class ngraph::op::v0::LSTMCell

Overview

Class for single lstm cell node. More…

#include <lstm_cell.hpp>

class LSTMCell: public ngraph::op::util::RNNCellBase
{
public:
    // fields

    static constexpr NodeTypeInfo type_info {"LSTMCell", 0};

    // construction

    LSTMCell();

    LSTMCell(
        const Output<Node>& X,
        const Output<Node>& initial_hidden_state,
        const Output<Node>& initial_cell_state,
        const Output<Node>& W,
        const Output<Node>& R,
        std::size_t hidden_size,
        LSTMWeightsFormat weights_format = LSTMWeightsFormat::IFCO,
        const std::vector<std::string>& activations = std::vector<std::string>{"sigmoid", "tanh", "tanh"},
        const std::vector<float>& activations_alpha = {},
        const std::vector<float>& activations_beta = {},
        float clip = 0.f,
        bool input_forget = false
        );

    LSTMCell(
        const Output<Node>& X,
        const Output<Node>& initial_hidden_state,
        const Output<Node>& initial_cell_state,
        const Output<Node>& W,
        const Output<Node>& R,
        const Output<Node>& B,
        std::size_t hidden_size,
        LSTMWeightsFormat weights_format = LSTMWeightsFormat::IFCO,
        const std::vector<std::string>& activations = std::vector<std::string>{"sigmoid", "tanh", "tanh"},
        const std::vector<float>& activations_alpha = {},
        const std::vector<float>& activations_beta = {},
        float clip = 0.f,
        bool input_forget = false
        );

    LSTMCell(
        const Output<Node>& X,
        const Output<Node>& initial_hidden_state,
        const Output<Node>& initial_cell_state,
        const Output<Node>& W,
        const Output<Node>& R,
        const Output<Node>& B,
        const Output<Node>& P,
        std::size_t hidden_size,
        LSTMWeightsFormat weights_format = LSTMWeightsFormat::IFCO,
        const std::vector<std::string>& activations = std::vector<std::string>{"sigmoid", "tanh", "tanh"},
        const std::vector<float>& activations_alpha = {},
        const std::vector<float>& activations_beta = {},
        float clip = 0.f,
        bool input_forget = false
        );

    // methods

    virtual const NodeTypeInfo& get_type_info() const;
    virtual void validate_and_infer_types();
    virtual bool visit_attributes(AttributeVisitor& visitor);
    virtual std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const;
    bool get_input_forget() const;
    LSTMWeightsFormat get_weights_format() const;
};

Inherited Members

public:
    // typedefs

    typedef DiscreteTypeInfo type_info_t;
    typedef std::map<std::string, std::shared_ptr<Variant>> RTMap;

    // fields

    NGRAPH_DEPRECATED("The tensor name was deprecated. Use get_input_tensor(i).get_names() instead.") const std std::unordered_set<descriptor::Tensor*> liveness_new_list;
    std::unordered_set<descriptor::Tensor*> liveness_free_list;

    // methods

    virtual void validate_and_infer_types();
    void constructor_validate_and_infer_types();
    virtual bool visit_attributes(AttributeVisitor&);
    virtual const op::AutoBroadcastSpec& get_autob() const;
    virtual bool has_evaluate() const;

    virtual bool evaluate(
        const HostTensorVector& output_values,
        const HostTensorVector& input_values
        ) const;

    virtual bool evaluate(
        const HostTensorVector& output_values,
        const HostTensorVector& input_values,
        const EvaluationContext& evaluationContext
        ) const;

    virtual bool evaluate_lower(const HostTensorVector& output_values) const;
    virtual bool evaluate_upper(const HostTensorVector& output_values) const;

    virtual bool constant_fold(
        OutputVector& output_values,
        const OutputVector& inputs_values
        );

    virtual OutputVector decompose_op() const;
    virtual const type_info_t& get_type_info() const = 0;
    const char* get_type_name() const;
    void set_arguments(const NodeVector& arguments);
    void set_arguments(const OutputVector& arguments);
    void set_argument(size_t position, const Output<Node>& argument);

    void set_output_type(
        size_t i,
        const element::Type& element_type,
        const PartialShape& pshape
        );

