cavis/docs/samediff/templates/ops.md

17 KiB

title short_title description category weight
Operations in SameDiff Ops What kind of operations is there in `SameDiff` and how to use them SameDiff 4

SameDiff operations

Operations in SameDiff work mostly the way you'd expect them to. You take variables - in our framework, those are objects of type SDVariable - apply operations to them, and thus produce new variables. Before we proceed to the overview of the available operations, let us list some of their common properties.

Common properties of operations

  • Variables of any variable type may be used in any operation, as long as their data types match those that are required by the operation (again, see our variables section for what variable types are). Most often an operation will require its SDVariable to have a floating point data type.
  • Variables created by operations have ARRAY variable type.
  • For all operations, you may define a String name of your resulting variable, although for most operations this is not obligatory. The name goes as the first argument in each operation, like so:
SDVariable linear = weights.mmul("matrix_product", input).add(bias); 
SDVariable output = sameDiff.nn.sigmoid("output", linear);

Named variables may be accessed from outside using a SameDiff method getVariable(String name). For the code above, this method will allow you to infer the value of both output as well as the result of mmul operation. Note that we haven't even explicitly defined this result as a separate SDVariable, and yet a corresponding SDVariable will be created internally and added to our instance of SameDiff under the String name "matrix_product". In fact, a unique String name is given to every SDVariable you produce by operations: if you don't give a name explicitly, it is assigned to the resulting SDVariable automatically based on the operation's name.

Overview of operations

The number of currently available operations, including overloads totals several hundreds, they range in complexity from s imple additions and multiplications via producing outputs of convolutional layers to creation of dedicated recurrent neural network modules, and much more. The sheer number of operations would've made it cumbersome to list them all on a single page. So, if you are already looking for something specific, you'll be better off checking our javadoc, which already contains a detailed information on each operation, or by simply browsing through autocompletion suggestions (if your IDE supports that). Here we rather try to give you an idea of what operations you may expect to find and where to seek for them.

All operations may be split into two major branches: those which are methods of SDVariable and those of SameDiff classes. Let us have a closer look at each:

SDVariable operations

We have already seen SDVariable operations in previous examples, in expressions like

SDVariable z = x.add(y);

where x and y are SDVariable's.

Among SDVariable methods, you will find:

  • BLAS-type operations to perform linear algebra: things like add, neg, mul (used for both scaling and elementwise multiplication) and mmul (matrix multiplication), dot, rdiv, etc.;
  • comparison operations like gt or lte, used both to compare each element to a fixed double value as well as for elementwise comparison with another SDVariable of the same shape, and alike;
  • basic reduction operations: things like min, sum, prod (product of elements in array), mean, norm2, argmax (index of the maximal element), squaredDifference and so on, which may be taken along specified dimensions;
  • basic statistics operations for computing mean and standard deviation along given dimensions: mean and std.
  • operations for restructuring of the underlying array: reshape and permute, along with shape - an operation that delivers the shape of a variable as an array of integers - the dimension sizes;

SDVariable operations may be easily chained, producing lines like:

SDVariable regressionCost = weights.mmul(input).add("regression_prediction", bias).squaredDifference(labels);

SameDiff operations

The operations that are methods of SameDiff are called via one of 6 auxiliary objects present in each SameDiff, which split all operations into 6 uneven branches:

  • math - for general mathematical operations;
  • random - creating different random number generators;
  • nn - general neural network tools;
  • cnn - convolutional neural network tools;
  • rnn - recurrent neural network tools;
  • loss - loss functions; In order to use a particular operation, you need to call one of these 6 objects form your SameDiff instance, and then an operation itself, like that:
SDVariable y = sameDiff.math.sin(x);

or

SDVariable y = samediff.math().sin(x);

The distribution of operations among the auxiliary objects has no structural bearing beyond organizing things in a more intuitive way. So, for instance, if you're not sure whether to seek for, say, tanh operation in math or in nn, don't worry: we have it in both.

Let us briefly describe what kinds of operations you may expect to find in each of the branches:

math - basic mathematical operations

Math module mostly consists of general mathematical functions and statistics methods. Those include:

