Layers Library Reference

Note: This documentation has not yet been completely updated with respect to the latest update of the Layers library. It should be correct but misses several new options and layer types.

CNTK predefines a number of common “layers,” which makes it very easy to write simple networks that consist of standard layers layered on top of each other. Layers are function objects that can be used like a regular Function but hold learnable parameters and have an additional pair of () to pass construction parameters or attributes.

For example, this is the network description for a simple 1-hidden layer model using the Dense() layer:

h = Dense(1024, activation=relu)(features)
p = Dense(9000, activation=softmax)(h)

which can then, e.g., be used for training against a cross-entropy criterion:

ce = cross_entropy(p, labels)

If your network is a straight concatenation of operations (many are), you can use the alternative Sequential() notation:

from cntk.layers import *
my_model = Sequential ([
    Dense(1024, activation=relu),
    Dense(9000, activation=softmax)

and invoke it like this:

p = my_model(features)

Built on top of Sequential() is For(), which allows to easily create models with repetitions. For example, a 2011-style feed-forward speech-recognition network with 6 hidden sigmoid layers of identical dimensions can be written like this:

my_model = Sequential ([
    For(range(6), lambda: \
        Dense(2048, activation=sigmoid))
    Dense(9000, activation=softmax)

Note that for most real-life inference scenarios, the output layer’s softmax non-linearity is not needed (it is instead made part of the training criterion).

General patterns

Specifying the same options to multiple layers

Often, many layers share options. For example, typical image-recognition systems use the relu activation function throughout. You can use the Python with statement with the CNTK default_options() function to define scopes with locally changed defaults, using one of the following two forms:

with default_options(OPT1=VAL1, OPT2=VAL2, ...):
    # scope with modified defaults

with default_options_for(FUNCTION, OPT1=VAL1, OPT2=VAL2, ...):
    # scope with modified defaults for FUNCTION only

The following options can be overridden with the with statement:

  • init (default: glorot_uniform()): initializer specification, for Dense(), Convolution(), and Embedding()
  • activation (default: None): activation function, for Dense() and Convolution()
  • bias (default: True): have a bias, for Dense() and Convolution()
  • init_bias (default: 0): initializer specification for the bias, for Dense() and Convolution()
  • initial_state (default: None): initial state to use in Recurrence() Recurrence()
  • use_peepholes (default: False): use peephole connections in LSTM() LSTM(), GRU(), RNNStep()

The second form allows to set default options on a per-layer type. This is, for example, valuable for the pad parameter which enables padding in convolution and pooling, but is not always set to the same for these two layer types.

Weight sharing

If you assign a layer to a variable and use it in multiple places, the weight parameters will be shared. If you say

lay = Dense(1024, activation=sigmoid)
h1 = lay(x)
h2 = lay(h1)  # same weights as h1

h1 and h2 will share the same weight parameters, as lay() is the same function in both cases. In the above case this is probably not what was desired, so be aware. If both invocations of lay() above are meant to have different parameters, remember to define two separate instances, for example lay1 = Dense(...) and lay2 = Dense(...).

So why this behavior? Layers allow to share parameters across sections of a model. Consider a DSSM model which processes two input images, say doc and query identically with the same processing chain, and compares the resulting hidden vectors:

with default_options(activation=relu):
    image_to_vec = Sequential([
        Convolution((5,5), 32, pad=True), MaxPooling((3,3), strides=2),
        Convolution((5,5), 64, pad=True), MaxPooling((3,3), strides=2),
        Dense(10, activation=None)
z_doc   = image_to_vec (doc)
z_query = image_to_vec (query)  # same model as for z_doc
sim = cos_distance(zdoc, z_query)

where image_to_vec is the part of the model that converts images into flat vector. image_to_vec is a function object that in turn contains several function objects (e.g. three instances of Convolution()). image_to_vec is instantiated once, and this instance holds the learnable parameters of all the included function objects. Both invocations of model() will share these parameters in application, and their gradients will be the sum of both invocations.

Lastly, note that if in the above example query and doc must have the same dimensions, since they are processed through the same function object, and that function object’s first layer has its input dimension inferred to match that of both query and doc. If their dimensions differ, then this network is malformed, and dimension inference/validation will fail with an error message.