    void set_output_size(size_t output_size);
    void invalidate_values();
    virtual void revalidate_and_infer_types();
    virtual std::string description() const;
    const std::string& get_name() const;
    void set_friendly_name(const std::string& name);
    const std::string& get_friendly_name() const;
    virtual bool is_dynamic() const;
    size_t get_instance_id() const;
    virtual std::ostream& write_description(std::ostream& os, uint32_t depth = 0) const;
    const std::vector<std::shared_ptr<Node>>& get_control_dependencies() const;
    const std::vector<Node*>& get_control_dependents() const;
    void add_control_dependency(std::shared_ptr<Node> node);
    void remove_control_dependency(std::shared_ptr<Node> node);
    void clear_control_dependencies();
    void clear_control_dependents();
    void add_node_control_dependencies(std::shared_ptr<Node> source_node);
    void add_node_control_dependents(std::shared_ptr<Node> source_node);
    void transfer_control_dependents(std::shared_ptr<Node> replacement);
    size_t get_output_size() const;
    const element::Type& get_output_element_type(size_t i) const;
    const element::Type& get_element_type() const;
    const Shape& get_output_shape(size_t i) const;
    const PartialShape& get_output_partial_shape(size_t i) const;
    Output<const Node> get_default_output() const;
    Output<Node> get_default_output();
    virtual size_t get_default_output_index() const;
    size_t no_default_index() const;
    const Shape& get_shape() const;
    descriptor::Tensor& get_output_tensor(size_t i) const;
    descriptor::Tensor& get_input_tensor(size_t i) const;
    NGRAPH_DEPRECATED("The tensor name was deprecated. Use get_output_tensor(i).get_names() instead.") const std std::set<Input<Node>> get_output_target_inputs(size_t i) const;
    size_t get_input_size() const;
    const element::Type& get_input_element_type(size_t i) const;
    const Shape& get_input_shape(size_t i) const;
    const PartialShape& get_input_partial_shape(size_t i) const;
    Node* get_input_node_ptr(size_t index) const;
    std::shared_ptr<Node> get_input_node_shared_ptr(size_t index) const;
    Output<Node> get_input_source_output(size_t i) const;
    virtual std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& inputs) const = 0;
    std::shared_ptr<Node> copy_with_new_inputs(const OutputVector& new_args) const;

    std::shared_ptr<Node> copy_with_new_inputs(
        const OutputVector& inputs,
        const std::vector<std::shared_ptr<Node>>& control_dependencies
        ) const;

    bool has_same_type(std::shared_ptr<const Node> node) const;
    RTMap& get_rt_info();
    const RTMap& get_rt_info() const;
    const std::unordered_set<std::string>& get_provenance_tags() const;
    void add_provenance_tag(const std::string& tag);

    template <typename T>
    void add_provenance_tags(T tag_set);

    void add_provenance_tags_above(
        const OutputVector& base,
        const std::unordered_set<std::string>& tag_set
        );

    void remove_provenance_tag(const std::string& tag);
    void add_provenance_group_member(const std::shared_ptr<Node>& node);
    void remove_provenance_group_member(const std::shared_ptr<Node>& node);

    void replace_provenance_group_member(
        const std::shared_ptr<Node>& current_node,
        const std::shared_ptr<Node>& replacement_node
        );

    const std::set<std::shared_ptr<Node>>& get_provenance_group_members() const;
    std::shared_ptr<Node> add_provenance_group_members_above(const OutputVector& base);
    void merge_provenance_tags_from(const std::shared_ptr<const Node>& source);
    void transfer_provenance_tags(const std::shared_ptr<Node>& replacement);
    NodeVector get_users(bool check_is_used = false) const;
    virtual size_t get_version() const;
    virtual std::shared_ptr<Node> get_default_value() const;
    bool operator < (const Node& other) const;
    std::vector<Input<Node>> inputs();
    std::vector<Input<const Node>> inputs() const;
    std::vector<Output<Node>> input_values() const;
    std::vector<Output<Node>> outputs();
    std::vector<Output<const Node>> outputs() const;
    Input<Node> input(size_t input_index);
    Input<const Node> input(size_t input_index) const;
    Output<Node> input_value(size_t input_index) const;
    Output<Node> output(size_t output_index);
    Output<const Node> output(size_t output_index) const;
    void set_op_annotations(std::shared_ptr<ngraph::op::util::OpAnnotations> op_annotations);
    std::shared_ptr<ngraph::op::util::OpAnnotations> get_op_annotations() const;

    virtual bool match_value(
        pattern::Matcher* matcher,
        const Output<Node>& pattern_value,
        const Output<Node>& graph_value
        );

    virtual bool match_node(
        pattern::Matcher* matcher,
        const Output<Node>& graph_value
        );

    void validate_input_rank_dimension(const std::vector<ngraph::PartialShape>& input);
    virtual bool visit_attributes(AttributeVisitor& visitor);
    std::size_t get_hidden_size() const;
    float get_clip() const;
    const std::vector<std::string>& get_activations() const;
    const std::vector<float>& get_activations_alpha() const;
    const std::vector<float>& get_activations_beta() const;

Detailed Documentation

Class for single lstm cell node.