  • power functions, e.g. square, cube, sqrt, pow, reciprocal etc.;
  • trigonometric functions, e.g. sin, atan etc.;
  • exponential/hyperbolic functions, like exp, sinh, log, atanh etc.;
  • miscellaneous elementwise operations, like taking absolute value, rounding and clipping, such as abs, sign, ceil, round, clipByValue, clipByNorm etc.;
  • reductions along specified dimensions: min, amax, mean, asum, logEntropy, and similar;
  • distance (reduction) operations, such as euclideanDistance, manhattanDistance, jaccardDistance, cosineDistance, hammingDistance, cosineSimilarity, along specified dimensions, for two identically shaped SDVariables;
  • specific matrix operations: matrixInverse, matrixDeterminant, diag (creating a diagonal matrix), trace, eye (creating identity matrix with variable dimensions), and several others;
  • more statistics operations: standardize, moment, normalizeMoments, erf and erfc (Gaussian error function and its complementary);
  • counting and indexing reductions: methods like conuntZero (number of zero elements), iamin (index of the element with the smallest absolute value), firstIndex (an index of the first element satisfying a specified Condition function);
  • reductions indicating properties of the underlying arrays. These include e.g. isNaN (elementwise checking), isMax (shape-preserving along specified dimensions), isNonDecreasing (reduction along specified dimensions);
  • elementwise logical operations: and, or, xor, not.

Most operations in math have very simple structure, and are inferred like that:

SDVariable activation = sameDiff.math.cube(input);

Operations may be chained, although in a more cumbersome way in comparison to the SDVariable operations, e.g.:

SDVariable matrixNorm1 = sameDiff.math.max(sameDiff.math.sum(sameDiff.math.abs(matrix), 1));

Observe that the (integer) argument 1 in the sum operation tells us that we have to take maximum absolute value along the 1's dimension, i.e. the column of the matrix.

random - creating random values Random

These operations create variables whose underlying arrays will be filled with random numbers following some distribution

  • say, Bernoulli, normal, binomial etc.. These values will be reset at each iteration. If you wish, for instance, to create a variable that will add a Gaussian noise to entries of the MNIST database, you may do something like:
double mean = 0.;
double deviation = 0.05;
long[] shape = new long[28, 28];
SDVariable noise_mnist = sameDiff.random.normal("noise_mnist", mean, deviation, shape);

The shape of you random variable may vary. Suppose, for instance, that you have audio signals of varying length, and you want to add noise to them. Then, you need to specify an SDVariable, say, windowShape with an integer data type, and proceed like that

SDVariabel noise_audio = sameDiff.random.normal("noise_audio", mean, deviation, windowShape);

nn - general neural network tools

Here we store methods for neural networks that are not necessarily associated with convolutional ones. Among them are

  • creation of dense linear and ReLU layers (with or without bias), and separate bias addition: linear, reluLayer, biasAdd;
  • popular activation functions, e.g. relu, sigmoid, tanh, softmax as well as their less used versions like leakyRelu, elu, hardTanh, and many more;
  • padding for 2d arrays with method pad, supporting several padding types, with both constant and variable padding width;
  • explosion/overfitting prevention, such as dropout, layerNorm and batchNorm for layer resp. batch normalization;

Some methods were created for internal use, but are openly available. Those include:

  • derivatives for several popular activation functions - these are mostly designed for speeding up backpropagation;
  • attention modules - basically, building blocks for recurrent neural networks we shall discuss below.

While activations in nn are fairly simple, other operations become more involved. Say, to create a linear or a ReLU layer, up to three predefined SDVariable objects may be required, as in the following code:

SDVariable denseReluLayer = sameDiff.nn.reluLayer(input, weights, bias);

where input, weights and bias need to have dimensions suiting each other.

To create, say, a dense layer with softmax activation, you may proceed as follows:

SDVariable linear = sameDiff.nn.linear(input, weight, bias);
SDVariable output = sameDiff.nn.softmax(linear);

cnn - convolutional neural networks tools

The cnn module contains layers and operations typically used in convolutional neural networks - different activations may be picked up from the nn module. Among cnn operations we currently have creation of:

  • linear convolution layers, currently for tensors of dimension up to 3 (minibatch not included): conv1d, conv2d, conv3d, depthWiseConv2d, separableConv2D/sconv2d;
  • linear deconvolution layers, currently deconv1d, deconv2d, deconv3d;
  • pooling, e.g. maxPoooling2D, avgPooling1D;
  • specialized reshaping methods: batchToSpace, spaceToDepth, col2Im and alike;
  • upsampling, currently presented by upsampling2d operation;
  • local response normalization: localResponseNormalization, currently for 2d convolutional layers only;

Convolution and deconvolution operations are specified by a number of static parameters like kernel size, dilation, having or not having bias etc.. To facilitate the creation process, we pack the required parameters into easily constructable and alterable configuration objects. Desired activations may be borrowed from the nn module. So, for example, if we want to create a 3x3 convolutional layer with relu activation, we may proceed as follows:

Conv2DConfig config2d = new Conv2DConfig().builder().kW(3).kH(3).pW(2).pH(2).build();
SDVariable convolution2dLinear = sameDiff.cnn.conv2d(input, weights, config2d);
SDVariable convolution2dOutput = sameDiff.nn.relu(convolution2dLinear);

