Fork me on GitHub


Layer is a core abstraction in SINGA. It performs a variety of feature transformations for extracting high-level features, e.g., loading raw features, parsing RGB values, doing convolution transformation, etc.

The Basic user guide section introduces the configuration of a built-in layer. Advanced user guide explains how to extend the base Layer class to implement users’ functions.

Basic user guide

Layer configuration

Configuration of two example layers are shown below,

layer {
  name: "data"
  type: kCSVRecord
  store_conf { }
  name: "fc1"
  type: kInnerProduct
  srclayers: "data"
  innerproduct_conf{ }
  param{ }

There are some common fields for all kinds of layers:

  • name: a string used to differentiate two layers in a neural net.
  • type: an integer used for identifying a specific Layer subclass. The types of built-in layers are listed in LayerType (defined in job.proto). For user-defined layer subclasses, user_type should be used instead of type.
  • srclayers: names of the source layers. In SINGA, all connections are converted to directed connections.
  • param: configuration for a Param instance. There can be multiple Param objects in one layer.

Different layers may have different configurations. These configurations are defined in <type>_conf. E.g., “fc1” layer has innerproduct_conf. The subsequent sections explain the functionality of each built-in layer and how to configure it.

Built-in Layer subclasses

SINGA has provided many built-in layers, which can be used directly to create neural nets. These layers are categorized according to their functionalities,

  • Input layers for loading records (e.g., images) from disk files, HDFS or network into memory.
  • Neuron layers for feature transformation, e.g., convolution, pooling, dropout, etc.
  • Loss layers for measuring the training objective loss, e.g., Cross Entropy loss or Euclidean loss.
  • Output layers for outputting the prediction results (e.g., probabilities of each category) or features into persistent storage, e.g., disk or HDFS.
  • Connection layers for connecting layers when the neural net is partitioned.

Input layers

Input layers load training/test data from disk or other places (e.g., HDFS or network) into memory.


StoreInputLayer is a base layer for loading data from data store. The data store can be a KVFile or TextFile (LMDB, LevelDB, HDFS, etc., will be supported later). Its ComputeFeature function reads batchsize (string:key, string:value) tuples. Each tuple is parsed by a Parse function implemented by its subclasses.

The configuration for this layer is in store_conf,

store_conf {
  backend: # "kvfile" or "textfile"
  path: # path to the data store
  batchsize : 32
  prefetching: true #default value is false

It is a subclass of StoreInputLayer. It assumes the (key, value) tuple loaded from a data store contains a feature vector (and a label) for one data instance. All feature vectors are of the same fixed length. The shape of one instance is configured through the shape field, e.g., the following configuration specifies the shape for the CIFAR10 images.

store_conf {
  shape: 3  #channels
  shape: 32 #height
  shape: 32 #width

It may do some preprocessing like standardization. The data for preprocessing is loaded by and parsed in a virtual function, which is implemented by its subclasses.


It is a subclass of SingleLabelRecordLayer. It parses the value field from one tuple into a RecordProto, which is generated by Google Protobuf according to common.proto. It can be used to store features for images (e.g., using the pixel field) or other objects (using the data field). The key field is not parsed.

type: kRecordInput
store_conf {
  has_label: # default is true

It is a subclass of SingleLabelRecordLayer. The value field from one tuple is parsed as a CSV line (separated by comma). The first number would be parsed as a label if has_label is configured in store_conf. Otherwise, all numbers would be parsed into one row of the data_ Blob.

type: kCSVInput
store_conf {
  has_label: # default is true

This layer does image preprocessing, e.g., cropping, mirroring and scaling, against the data Blob from its source layer. It deprecates the RGBImageLayer which works on the Record from ShardDataLayer. It still uses the same configuration as RGBImageLayer,

type: kImagePreprocess
rgbimage_conf {
  scale: float
  cropsize: int  # cropping each image to keep the central part with this size
  mirror: bool  # mirror the image by set image[i,j]=image[i,len-j]
  meanfile: "Image_Mean_File_Path"
ShardDataLayer (Deprected)

Deprected! Please use ProtoRecordInputLayer or CSVRecordInputLayer.

