Network Structure


The network structure which defines the model topology is defined by the config network option, which is a dict, where each entry is a layer specification, which itself is a dict containing the kwargs for the specific layer class. E.g.:

network = {
    "fw1": {"class": "linear", "activation": "relu", "dropout": 0.1, "n_out": 500, "from": "data"},
    "fw2": {"class": "linear", "activation": "relu", "dropout": 0.1, "n_out": 500, "from": "fw1"},
    "output": {"class": "softmax", "loss": "ce", "from": ["fw2"], "target": "classes"}

The "class" key will get extracted from the layer arguments and the specific layer class will be used. Some arguments are available for all layer classes, such as dropout. A list of all general arguments can be found below in Defining Layers. For the layer specific arguments such as activation``for the linear layer please have a look at the :ref:`layer_reference`. The ``from argument, which is also available for all layers, is a list of all input layers or datasets. "data" denotes the default data input. More details on how to connect layers and datasets can be found below at Connecting Layers.

For Theano, the base layer class is returnn.theano.layers.base.Container and returnn.theano.layers.base.Layer; for TensorFlow, it is E.g. that would use the class, and the LinearLayer.__init__ will accepts arguments like activation. In the given example, all the remaining arguments will get handled by the base layer.

The construction itself can be found for TensorFlow in, which starts from the output layers goes over the sources of a layer, which are defined by "from". If a layer does not define "from", it will automatically get the input from the dataset data.

Here is a 2 layer unidirectional LSTM network:

network = {
    "lstm1": {"class": "rec", "unit": "lstm", "dropout": 0.1, "n_out": 500, "from": "data"},
    "lstm2": {"class": "rec", "unit": "lstm", "dropout": 0.1, "n_out": 500, "from": "lstm1"},
    "output": {"class": "softmax", "loss": "ce", "from": "lstm2", "target": "classes"}

In TensorFlow, that would use the layer class which will handle the argument unit.

And here is a 3 layer bidirectional LSTM network:

network = {
"lstm0_fw" : { "class": "rec", "unit": "lstm", "n_out" : 500, "dropout": 0.1, "L2": 0.01, "direction": 1, "from": "data" },
"lstm0_bw" : { "class": "rec", "unit": "lstm", "n_out" : 500, "dropout": 0.1, "L2": 0.01, "direction": -1, "from": "data" },

"lstm1_fw" : { "class": "rec", "unit": "lstm", "n_out" : 500, "dropout": 0.1, "L2": 0.01, "direction": 1, "from" : ["lstm0_fw", "lstm0_bw"] },
"lstm1_bw" : { "class": "rec", "unit": "lstm", "n_out" : 500, "dropout": 0.1, "L2": 0.01, "direction": -1, "from" : ["lstm0_fw", "lstm0_bw"] },

"lstm2_fw" : { "class": "rec", "unit": "lstm", "n_out" : 500, "dropout": 0.1, "L2": 0.01, "direction": 1, "from" : ["lstm1_fw", "lstm1_bw"] },
"lstm2_bw" : { "class": "rec", "unit": "lstm", "n_out" : 500, "dropout": 0.1, "L2": 0.01, "direction": -1, "from" : ["lstm1_fw", "lstm1_bw"] },

"output" :   { "class" : "softmax", "loss" : "ce", "from" : ["lstm2_fw", "lstm2_bw"] }

Defining Layers

Every usable layer with the TensorFlow backend inherits from This class provides most of the parameters that can be set for each layer.

Every layer accepts the following dictionary entries:

class [str] specifies the type of the layer. Each layer class defines a layer_class attribute which defines the layer name.

from [list[str]] specifies the inputs of a layer, usually refering to the layer name. Many layers automatically concatenate their inputs, as provided by TFNetworkLayer._ConcatInputLayer. For more details on how to connect layers, see Connecting Layers.

n_out [int] specifies the output feature dimension, but the argument is usually not strictly required, except if there is some transformation like for Otherwise the output dimension is predefined (determined by If an explicit output feature dimension is required (like for and if n_out is not specified or set to None, it will try to determine the output size by a provided target. If a loss is given, it will set n_out to the value provided by See out_type for a more generic parameter.