In the first line, we construct a convolution configuration using its default constructor. Then we specify the kernel size (this is mandatory) and optional padding size, keeping other settings default (unit stride, no dilation, no bias, NCHW data format). We then employ this configuration to create a linear convolution with predefined SDVariables for input and weights; the shape of weights is to be tuned to that of input and to config beforehand. Thus, if in the above example input has shape, say, [-1, nIn, height, width], then weights are to have a form [nIn, nOut, 3, 3] (because we have 3x3 convolution kernel). The shape of the resulting variable convoluton2d will be predetermined by these parameters (in our case, it will be [-1, nOut, height, width]). Finally, in the last line we apply a relu activation.

rnn - Recurrent neural networks

This module contains arguably the most sophisticated methods in the framework. Currently it allows you to create

  • simple recurrent units, using sru and sruCell methods;
  • LSTM units, using lstmCell, lstmBlockCell and lstmLayer;
  • Graves LSTM units, using gru methods.

As of now, recurrent operations require special configuration objects as input, in which you need to pack all the variables that will be used in a unit. This is subject to change in the later versions. For instance, to create a simple recurrent unit, you need to proceed like that:

SRUConfiguration sruConfig = new SRUConfiguration(input, weights, bias, init);
SDVariable sruOutput = samediff.rnn().sru(sruConfig);

Here, the arguments in the SRUConfiguration constructor are variables that are to be defined beforehand. Obviously their shapes should be matching, and these shapes predetermine the shape of output.

loss - Loss functions

In this branch we keep common loss functions. Most loss functions may be created quite simply, like that:

SDVariable logLoss = sameDiff.loss.logLoss("logLoss", label, predictions);

where labels and predictions are SDVariable's. A String name is a mandatory parameter in most loss methods, yet it may be set to null - in this case, the name will be generated automatically. You may also create weighted loss functions by adding another SDVariable parameters containing weights, as well as specify a reduction method (see below) for the loss over the minibatch. Thus, a full-fledged logLoss operation may look like:

SDVariable wLogLossMean = sameDiff.loss.logLoss("wLogLossMean", label, predictions, weights, LossReduce.MEAN_BY_WEIGHT);

Some loss operations may allow/require further arguments, depending on their type: e.g. a dimension along which the loss is to be computed (as in cosineLoss), or some real-valued parameters.

As for reduction methods, over the minibatch, there are currently 4 of them available. Thus, initially loss values for each sample of the minibatch are computed, then they are multiplied by weights (if specified), and finally one of the following routines takes place:

  • NONE - leaving the resulting (weighted)loss values as-is; the result is an INDArray with the length of the minibatch: sum_loss = sum(weights * loss_per_sample).
  • SUM - summing the values, producing a scalar result.
  • MEAN_BY_WEIGHT - first computes the sum as above, and then divides it by the sum of all weights, producing a scalar value: mean_loss = sum(weights * loss_per_sample) / sum(weights). If weights are not specified, they all are set to 1.0 and this reduction is equivalent to getting mean loss value over the minibatch.
  • MEAN_BY_NONZERO_WEIGHT_COUNT - divides the weighted sum by the number of nonzero weight, producing a scalar: mean_count_loss = sum(weights * loss_per_sample) / count(weights != 0). Useful e.g. when you want to compute the mean only over a subset of valid samples, setting weights by either 0. or 1.. When weights are not given, it just produces mean, and thus equivalent to MEAN_BY_WEIGHT.

The don'ts of operations

In order for SameDiff operations to work properly, several main rules are to be upheld. Failing to do so may result in an exception or, worse even, to a working code producing undesired results. All the things we mention in the current section describe what you better not do.

  • All variables in an operation have to belong to the same instance of SamdeDiff (see the variables section on how variables are added to a SameDiff instance). In other words, you better not
SDVariable x = sameDiff0.var(DataType.FLOAT, 1);
SDVariable y = sameDiff1.placeHolder(DataType.FLOAT, 1);
SDVariable z = x.add(y);
  • At best, a new variable is to be created for a result of an operation or a chain of operations. In other words, you better not redefine existing variables and better not leave operations returning no result. In other words, try to avoid the code like this:
SDVariable z = x.add(y);
//DON'T!!!
z.mul(2);
x = z.mul(y);

A properly working version of the above code (if we've desired to obtain 2xy+2y2 in an unusual way) will be

SDVariable z = x.add(y);
SDVariable _2z = z.mul(2);
w = _2z.mul(y);

To learn more why it functions like that, see our graph section.