ShardDataLayer is a subclass of DataLayer, which reads Records from disk file. The file should be created using DataShard class. With the data file prepared, users configure the layer as

type: kShardData
sharddata_conf {
  path: "path to data shard folder"
  batchsize: int
  random_skip: int

batchsize specifies the number of records to be trained for one mini-batch. The first rand() % random_skip Records will be skipped at the first iteration. This is to enforce that different workers work on different Records.

LMDBDataLayer (Deprected)

Deprected! Please use ProtoRecordInputLayer or CSVRecordInputLayer.

[LMDBDataLayer] is similar to ShardDataLayer, except that the Records are loaded from LMDB.

type: kLMDBData
lmdbdata_conf {
  path: "path to LMDB folder"
  batchsize: int
  random_skip: int
ParserLayer (Deprected)

Deprected! Please use ProtoRecordInputLayer or CSVRecordInputLayer.

It get a vector of Records from DataLayer and parse features into a Blob.

virtual void ParseRecords(Phase phase, const vector<Record>& records, Blob<float>* blob) = 0;
LabelLayer (Deprected)

Deprected! Please use ProtoRecordInputLayer or CSVRecordInputLayer.

LabelLayer is a subclass of ParserLayer. It parses a single label from each Record. Consequently, it will put $b$ (mini-batch size) values into the Blob. It has no specific configuration fields.

MnistImageLayer (Deprected)

Deprected! Please use ProtoRecordInputLayer or CSVRecordInputLayer. [MnistImageLayer] is a subclass of ParserLayer. It parses the pixel values of each image from the MNIST dataset. The pixel values may be normalized as x/norm_a - norm_b. For example, if norm_a is set to 255 and norm_b is set to 0, then every pixel will be normalized into [0, 1].

type: kMnistImage
mnistimage_conf {
  norm_a: float
  norm_b: float
RGBImageLayer (Deprected)

Deprected! Please use the ImagePreprocessLayer. RGBImageLayer is a subclass of ParserLayer. It parses the RGB values of one image from each Record. It may also apply some transformations, e.g., cropping, mirroring operations. If the meanfile is specified, it should point to a path that contains one Record for the mean of each pixel over all training images.

type: kRGBImage
rgbimage_conf {
  scale: float
  cropsize: int  # cropping each image to keep the central part with this size
  mirror: bool  # mirror the image by set image[i,j]=image[i,len-j]
  meanfile: "Image_Mean_File_Path"

PrefetchLayer embeds other input layers to do data prefeching. It will launch a thread to call the embedded layers to load and extract features. It ensures that the I/O task and computation task can work simultaneously. One example PrefetchLayer configuration is,

layer {
  name: "prefetch"
  type: kPrefetch
  sublayers {
    name: "data"
    type: kShardData
    sharddata_conf { }
  sublayers {
    name: "rgb"
    type: kRGBImage
    rgbimage_conf { }
  sublayers {
    name: "label"
    type: kLabel
    srclayers: "data"

The layers on top of the PrefetchLayer should use the name of the embedded layers as their source layers. For example, the “rgb” and “label” should be configured to the srclayers of other layers.

Output Layers

Output layers get data from their source layers and write them to persistent storage, e.g., disk files or HDFS (to be supported).


This layer gets data (and label if it is available) from its source layer and converts it into records of type RecordProto. Records are written as (key = instance No., value = serialized record) tuples into Store, e.g., KVFile. The configuration of this layer should include the specifics of the Store backend via store_conf.

layer {
  name: "output"
  type: kRecordOutput
  store_conf {
    backend: "kvfile"

This layer gets data (and label if it available) from its source layer and converts it into a string per instance with fields separated by commas (i.e., CSV format). The shape information is not kept in the string. All strings are written into Store, e.g., text file. The configuration of this layer should include the specifics of the Store backend via store_conf.

layer {
  name: "output"
  type: kCSVOutput
  store_conf {
    backend: "textfile"

Neuron Layers

Neuron layers conduct feature transformations.


type: kActivation
activation_conf {

ConvolutionLayer conducts convolution transformation.

type: kConvolution
convolution_conf {
  num_filters: int
  kernel: int
  stride: int
  pad: int
param { } # weight/filter matrix
param { } # bias vector

The int value num_filters stands for the count of the applied filters; the int value kernel stands for the convolution kernel size (equal width and height); the int value stride stands for the distance between the successive filters; the int value pad pads each with a given int number of pixels border of zeros.