Example models

The following shows a slot tagger that embeds a word sequence, processes it with a recurrent LSTM, and then classifies each word:

from cntk.layers import *
tagging_model = Sequential ([
    Embedding(150),         # embed into a 150-dimensional vector
    Recurrence(LSTM(300)),  # forward LSTM
    Dense(labelDim)         # word-wise classification

And the following is a simple convolutional network for image recognition, using the with default_options(...): Specifying the same options to multiple layers pattern):

with default_options(activation=relu):
    conv_net = Sequential ([
        # 3 layers of convolution and dimension reduction by pooling
        Convolution((5,5), 32, pad=True), MaxPooling((3,3), strides=2),
        Convolution((5,5), 32, pad=True), MaxPooling((3,3), strides=2),
        Convolution((5,5), 64, pad=True), MaxPooling((3,3), strides=2),
        # 2 dense layers for classification
        Dense(10, activation=None)


Many layers are wrappers around underlying CNTK primitives, along with the respective required learnable parameters. For example, `Convolution() Convolution() wraps the convolution() primitive. The benefits of using layers are: * layers contain learnable parameters of the correct dimension * layers are composable (cf. `Sequential() Sequential())

However, since the layers themselves are implemented in Python using the same CNTK primitives that are available to the user, if you find that a layer you need is not available, you can always write it yourself or write the formula directly as a CNTK expression.

The Python library described here is the equivalent of BrainScript’s Layers Library.


Factory function to create a fully-connected layer. Dense() takes an optional activation function.

Dense(shape, activation=default_override_or(identity), init=default_override_or(glorot_uniform()),
      input_rank=None, map_rank=None,
      bias=default_override_or(True), init_bias=default_override_or(0),


  • shape: output dimension of this layer
  • activation (default: None: pass a function here to be used as the activation function, such as activation=relu
  • input_rank: if given, number of trailing dimensions that are transformed by Dense() (map_rank must not be given)
  • map_rank: if given, the number of leading dimensions that are not transformed by Dense() (input_rank must not be given)
  • init (default: glorot_uniform()): initializer descriptor for the weights. See cntk.initializer for a full list of random-initialization options.
  • bias: if False, do not include a bias parameter
  • init_bias (default: 0): initializer for the bias

Return Value

A function that implements the desired fully-connected layer. See description.


Use these factory functions to create a fully-connected layer. It creates a function object that contains a learnable weight matrix and, unless bias=False, a learnable bias. The function object can be used like a function, which implements one of these formulas (using Python 3.5 @ operator for matrix multiplication):

Dense(...)(v) = activation (v @ W + b)
Dense(...)(v) = v @ W + b      # if activation is None

where W is a weight matrix of dimension ((dimension of v), shape), b is the bias of dimension (outdim,), and the resulting value has dimension (or tensor dimensions) as given by shape.

Tensor support

If the returned function is applied to an input of a tensor rank > 1, e.g. a 2D image, W will have the dimension (..., (second dimension of input), (first dimension of input), shape).

On the other hand, shape can be a vector that specifies tensor dimensions, for example (10,10). In that case, W will have the dimension ((dimension of input), ..., shape[1], shape[0]), and b will have the tensor dimensions (..., shape[1], shape[0]).

CNTK’s matrix product will interpret these extra output or input dimensions as if they were flattened into a long vector. For more details on this, see the documentation of Times().

The options input_rank and map_rank, which are mutually exclusive, can specify that not all of the input axes of a tensor should be transformed. map_rank specifies how many leading axes are kept as dimensions in the result; those axes are not part of the projection, but rather each element along these axes is transformed independently (aka mapped). input_rank is an alternative that instead specifies the how many trailing axes are to be transformed (the remaining are mapped).


h = Dense(1024, activation=sigmoid)(v)

or alternatively:

layer = Dense(1024, activation=sigmoid)
h = layer(v)


Creates a convolution layer with optional non-linearity.

