package syntaxnet.dragnn

Get desktop application:
View/edit binary Protocol Buffers messages

Complete specification for a single task.

Used in: MasterSpec

• optional string name = 1

  Name for this component: this is used in linked features via the "source_component" field.

• optional RegisteredModuleSpec transition_system = 2

  TransitionSystem to use.

• repeated Resource resource = 3

  Resources that this component depends on. These are copied to TaskInputs when calling SAFT code.

• repeated FixedFeatureChannel fixed_feature = 4

  Feature space configurations.

• repeated LinkedFeatureChannel linked_feature = 5

• optional RegisteredModuleSpec network_unit = 6

  Neural Network builder specification.

• optional RegisteredModuleSpec backend = 7

  The registered C++ implementation of the dragnn::Component class; e.g. "SyntaxNetComponent".

• optional int32 num_actions = 8

  Number of possible actions from every state.

• optional string attention_component = 9

  Specify the name of the lower level component on which it has attention.

• optional RegisteredModuleSpec component_builder = 10

  Options for the ComponentBuilder. If this is empty, the regular tf.while_loop based builder is assumed.

• optional int32 training_beam_size = 11

  Default max number of active states for beam training.

• optional int32 inference_beam_size = 12

  Default max number of active states for beam inference.

The trace for a single step of a single Component.

Used in: ComponentTrace

• optional string caption = 1

  A caption/description to describe this step. This should fit in a graphical node rendered to the screen.

• repeated FixedFeatureChannelTrace fixed_feature_trace = 2

• repeated LinkedFeatureChannelTrace linked_feature_trace = 3

• optional string html_representation = 4

  An *HTML-language* representation of the current state.

• repeated double outcome_score = 5

  The scores for each potential decision. (The mapping from index to name is managed by the component.)

• optional bool step_finished = 6

  Set to true once the step is finished. (This allows us to open a step after each transition, without having to know if it will be used.)

The traces for all steps for a single Component.

Used in: MasterTrace

• optional string name = 1

  Name of the component; should match the ComponentSpec.

• repeated ComponentStepTrace step_trace = 2

  The steps that have been taken by this Component.

Main proto being used to trace parsing.

• repeated MasterTrace master_trace = 1

  For each sentence, there is a sequence of state sets storing tracing information.

Specification for a feature channel that is a *fixed* function of the input. NEXT_ID: 10

Used in: ComponentSpec

• optional string name = 1

  Interpretable name for this feature channel. NN builders might depend on this to determine how to hook different channels up internally.

• optional string fml = 2

  String describing the FML for this feature channel.

• optional int32 embedding_dim = 3

  Dimensions of embedding space, or -1 if the feature should not be embedded.

• optional int32 vocabulary_size = 4

  No. of possible values returned.

• optional int32 size = 5

  No. of different feature templates in the channel, i.e. the # of features that will be concatenated but share the embedding for this channel.

• optional bool is_constant = 9

  Whether the embeddings for this channel should be held constant at their pretrained values, instead of being trained. Pretrained embeddings are required when true.

• optional string predicate_map = 6

  Predicate map for compacting feature values.

• optional Resource pretrained_embedding_matrix = 7

  Pointer to a pretrained embedding matrix for this feature set.

• optional Resource vocab = 8

  Vocab file, containing all vocabulary words one per line.

Describes single embedding "group", e.g., 'words', 'tags'. Each group shares an embedding space.

Used in: ComponentStepTrace

• optional string name = 1

  string-valued name of the group, e.g., 'words'.

• repeated FixedFeatures value_trace = 2

  The feature functions active in this embedding group.

A fixed sparse bag of features in DRAGNN. The id, weight, and description fields are all aligned if present (ie, any of these that are non-empty should have the same # items). If weight is omitted, 1.0 is used. These features as interepreted as multiple firings of a single feature template: e.g., for a single focus word, a bag of ngrams.

Used in: FixedFeatureChannelTrace

• repeated uint64 id = 1

• repeated float weight = 2

• repeated string value_name = 3

  string-valued description of each *feature value*. (Only used for debugging.)

• optional string feature_name = 4

  string-valued name of feature. (Only used for debugging.)

A hyperparameter configuration for a training run. NEXT ID: 22

Used in: GridPoint.CompositeOptimizerSpec, TrainingGridSpec

• optional double learning_rate = 1

  Global learning rate initialization point.

• optional double momentum = 2

  Momentum coefficient when using MomentumOptimizer.

• optional double decay_base = 16

  Decay rate and base for global learning rate decay. The learning rate is reduced by a factor of |decay_base| every |decay_steps|.

• optional int32 decay_steps = 3

• optional bool decay_staircase = 17

  Whether to decay the learning rate in a "staircase" manner. If true, the rate is adjusted exactly once every |decay_steps|. Otherwise, the rate is adjusted in smaller increments on every step, such that the overall rate of decay is still |decay_base| every |decay_steps|.

• optional int32 seed = 4

  Random seed to initialize parameters.

• optional string learning_method = 7

  Specify the optimizer used in training, the default is MomentumOptimizer.

• optional bool use_moving_average = 8

  Whether or not to use a moving average of the weights in inference time.