InnerProductLayer is fully connected with its (single) source layer. Typically, it has two parameter fields, one for weight matrix, and the other for bias vector. It rotates the feature of the source layer (by multiplying with weight matrix) and shifts it (by adding the bias vector).

type: kInnerProduct
innerproduct_conf {
  num_output: int
param { } # weight matrix
param { } # bias vector

PoolingLayer is used to do a normalization (or averaging or sampling) of the feature vectors from the source layer.

type: kPooling
pooling_conf {
  pool: AVE|MAX // Choose whether use the Average Pooling or Max Pooling
  kernel: int   // size of the kernel filter
  pad: int      // the padding size
  stride: int   // the step length of the filter

The pooling layer has two methods: Average Pooling and Max Pooling. Use the enum AVE and MAX to choose the method.

  • Max Pooling selects the max value for each filtering area as a point of the result feature blob.
  • Average Pooling averages all values for each filtering area at a point of the result feature blob.

ReLuLayer has rectified linear neurons, which conducts the following transformation, f(x) = Max(0, x). It has no specific configuration fields.


STanhLayer uses the scaled tanh as activation function, i.e., f(x)=1.7159047* tanh(0.6666667 * x). It has no specific configuration fields.


[SigmoidLayer] uses the sigmoid (or logistic) as activation function, i.e., f(x)=sigmoid(x). It has no specific configuration fields.

Dropout Layer

DropoutLayer is a layer that randomly dropouts some inputs. This scheme helps deep learning model away from over-fitting.

type: kDropout
dropout_conf {
  dropout_ratio: float # dropout probability

LRNLayer, (Local Response Normalization), normalizes over the channels.

type: kLRN
lrn_conf {
  local_size: int
  alpha: float  // scaling parameter
  beta: float   // exponential number

local_size specifies the quantity of the adjoining channels which will be summed up. For WITHIN_CHANNEL, it means the side length of the space region which will be summed up.

CuDNN layers

CuDNN v3 and v4 are supported in SINGA, which include the following layers,

  • CudnnActivationLayer (activation functions are SIGMOID, TANH, RELU)
  • CudnnConvLayer
  • CudnnLRNLayer
  • CudnnPoolLayer
  • CudnnSoftmaxLayer

These layers have the same configuration as the corresponding CPU layers. For CuDNN v4, the batch normalization layer is added, which is named as CudnnBMLayer.

Loss Layers

Loss layers measures the objective training loss.


SoftmaxLossLayer is a combination of the Softmax transformation and Cross-Entropy loss. It applies Softmax firstly to get a prediction probability for each output unit (neuron) and compute the cross-entropy against the ground truth. It is generally used as the final layer to generate labels for classification tasks.

type: kSoftmaxLoss
softmaxloss_conf {
  topk: int

The configuration field topk is for selecting the labels with topk probabilities as the prediction results. It is tedious for users to view the prediction probability of every label.


Subclasses of ConnectionLayer are utility layers that connects other layers due to neural net partitioning or other cases.


ConcateLayer connects more than one source layers to concatenate their feature blob along given dimension.

type: kConcate
concate_conf {
  concate_dim: int  // define the dimension

SliceLayer connects to more than one destination layers to slice its feature blob along given dimension.

type: kSlice
slice_conf {
  slice_dim: int

SplitLayer connects to more than one destination layers to replicate its feature blob.

type: kSplit
split_conf {
  num_splits: int
BridgeSrcLayer & BridgeDstLayer

BridgeSrcLayer & BridgeDstLayer are utility layers assisting data (e.g., feature or gradient) transferring due to neural net partitioning. These two layers are added implicitly. Users typically do not need to configure them in their neural net configuration.


It write the prediction results or the extracted features into file, HTTP stream or other places. Currently SINGA has not implemented any specific output layer.

Advanced user guide

The base Layer class is introduced in this section, followed by how to implement a new Layer subclass.

Base Layer class


LayerProto layer_conf_;
vector<Blob<float>> datavec_, gradvec_;
vector<AuxType> aux_data_;

The base layer class keeps the user configuration in layer_conf_. datavec_ stores the features associated with this layer. There are layers without feature vectors; instead, they share the data from source layers. The gradvec_ is for storing the gradients of the objective loss w.r.t. the datavec_. The aux_data_ stores the auxiliary data, e.g., image label (set AuxType to int). If images have variant number of labels, the AuxType can be defined to vector<int>. Currently, we hard code AuxType to int. It will be added as a template argument of Layer class later.