out_type [dict[str]] specifies the output shape in more details (i.e. a more generic version than n_out). The keys are dim and shape and others from Usually it is automatically derived via

loss [str] every layer can have its output connected to a loss function. For available loss functions, see Loss Functions. When specifying a loss, also target has to be set (see below). In addition, loss_scale (defaults to 1) and loss_opts can be specified.

target [str] specifies the loss target in the dataset. If the target is not part of extern_data, but another layer in the network, add ‘layer:’ as prefix.

loss_scale [float] specifies a loss scale. Before adding all losses, this factor will be used as scaling.

loss_opts [dict] specifies additional loss arguments. For details, see the documentation of the loss functions Loss Functions

loss_only_on_non_search [bool] specifies that the loss should not be calculated during search.

trainable [bool] (default True) if set to False, the layer parameters will not be updated during training (parameter freezing).

L2 [float] if specified, add the L2 norm of the parameters with the given factor to the total constraints.

darc1 [float] if specified, add darc1 loss of the parameters with the given factor to the total constraints.

dropout [float] if specified, applies dropout in the input of the layer.

dropout_noise_shape [None | dict | list | tuple] Specify for which axes the dropout mask will be broadcasted (= re-used). Use 1 for broadcasting and None otherwise. When using a dict, the default axis labels can be used (see Managing Axes below). To disable broadcasting for all axes {"*": None} can be used. Note that the the dropout mask will always be shared inside a recurrent layer for all recurrent steps.

dropout_on_forward [bool] if set to true, will also apply dropout during all tasks, and not only during training.

spatial_smoothing [float] if specified, add spatial-smoothing loss of the layer output with the given factor to the total constraints.

register_as_extern_data [str] register the output of the layer as an accessable entry of extern_data.

Connecting Layers

In most cases it is sufficient to just specify a list of layer names for the from attribute. When no input is specified, it will automatically fallback to "data", which is the default input-data of the provided dataset. Depending on the definition of the feature and target keys (see Dataset.DatasetSeq), the data can be accessed via from["data:DATA_KEY"]. When specifying layers inside a recurrent unit (see Recurrent Layers), two additional input prefixes are available, base and prev. When trying to access layers from outside the recurrent unit, the prefix base as to be used. Otherwise, only other layers inside the recurrent unit are recognised. prev can be used to access the layer output from the previous recurrent step (e.g. for target embedding feedback).

Layer Initialization

RETURNN offers multiple methods of initializing layers. This is usually done by setting the parameter "forward_weights_init" in layers that have trainable parameters. The methods for initializations include, but are not limited to:

  • providing a single value (will map to tf.initializers.constant)
  • providing the (lowercase) name of a given tensorflow initializer <>`_, which can be e.g.:
    • "glorot_normal"
    • "glorot_uniform"
    • "orthogonal"
  • providing a dictionary for the initializer classes:
    • Example: "forward_weights_init": {'class': 'VarianceScaling', 'scale': 0.5, 'mode': 'fan_out'}

The initialization is performed in TFUtil.get_initializer().

Note: the initalizers can be accessed both as e.g. "glorot_normal" or "glorot_normal_initializer".

Managing Axes

In the default case, the axes of data that is passed between layers (such as batch, time, spatial and feature) are not visible to the user, and handled by RETURNN internally with the help of objects. For layers that operate on specific axes, meaning they have an axis or axes parameter, different identifier (strings) can be used to select the correct axes. These identifier are e.g.

  • *: select all axes
  • B|batch: select the batch axis
  • T|time: select the time axis
  • F|feature select the feature axis
  • spatial select all spatial axes (not batch and not feature)
  • S:<int>|spatial:<int> select a single spatial axis from the list of all spatial axes (zero-based, can be negative)
  • dyn|dynamic select all dynamic axes (all spacial axes with dynamic time and time even if it has no dynamic length)
  • D:<int>|dyn:<int>|dynamic:<int> select a specific dynamic axis (zero-based, can be negative)
  • static select all static axes (not batch, and has a fixed dimension)
  • static:<int> select a specific static axis
  • T? select time axis if existing, none otherwise
  • spatial_except_time select all spatial axes but also not the time axis
  • except_time select all axes except time and batch axis
  • except_batch select all axes except batch axis

Note that all identifier can be used case-insensitive. For axes parameter it is also possible to provide a tuple or list of the above identifiers. If something is unclear, or not working as intended, please refer to Data.get_axes_from_description().