• optional double average_weight = 9

  Rolling average update co-efficient.

• optional double dropout_rate = 10

  The dropout *keep* probability rate used in the model. 1.0 = no dropout.

• optional double recurrent_dropout_rate = 20

  The dropout *keep* probability rate for recurrent connections. If < 0.0, recurrent connections should use |dropout_rate| instead. 1.0 = no dropout.

• optional double gradient_clip_norm = 11

  Gradient clipping threshold, applied if greater than zero. A value in the range 1-20 seems to work well to prevent large learning rates from causing problems for updates at the start of training.

• optional GridPoint.CompositeOptimizerSpec composite_optimizer_spec = 12

• optional double adam_beta1 = 13

  Parameters for Adam training.

• optional double adam_beta2 = 14

• optional double adam_eps = 15

• optional double l2_regularization_coefficient = 18

  Coefficient for global L2 regularization.

• optional double self_norm_alpha = 19

  Coefficient for global self normalization regularization. A value of zero turns it off.

• optional string self_norm_components_filter = 21

  Comma separated list of components to which self_norm_alpha should be restricted. If left empty, no filtering will take place. Typically a single component.

A spec for using multiple optimization methods.

Used in: GridPoint

• optional GridPoint method1 = 1

  First optimizer.

• optional GridPoint method2 = 2

  Second optimizer.

• optional int32 switch_after_steps = 3

  After this number of steps, switch from first to second.

A feature in DRAGNN thats link a component to another or a component to itself recurrently. If batch_idx or beam_idx are omitted, 0 is used.

Used in: LinkedFeatureChannelTrace

• optional int64 batch_idx = 1

  Index into the {step x batch x beam} activations workspace generated by the previous computation.

• optional int64 beam_idx = 2

• optional int64 step_idx = 3

• optional int64 feature_value = 4

  Values in the original feature space. This is ignored in TensorFlow.

• optional string feature_name = 5

  string-valued name of feature. (Only used for debugging.)

Specification for a feature channel that *links* to component activations. Note that the "vocabulary" of these features is the activations that they are linked to, so it is determined by the other components in the spec.

Used in: ComponentSpec

• optional string name = 1

  Interpretable name for this feature channel. NN builders might depend on this to determine how to hook different channels up internally.

• optional string fml = 2

  Feature function specification. Note: these should all be of type LinkedFeatureType.

• optional int32 embedding_dim = 3

  Embedding dimension, or -1 if the link should not be embedded.

• optional int32 size = 4

  No. of different feature templates in the channel, i.e. the # of features that will be concatenated but share the embedding for this channel.

• optional string source_component = 5

  Component to use for translation, e.g. "tagger"

• optional string source_translator = 6

  Translator target, e.g. "token" or "last_action", to translate raw feature values into indices. This must be interpretable by the Component referenced by source_component.

• optional string source_layer = 7

  Layer that these features should connect to.

Trace for an entire linked feature channel.

Used in: ComponentStepTrace

• optional string name = 1

  Name of the embedding space.

• optional string source_component = 2

  The component that this feature links to.

• optional string source_translator = 3

  The string-valued name of the translator function that maps a feature value to a step index.

• optional string source_layer = 4

  The name of the layer that we are extracting from the identified step.

• repeated LinkFeatures value_trace = 5

  Individual features within this group.

Proto to specify a set of DRAGNN components (transition systems) that are trained and evaluated jointly. Each component gets one ComponentSpec. The order of component is important: a component can only link to components that come before (for now.) NEXT ID: 6

• repeated ComponentSpec component = 1

• optional bool debug_tracing = 4

  Whether to extract debug traces.

The traces for all Components.

Used in: DragnnTrace

• repeated ComponentTrace component_trace = 1

The Parts here should be more or less compatible with TaskInput.

Used in: Resource

• optional string file_pattern = 1

• optional string file_format = 2

• optional string record_format = 3

Super generic container for any registered sub-piece of DRAGNN.

Used in: ComponentSpec

• optional string registered_name = 1

  Name of the registered class.

• map<string, string> parameters = 2

  Parameters to set while initializing this system; these are copied to Parameters in a TaskSpec when calling SAFT code, or via kwargs in TF Python code.

Fixed resources that will be converted into TaskInput's when calling SAFT code.

Used in: ComponentSpec, FixedFeatureChannel

• optional string name = 1

• repeated Part part = 2

Training target to be built into the graph.

Used in: TrainingGridSpec

• optional string name = 1

  Name for this target. This should be unique across all targets.

• repeated double component_weights = 2

  Specify the weights for different components. This should be the same size as the number of components in the spec, or empty (defaults to equal weights). Weights are normalized across the components being trained to sum to one.

• repeated bool unroll_using_oracle = 3

  Specify whether to train a component using supervised signal or not. This should be the same size as the number of components in the spec, or empty (defaults to all true).

• optional int32 max_index = 4

  Maximum length of the pipeline to train. E.g. if max_index is 1, then only the first component will be trained via this target.

A vector of hyperparameter configurations to search over.

• repeated GridPoint grid_point = 1

  Grid points to search over.

• repeated TrainTarget target = 2

  Training targets to create in the graph builder stage.