Convolution(filter_shape,     # shape of receptive field, e.g. (3,3)
            num_filters=None, # e.g. 64 or None (which means 1 channel and don't add a dimension)
            sequential=False, # time convolution if True (filter_shape[0] corresponds to dynamic axis)
            reduction_rank=1, # (0 means input has no depth dimension, e.g. audio signal or B&W image)


  • filter_shape: shape of receptive field of the filter, e.g. (5,5) for a 2D filter (not including the input feature-map depth)
  • num_filters: number of output channels (number of filters)
  • activation: optional non-linearity, e.g. activation=relu
  • init: initializer descriptor for the weights, e.g. glorot_uniform(). See cntk.initializer for a full list of random-initialization options.
  • pad: if False (default), then the filter will be shifted over the “valid” area of input, that is, no value outside the area is used. If pad is True on the other hand, the filter will be applied to all input positions, and values outside the valid region will be considered zero.
  • strides: increment when sliding the filter over the input. E.g. (2,2) to reduce the dimensions by 2
  • bias: if False, do not include a bias parameter
  • init_bias: initializer for the bias
  • use_correlation: currently always True and cannot be changed. It indicates that Convolution() actually computes the cross-correlation rather than the true convolution

Return Value

A function that implements the desired convolution operation.


Use these factory functions to create a convolution layer.

The resulting layer applies a convolution operation on N-dimensional feature maps. The caller specifies the receptive field of the filter and the number of filters (output feature maps).

A set of filters for a given receptive field (e.g. (5,5)) is correlated with every location of the input (e.g. a (480, 640)-sized image). Assuming padding is enabled (pad) and strides are 1, this will generate an output of the same dimension ((480, 640)).

Typically, many filters are applied at the same time, to create “per-pixel activation vectors”. num_filters specifies the number: For every input location, an entire vector of num_filters is produced. For our example above, setting num_filters to 64 would in a (64, 480, 640)-sized tensor. That first axis is also called the channel dimension or the feature-map axis.

When convolution is applied to an input with a channel dimension, each filter will also consist of vectors of the input’s channel dimension. E.g. when applying convolution with a specified receptive field of (5,5) to a (3, 480, 640)-sized color image, each filter will be a (3, 5, 5)] tensor.

All num_filters filters are stacked together into the so-called convolution kernel, which is a parameter tensor owned by and held inside this layer. In our example, the kernel shape will be (64, 3, 5, 5).

The following summarizes the relationship between the various dimensions and shapes:

input shape   : (               num_input_channels, (spatial dims) )
filter shape  : (                                   (filter_shape) )
output shape  : ( num_filters,                      (spatial dims) )
kernel shape  : ( num_filters,  num_input_channels, (filter_shape)     )

which in our example are:

input shape   : (              3, 480, 640 )
filter shape  : (                   5, 5   )
output shape  : ( num_filters,    480, 640 )
kernel shape  : ( num_filters, 3,   5, 5   )


If padding is not enabled (pad not given or False for all dimensions), then the output size will be reduced by stripping the boundary locations to which the full filter extent cannot be applied. E.g. applying a (5,5)-extent filter to an image without padding, the outermost 2 rows and columns of pixels would cause the filter to be applied out of bounds. Hence, Convolution() will reduce the dimensions accordingly: An (480, 640) image convolved with a (5,5) filter without padding will leave a (476, 636)-sized output.


The strides parameters specify the increment of filters. Stride values greater than one will lead to a sub-sampling of the output region. E.g. filtering a (480, 640) image with a stride of (2,2) will result in a (240, 320)-sized region with padding, and (238, 318) without padding.


This layer is a wrapper around the convolution() primitive.

The filter kernel parameters’ name as shown in the log’s validation section will end in .W.

Atrous convolution is at present not supported but planned for the near future.


c = Convolution((3,3), 64, pad=True, strides=(1,1), bias=False)(x)

MaxPooling(), AveragePooling()

Factory functions to create a max- or average-pooling layer.

MaxPooling(filter_shape,      # shape of receptive field, e.g. (3,3)
AveragePooling(filter_shape,  # shape of receptive field, e.g. (3,3)


  • filter_shape: receptive field (window) to pool over, e.g. (2,2) (not including the input feature-map depth)
  • strides: increment when sliding the pool over the input. E.g. (2,2) to reduce the dimensions by 2
  • pad: if False (default), then the pool will be shifted over the “valid” area of input, that is, no value outside the area is used. If pad is True on the other hand, the pool will be applied to all input positions, and values outside the valid region will be considered zero. For average pooling, count for average does not include padded values.

Return Value

A function that implements the desired pooling layer.


Use this factory function to create a pooling operation. Use MaxPooling() to compute the maximum over the values in the sliding pooling window, and AveragePooling() to take their average.

The pooling operation slides a receptive field, or pooling window, over the input, and computes either the maximum or the average of the values in the respective window. In case of average with pad being True, the padding regions are not included in the average.