If a layer has parameters, these parameters are declared using type Param. Since some layers do not have parameters, we do not declare any Param in the base layer class.


virtual void Setup(const LayerProto& conf, const vector<Layer*>& srclayers);
virtual void ComputeFeature(int flag, const vector<Layer*>& srclayers) = 0;
virtual void ComputeGradient(int flag, const vector<Layer*>& srclayers) = 0;

The Setup function reads user configuration, i.e. conf, and information from source layers, e.g., mini-batch size, to set the shape of the data_ (and grad_) field as well as some other layer specific fields. Memory will not be allocated until computation over the data structure happens.

The ComputeFeature function evaluates the feature blob by transforming (e.g. convolution and pooling) features from the source layers. ComputeGradient computes the gradients of parameters associated with this layer. These two functions are invoked by the TrainOneBatch function during training. Hence, they should be consistent with the TrainOneBatch function. Particularly, for feed-forward and RNN models, they are trained using BP algorithm, which requires each layer’s ComputeFeature function to compute data_ based on source layers, and requires each layer’s ComputeGradient to compute gradients of parameters and source layers’ grad_. For energy models, e.g., RBM, they are trained by CD algorithm, which requires each layer’s ComputeFeature function to compute the feature vectors for the positive phase or negative phase depending on the phase argument, and requires the ComputeGradient function to only compute parameter gradients. For some layers, e.g., loss layer or output layer, they can put the loss or prediction result into the metric argument, which will be averaged and displayed periodically.

Implementing a new Layer subclass

Users can extend the Layer class or other subclasses to implement their own feature transformation logics as long as the two virtual functions are overridden to be consistent with the TrainOneBatch function. The Setup function may also be overridden to read specific layer configuration.

The RNNLM provides a couple of user-defined layers. You can refer to them as examples.

Layer specific protocol message

To implement a new layer, the first step is to define the layer specific configuration. Suppose the new layer is FooLayer, the layer specific google protocol message FooLayerProto should be defined as

# in user.proto
package singa
import "job.proto"
message FooLayerProto {
  optional int32 a = 1;  // specific fields to the FooLayer

In addition, users need to extend the original LayerProto (defined in job.proto of SINGA) to include the foo_conf as follows.

extend LayerProto {
  optional FooLayerProto foo_conf = 101;  // unique field id, reserved for extensions

If there are multiple new layers, then each layer that has specific configurations would have a <type>_conf field and takes one unique extension number. SINGA has reserved enough extension numbers, e.g., starting from 101 to 1000.

# job.proto of SINGA
LayerProto {
  extensions 101 to 1000;

With user.proto defined, users can use protoc to generate the and user.pb.h files. In users’ code, the extension fields can be accessed via,

auto conf = layer_proto_.GetExtension(foo_conf);
int a = conf.a();

When defining configurations of the new layer (in job.conf), users should use user_type for its layer type instead of type. In addition, foo_conf should be enclosed in brackets.

layer {
  name: "foo"
  user_type: "kFooLayer"  # Note user_type of user-defined layers is string
  [foo_conf] {      # Note there is a pair of [] for extension fields
    a: 10

New Layer subclass declaration

The new layer subclass can be implemented like the built-in layer subclasses.

class FooLayer : public singa::Layer {
  void Setup(const LayerProto& conf, const vector<Layer*>& srclayers) override;
  void ComputeFeature(int flag, const vector<Layer*>& srclayers) override;
  void ComputeGradient(int flag, const vector<Layer*>& srclayers) override;

  //  members

Users must override the two virtual functions to be called by the TrainOneBatch for either BP or CD algorithm. Typically, the Setup function will also be overridden to initialize some members. The user configured fields can be accessed through layer_conf_ as shown in the above paragraphs.

New Layer subclass registration

The newly defined layer should be registered in by adding

driver.RegisterLayer<FooLayer, std::string>("kFooLayer"); // "kFooLayer" should be matched to layer configurations in job.conf.

After that, the NeuralNet can create instances of the new Layer subclass.