used to sample noise samples via a fast implementation of torch.multinomial. The Z is the normalization constant of the approximated SoftMax. The default is math.exp(9) as specified in Ref. A.

For inference, or measuring perplexity, the full Linear + SoftMax will need to be computed. The NCEModule can do this by switching on the following :

ncem:evaluate()
ncem.normalized = true

Furthermore, to simulate Linear + LogSoftMax instead, one need only add the following to the above:

ncem.logsoftmax = true

An example is provided via the rnn package.

NCECriterion

ncec = nn.NCECriterion()

This criterion only works with an NCEModule on the output layer. Together, they implement noise-contrastive estimation.

Reinforce

Ref A. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

Abstract class for modules that implement the REINFORCE algorithm (ref. A).

module = nn.Reinforce([stochastic])

The reinforce(reward) method is called by a special Reward Criterion (e.g. VRClassReward). After which, when backward is called, the reward will be used to generate gradInputs. When stochastic=true, the module is stochastic (i.e. samples from a distribution) during evaluation and training. When stochastic=false (the default), the module is only stochastic during training.

The REINFORCE rule for a module can be summarized as follows :

            d ln(f(output,input))
gradInput = ---------------------  * reward
                  d input

where the reward is what is provided by a Reward criterion like VRClassReward via the reinforce method. The criterion will normally be responsible for the following formula :

reward = a*(R - b)

where a is the alpha of the original paper, i.e. a reward scale, R is the raw reward (usually 0 or 1), and b is the baseline reward, which is often taken to be the expected raw reward R.

The output is usually sampled from a probability distribution f() parameterized by the input. See ReinforceBernoulli for a concrete derivation.

Also, as you can see, the gradOutput is ignored. So within a backpropagation graph, the Reinforce modules will replace the backpropagated gradients (gradOutput) with their own obtained from the broadcasted reward.

ReinforceBernoulli

Ref A. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

module = nn.ReinforceBernoulli([stochastic])

A Reinforce subclass that implements the REINFORCE algorithm (ref. A p.230-236) for the Bernoulli probability distribution. Inputs are bernoulli probabilities p. During training, outputs are samples drawn from this distribution. During evaluation, when stochastic=false, outputs are the same as the inputs. Uses the REINFORCE algorithm (ref. A p.230-236) which is implemented through the reinforce interface (gradOutputs are ignored).

Given the following variables :

  • f : bernoulli probability mass function
  • x : the sampled values (0 or 1) (i.e. self.output)
  • p : probability of sampling a 1

the derivative of the log bernoulli w.r.t. probability p is :

d ln(f(output,input))   d ln(f(x,p))    (x - p)
--------------------- = ------------ = ---------
      d input               d p         p(1 - p)

ReinforceNormal

Ref A. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

module = nn.ReinforceNormal(stdev, [stochastic])

A Reinforce subclass that implements the REINFORCE algorithm (ref. A p.238-239) for a Normal (i.e. Gaussian) probability distribution. Inputs are the means of the normal distribution. The stdev argument specifies the standard deviation of the distribution. During training, outputs are samples drawn from this distribution. During evaluation, when stochastic=false, outputs are the same as the inputs, i.e. the means. Uses the REINFORCE algorithm (ref. A p.238-239) which is implemented through the reinforce interface (gradOutputs are ignored).

Given the following variables :

  • f : normal probability density function
  • x : the sampled values (i.e. self.output)
  • u : mean (input)
  • s : standard deviation (self.stdev)

the derivative of log normal w.r.t. mean u is :

d ln(f(x,u,s))   (x - u)
-------------- = -------
     d u           s^2

As an example, it is used to sample locations for the RecurrentAttention module (see this example).

ReinforceGamma

Ref A. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

module = nn.ReinforceGamma(scale, [stochastic])

A Reinforce subclass that implements the REINFORCE algorithm (ref. A) for a Gamma probability distribution parametrized by shape (k) and scale (theta) variables. Inputs are the shapes of the gamma distribution. During training, outputs are samples drawn from this distribution. During evaluation, when stochastic=false, outputs are equal to the mean, defined as the product of shape and scale ie. k*theta. Uses the REINFORCE algorithm (ref. A) which is implemented through the reinforce interface (gradOutputs are ignored).