This operation is structurally very similar to convolution, except that the operation applied to the sliding window is of a different nature.

All considerations regarding input dimensions, padding, and strides apply, so please see Convolution() for more detail.


p = MaxPooling((3,3), strides=(2,2))(c)

GlobalMaxPooling(), GlobalAveragePooling()

Factory functions to create a global-max-pooling or global-average-pooling layer.


Return Value

A function that implements the desired pooling layer.


Use this factory function to create a global pooling operation. Use GlobalMaxPooling() to compute the maximum over all spatial data, or GlobalAveragePooling() to take their average.

The global pooling operation infer the pooling window shape from the input tensor and create a pooling function with pooling window size that matches the input spatial dimension. It then computes either the maximum or the average of all the values inside the inferred pooling window.


p = GlobalMaxPooling()(c)


Factory functions to create a dropout layer.

Dropout(dropout_rate=None, keep_prob=None, name='')


  • dropout_rate: a fraction between [0, 1) that specifies the probability by which the dropout operation will randomly set elements of the input to zero. 0 mean select everything and close to 1 mean drop every element.

Return Value

A function that implements the desired dropout layer.


Use this factory function to create a dropout operation with a specific dropout rate.


p = Dropout(0.5)(c)


Factory function to create a linear embedding layer, which is either learned or a constant passed from outside.

Embedding(shape=None, init=default_override_or(glorot_uniform()), weights=None, name='')


  • shape: the dimension of the desired embedding vector. Must not be None unless weights are passed
  • init: initializer descriptor for the weights to be learned. See cntk.initializer for a full list of initialization options.
  • weights (numpy array): if given, embeddings are not learned but specified by this array (which could be, e.g., loaded from a file) and not updated further during training

Return Value

A function that implements the embedding layer. See description.


“Embedding” refers to representing words or other discrete items by dense continuous vectors. This layer assumes that the input is in one-hot form. E.g., for a vocabulary size of 10,000, each input vector is expected to have dimension 10,000 and consist of zeroes except for one position that contains a 1. The index of that location is the index of the word or item it represents.

In CNTK, the corresponding embedding vectors are stored as rows of a matrix. Hence, mapping an input word to its embedding is implemented as a matrix product. For this to be very efficient, it is important that the input vectors are stored in sparse format (specify is_sparse=True in the corresponding Input()).

Fun fact: The gradient of an embedding matrix has the form of gradient vectors that are only non-zero for words seen in a minibatch. Since for realistic vocabularies of tens or hundreds of thousands, the vast majority of rows would be zero, CNTK implements a specific optimization to represent the gradient in “row-sparse” form.

Known issue: The above-mentioned row-sparse gradient form is currently not supported by our 1-bit SGD parallelization technique. Please use the block-momentum technique instead.


A learned embedding that represents words from a vocabulary of 87636 as a 300-dimensional vector:

input = Input(87636, is_sparse=True)  # word sequence, as one-hot vector, sparse format
embEn = Embedding(300)(input)         # embed word as a 300-dimensional continuous vector

In addition to is_sparse=True, one would also typically read sparse data from disk. Here is an example of reading sparse text input with the CNTKTextFormatReader:

source = MinibatchSource(CTFDeserializer('en2fr.ctf', StreamDefs(
    input   = StreamDef(field='E', shape=87636, is_sparse=True),
    labels  = StreamDef(field='F', shape=98624, is_sparse=True)

If, instead, the embedding vectors already exist and should be loaded from a file, it would look like this:

input = Input(87636, is_sparse=True)   # word sequence, as one-hot vector, sparse format
embEn = Embedding(300, weights=np.load_txt('embedding-en.txt'))(w) # embedding from disk

where the file 'embedding-en.txt' is the name of a file that would be expected to consist of 87,636 text rows, each of which consisting of 300 space-separated numbers.


Factory function to create a single-layer or multi-layer recurrence.

Recurrence(step_function, go_backwards=default_override_or(False), initial_state=default_override_or(0), return_full_state=False, name='')
RecurrenceFrom(step_function, go_backwards=default_override_or(False), return_full_state=False, name='')
Fold(folder_function, go_backwards=default_override_or(False), initial_state=default_override_or(0), return_full_state=False, name='')
UnfoldFrom(generator_function, until_predicate=None, length_increase=1, name='')


  • step_function: the Function to recur over, for example LSTM()
  • go_backwards (optional): if True, the recurrence is run backwards
  • initial_state (optional, default 0): initial value of the hidden variable that is recurred over. Currently, initial_state cannot have a dynamic axis.

Return Value

Recurrence() creates a function that implements the desired layer that recurrently applies a model, such as an LSTM, to an input sequence. This layer maps an input sequence to a sequence of hidden states of the same length.


This implements the recurrence to be applied to an input sequence along a dynamic axis. This operation automatically handles batches of variable-length input sequences. The initial value(s) of the hidden state variable(s) are 0 unless specified by initial_state. A recurrence layer’s operation can be best described by pseudo-code (but note that the real implementation is more complicated since it handles automatic minibatching even if not all sequences are of the same length):

# pseudo-code for y = Recurrence(step_function)(x)
#  x: input sequence of tensors along the dynamic axis
#  y: resulting sequence of outputs along the same dynamic axis
y = []              # result sequence goes here
s = initial_state   # s = output of previous step ("state")
for x_n in x:       # pseudo-code for looping over all steps of input sequence along its dynamic axis
    s = step_function(s, x_n)  # pass previous state and new data to step_function -> new state

The step_function must be a CNTK Function that takes the previous state and a new input, and outputs a new state. State may consist of multiple variables (e.g. h and c in the case of the LSTM).

Applying this layer to an input sequence will return the sequence of the hidden states of the Function to recur over (in case of an LSTM, the LSTM’s memory cell’s value is not returned). The returned sequence has the same length as the input. If only the last state is desired, as in sequence-classification or some sequence-to-sequence scenarios, use Fold() instead of Recurrence().

Any function with such a signature can be used. For example, Recurrence(plus, initial_value=0) is a layer that computes a cumulative sum over the input data, while Fold(element_max) is a layer that performs max-pooling over a sequence.

To create a bidirectional model with Recurrence(), use two layers, one with go_backwards=True, and splice() the two outputs together.

initial_state may have a dynamic batch axis. In that case, the preferred pattern is RecurrentFrom(), which creates a function that takes the initial state as its first argument(s), followed by the inputs.


A simple text classifier, which runs a word sequence through a recurrence and then passes the last hidden state of the LSTM to a softmax classifer, could have this form:

w = Input(...)                          # word sequence (one-hot vectors)
e = Embedding(150)(w)                   # embed as a 150-dimensional dense vector
h = Recurrence(LSTM(300))(e)            # left-to-right LSTM with hidden and cell dim 300
t = select_last(h)                      # extract last hidden state
z = Dense(10000, activation=softmax)(t) # softmax classifier

To create a bidirectional one-layer LSTM (e.g. using half the hidden dimension compared to above), use this:

h_fwd = Recurrence(LSTM(150))(e)
h_bwd = Recurrence(LSTM(150), go_backwards=True)(e)
h = splice (h_fwd, h_bwd)

LSTM(), GRU(), RNNStep()

Factory functions to create a stateless LSTM/GRU/RNN Function, typically for use with Recurrence().

LSTM(shape, cell_shape=None, activation=default_override_or(tanh), use_peepholes=default_override_or(False),
     init=default_override_or(glorot_uniform()), init_bias=default_override_or(0),
GRU(shape, cell_shape=None, activation=default_override_or(tanh),
    init=default_override_or(glorot_uniform()), init_bias=default_override_or(0),
RNNStep(shape, cell_shape=None, activation=default_override_or(sigmoid),
        init=default_override_or(glorot_uniform()), init_bias=default_override_or(0),


  • shape: dimension of the output
  • cell_shape (optional): the dimension of the LSTM’s cell. If None, the cell shape is identical to shape. If specified, an additional linear projection will be inserted to project from the cell dimension to the output shape.
  • use_peepholes (optional): if True, then use peephole connections in the LSTM
  • init: initializer descriptor for the weights. See cntk.initializer for a full list of initialization options.
  • enable_self_stabilization (optional): if True, insert a Stabilizer() for the hidden state and cell

Return Value

A Function that implements stateless Long-Short-Term-Memory, Gated Recurrent Unit, or plain recurrent unit, typically for use with the Recurrence() family of higher-order layers.


This creates a Function object that implements the LSTM, GRU, or a RNN block. It returns its current state, and takes the previous state as an additional input. The function is stateless; i.e., it is not a recurrent LSTM layer. Use Recurrence() to turn this into a recurrent layer that is applied along a dynamic axis.