Given the following variables :

  • f : gamma probability density function
  • g : digamma function
  • x : the sampled values (i.e. self.output)
  • k : shape (input)
  • t : scale

the derivative of log gamma w.r.t. shape k is :

d ln(f(x,k,t))
-------------- = ln(x) - g(k) - ln(t)
      d k

ReinforceCategorical

Ref A. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

module = nn.ReinforceCategorical([stochastic])

A Reinforce subclass that implements the REINFORCE algorithm (ref. A) for a Categorical (i.e. Multinomial with one sample) probability distribution. Inputs are the categorical probabilities of the distribution : p[1], p[2], ..., p[k]. These are usually the output of a SoftMax.
For n categories, both the input and output ares of size batchSize x n. During training, outputs are samples drawn from this distribution. The outputs are returned in one-hot encoding i.e. the output for each example has exactly one category having a 1, while the remainder are zero. During evaluation, when stochastic=false, outputs are the same as the inputs, i.e. the probabilities p. Uses the REINFORCE algorithm (ref. A) which is implemented through the reinforce interface (gradOutputs are ignored).

Given the following variables :

  • f : categorical probability mass function
  • x : the sampled indices (one per sample) (self.output is the one-hot encoding of these indices)
  • p : probability vector (p[1], p[2], ..., p[k]) (input)

the derivative of log categorical w.r.t. probability vector p is :

d ln(f(x,p))     1/p[i]    if i = x  
------------ =   
    d p          0         otherwise

VRClassReward

Ref A. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning

This Reward criterion implements the REINFORCE algoritm (ref. A) for classification models. Specifically, it is a Variance Reduces (VR) classification reinforcement leanring (reward-based) criterion.

vcr = nn.VRClassReward(module [, scale, criterion])

While it conforms to the Criterion interface (which it inherits), it does not backpropagate gradients (except for the baseline b; see below). Instead, a reward is broadcast to the module via the reinforce method.

The criterion implements the following formula :

reward = a*(R - b)

where a is the alpha described in Ref. A, i.e. a reward scale (defaults to 1), R is the raw reward (0 for incorrect and 1 for correct classification), and b is the baseline reward, which is often taken to be the expected raw reward R.

The target of the criterion is a tensor of class indices. The input to the criterion is a table {y,b} where y is the probability (or log-probability) of classes (usually the output of a SoftMax), and b is the baseline reward discussed above.

For each example, if argmax(y) is equal to the target class, the raw reward R = 1, otherwize R = 0.

As for b, its gradInputs are obtained from the criterion, which defaults to MSECriterion. The criterion's target is the commensurate raw reward R. Using a*(R-b) instead of a*R to obtain a reward is what makes this class variance reduced (VR). By reducing the variance, the training can converge faster (Ref. A). The predicted b can be nothing more than the expectation E(R).

Note : for RNNs with R = 1 for last step in sequence, encapsulate it in nn.ModuleCriterion(VRClassReward, nn.SelectTable(-1)).

For an example, this criterion is used along with the RecurrentAttention module to train a recurrent model for visual attention.

BinaryLogisticRegression

Ref A. Learning to Segment Object Candidates This criterion implements the score criterion mentioned in (ref. A).

criterion = nn.BinaryLogisticRegression()

BinaryLogisticRegression implements following cost function for binary classification.


 log( 1 + exp( -y_k * score(x_k) ) )

where y_k is binary target score(x_k) is the corresponding prediction. y_k has value {-1, +1} and score(x_k) has value in [-1, +1].

SpatialBinaryLogisticRegression

Ref A. Learning to Segment Object Candidates

This criterion implements the spatial component of the criterion mentioned in (ref. A).

criterion = nn.SpatialBinaryLogisticRegression()

SpatialBinaryLogisticRegression implements following cost function for binary pixel classification.

   1
_______ sum_ij [ log( 1 + exp( -m_ij * f_ij ) ) ]
 2*w*h

where m_ij is target binary image and f_ij is the corresponding prediction. m_ij has value {-1, +1} and f_ij has value in [-1, +1].