See Recurrence().


Factory function to create a layer that delays its input.

Delay(T=1, initial_state=default_override_or(0), name='')


  • T: the number of time steps to delay. To access future values, use a negative value
  • initial_state (optiona, default=0): value to use for the delayed frames at the boundaries

Return Value

A function that implements the desired delay operation.


This operation delays an input sequence by T steps (default 1). This useful, for example, to turn a word sequence into a sequence of overlapping word triples.

Consider an input sequence “a b c b”, which shall be encoded as a sequence of 3-dimensional one-hot vectors as follows:

1 0 0
0 1 0
0 0 1
0 1 0

Here, every row is a one-hot vector and corresponds to a word. Applying Delay(T=1) to this input will generate this sequence:

0 0 0
1 0 0
0 1 0
0 0 1

All tokens get delayed by one, and the first position gets filled in by initial_state which defaults to 0. Likewise, using Delay(T=-1) (negative delay) will give access to the future values, and pad from the end with a zero:

0 1 0
0 0 1
0 1 0
0 0 0


This layer is a wrapper around the sequence.past_value() and sequence.future_value() primitives.


The following shows how to stack three neighbor words into a trigram vector:

x  = ...                   # input value, e.g. a N-dimensional one-hot vector
xp = Delay()(x)            # previous value
xn = Delay(T=-1)(x)        # next value (negative delay)
tg = splice (xp, x, xn)    # concatenate all into a 3N-dimensional three-hot vector

BatchNormalization(), LayerNormalization(), Stabilizer()

Factory functions to create layers for batch normalization, layer normalization, and self-stabilization.

BatchNormalization(map_rank=default_override_or(None),  # if given then normalize only over this many dimensions. E.g. pass 1 to tie all (h,w) in a (C, H, W)-shaped input
                   normalization_time_constant=default_override_or(5000), blend_time_constant=0,
                   epsilon=default_override_or(0.00001), use_cntk_engine=default_override_or(False),
LayerNormalization(initial_scale=1, initial_bias=0, epsilon=default_override_or(0.00001), name='')
Stabilizer(steepness=4, enable_self_stabilization=default_override_or(True), name='')



  • map_rank: if given then normalize only over this many leading dimensions. E.g. 1 to tie all (h,w) in a (C, H, W)-shaped input. Currently, the only allowed values are None (no pooling) and 1 (e.g. pooling across all pixel positions of an image)
  • normalization_time_constant (default 5000): time constant in samples of the first-order low-pass filter that is used to compute mean/variance statistics for use in inference
  • initial_scale: initial value of scale parameter
  • epsilon: small value that gets added to the variance estimate when computing the inverse
  • use_cntk_engine: if True, use CNTK’s native implementation. If false, use cuDNN’s implementation (GPU only).


  • initial_scale: initial value of scale parameter
  • initial_bias: initial value of bias parameter


  • steepness: sharpness of the knee of the softplus function

Return Value

A function that implements a layer that performs the normalization operation.


BatchNormalization() implements the technique described in paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift (Sergey Ioffe, Christian Szegedy). It normalizes its inputs for every minibatch by the minibatch mean/variance, and de-normalizes it with a learned scaling factor and bias.

In inference, instead of using minibatch mean/variance, batch normalization uses a long-term running mean/var estimate. This estimate is computed during training by low-pass filtering minibatch statistics. The time constant of the low-pass filter can be modified by the normalization_time_constant parameter. We recommend to start with the default of (5000), but experiment with other values, typically on the order of several thousand to tens of thousand.

Batch normalization currently requires a GPU for training.

LayerNormalization() implements Layer Normalization (Jimmy Lei Ba, Jamie Ryan Kiros, Geoffrey E. Hinton). It normalizes each input sample by subtracting the mean across all elements of the sample, and then dividing by the standard deviation over all elements of the sample.

Stabilizer() implements a self-stabilizer per Self-stabilized deep neural network (P. Ghahremani, J. Droppo). This simple but effective technique multiplies its input with a learnable scalar (but unlike layer normalization, it does not first normalize the input, nor does it subtract a mean). Note that compared to the original paper, which proposes a linear scalar beta or an exponential one Exp (beta), we found it beneficial to use a sharpened softplus operation per the second author’s suggestion, which avoids both negative values and instability from the exponential.


BatchNormalization() is a wrapper around the batch_normalization() primitive. LayerNormalization() and Stabilizer() are expressed directly in Python as a CNTK expression.


A typical layer in a convolutional network with batch normalization:

def my_convo_layer(x, depth, init):
    c = Convolution(depth, (5,5), pad=True, init=init)(x)
    b = BatchNormalization(map_rank=1)(c)
    r = relu(b)
    p = MaxPooling((3,3), strides=(2,2))(r)
    return p


Composes an list of functions into a new function that calls these functions one after another (“forward function composition”).

Sequential(layers, name='')


layers: a list of functions which may be layer instances or single-argument primitives, e.g. [ LinearLayer(1024), sigmoid ]

Return value

This function returns another Function. That returned function takes one argument, and returns the result of applying all given functions in sequence to the input.


Sequential() is a powerful operation that allows to compactly express a very common situation in neural networks where an input is processed by propagating it through a progression of layers. You may be familiar with it from other neural-network toolkits.

Sequential() takes an array of functions as its argument, and returns a new function that invokes these function in order, each time passing the output of one to the next. Consider this example:

FGH = Sequential ([F, G, H])
y = FGH (x)

The FGH function defined above means the same as

y = H(G(F(x)))

This is known as function composition, and is especially convenient for expressing neural networks, which often have this form:

     +-------+   +-------+   +-------+
x -->|   F   |-->|   G   |-->|   H   |--> y
     +-------+   +-------+   +-------+

which is perfectly expressed by Sequential ([F, G, H]). (An even shorter alternative way of writing it is (F >> G >> H).)

Lastly, please be aware that the following expression:

layer1 = Dense(1024)
layer2 = Dense(1024)
z = Sequential([layer1, layer2])(x)

means something different from:

layer = Dense(1024)
z = Sequential([layer, layer])(x)

In the latter form, the same function with the same shared set of parameters is applied twice (typically not desired), while in the former, the two layers have separate sets of parameters.


Standard 4-hidden layer feed-forward network as used in the earlier deep-neural network work on speech recognition:

my_model = Sequential ([
    Dense(2048, activation=sigmoid),  # four hidden layers
    Dense(2048, activation=sigmoid),
    Dense(2048, activation=sigmoid),
    Dense(2048, activation=sigmoid),
    Dense(9000, activation=softmax)   # note: last layer is a softmax
features = Input(40)
p = my_model(features)


Repeats a layer multiple times.

For(rng, constructor, name='')


  • N: number of repetitions
  • constructor: a lambda with 0 or 1 argument that creates the layer

Return value

This function returns another Function. That returned function takes one argument, and returns the result of applying the repeated layers to the input, where each layer is a separate instance with a distinct set of model parameters.


For() creates a sequential model by repeatedly executing a constructor lambda passed to it; that is, you pass a Python function that creates a layer, e.g. using the Python lambda syntax.

For example, creating a stack of 3 Dense layers of identical shape:

     +------------+   +------------+   +------------+
x -->| Dense(128) |-->| Dense(128) |-->| Dense(128) |--> y
     +------------+   +------------+   +------------+

is as easy as:

model = For(range(3), lambda: Dense(128))

Note that because you pass in a lambda for creating the layer, each layer will be separately constructed. This is important, because this ensures that all layers have their own distinct set of model parameters.

That constructor lambda can optionally take one parameter, the layer counter. E.g. if the output dimension should double in each layer,

     +------------+   +------------+   +------------+
x -->| Dense(128) |-->| Dense(256) |-->| Dense(512) |--> y
     +------------+   +------------+   +------------+

the one-parameter lambda form allows you to say this (notice the lambda i, which defines a function that takes one parameter named i):

model = For(range(3), lambda i: Dense(128 * 2**i))

or this:

dims = [128,256,512]
model = For(range(3), lambda i: Dense(dims[i]))


The following creates a 9-hidden-layer VGG-style model. VGG is a popular architecture for image recognition:

with default_options(activation=relu):
    model = Sequential([
        For(range(3), lambda i: [  # lambda with one parameter
            Convolution((3,3), [64,96,128][i], pad=True),  # depth depends on i
            Convolution((3,3), [64,96,128][i], pad=True),
            MaxPooling((3,3), strides=(2,2))
        For(range(2), lambda : [   # lambda without parameter
        Dense(num_classes, activation=None)

The resulting model will have this structure (read this from top to bottom)

input: image
output: object