Following implementation supports:

It calculates following equations:

it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi)
ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf)
ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc)
Ct = ft (.) Ct-1 + it (.) ct
ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo)
Ht = ot (.) h(Ct)

*       - Is a dot product,
(.)     - is a Hadamard product (element-wise),
f, g, h - are activation functions.

This class represents only single cell (for current time step) and not the whole LSTM Sequence layer

See also:

LSTMSequence, RNNCell, GRUCell

Construction

LSTMCell(
    const Output<Node>& X,
    const Output<Node>& initial_hidden_state,
    const Output<Node>& initial_cell_state,
    const Output<Node>& W,
    const Output<Node>& R,
    std::size_t hidden_size,
    LSTMWeightsFormat weights_format = LSTMWeightsFormat::IFCO,
    const std::vector<std::string>& activations = std::vector<std::string>{"sigmoid", "tanh", "tanh"},
    const std::vector<float>& activations_alpha = {},
    const std::vector<float>& activations_beta = {},
    float clip = 0.f,
    bool input_forget = false
    )

Constructs LSTMCell node.

Parameters:

X

The input tensor with shape: [batch_size, input_size].

initial_hidden_state

The hidden state tensor at current time step with shape: [batch_size, hidden_size].

initial_cell_state

The cell state tensor at current time step with shape: [batch_size, hidden_size].

W

The gate weights tensor with shape: [4*hidden_size, input_size].

R

The recurrence weights tensor with shape: [4*hidden_size, hidden_size].

hidden_size

The number of hidden units for recurrent cell.

weights_format

The order of gates in weights tensors. The default format is IFCO since it is used by DNNL.

activations

The vector of activation functions used inside recurrent cell.

activations_alpha

The vector of alpha parameters for activation functions in order respective to activation list.

activations_beta

The vector of beta parameters for activation functions in order respective to activation list.

clip

The value defining clipping range [-clip, clip] on input of activation functions.

input_forget

Controls coupling input and forget gates.

LSTMCell(
    const Output<Node>& X,
    const Output<Node>& initial_hidden_state,
    const Output<Node>& initial_cell_state,
    const Output<Node>& W,
    const Output<Node>& R,
    const Output<Node>& B,
    std::size_t hidden_size,
    LSTMWeightsFormat weights_format = LSTMWeightsFormat::IFCO,
    const std::vector<std::string>& activations = std::vector<std::string>{"sigmoid", "tanh", "tanh"},
    const std::vector<float>& activations_alpha = {},
    const std::vector<float>& activations_beta = {},
    float clip = 0.f,
    bool input_forget = false
    )

Constructs LSTMCell node.

Parameters:

X

The input tensor with shape: [batch_size, input_size].

initial_hidden_state

The hidden state tensor at current time step with shape: [batch_size, hidden_size].

initial_cell_state

The cell state tensor at current time step with shape: [batch_size, hidden_size].

W

The weight tensor with shape: [4*hidden_size, input_size].

R

The recurrence weight tensor with shape: [4*hidden_size, hidden_size].

B

The bias tensor for gates with shape: [4*hidden_size].

hidden_size

The number of hidden units for recurrent cell.

weights_format

The order of gates in weights tensors. The default format is IFCO since it is used by DNNL.

activations

The vector of activation functions used inside recurrent cell.

activations_alpha

The vector of alpha parameters for activation functions in order respective to activation list.

activations_beta

The vector of beta parameters for activation functions in order respective to activation list.

clip

The value defining clipping range [-clip, clip] on input of activation functions.

input_forget

Controls coupling input and forget gates.

LSTMCell(
    const Output<Node>& X,
    const Output<Node>& initial_hidden_state,
    const Output<Node>& initial_cell_state,
    const Output<Node>& W,
    const Output<Node>& R,
    const Output<Node>& B,
    const Output<Node>& P,
    std::size_t hidden_size,
    LSTMWeightsFormat weights_format = LSTMWeightsFormat::IFCO,
    const std::vector<std::string>& activations = std::vector<std::string>{"sigmoid", "tanh", "tanh"},
    const std::vector<float>& activations_alpha = {},
    const std::vector<float>& activations_beta = {},
    float clip = 0.f,
    bool input_forget = false
    )

Constructs LSTMCell node.

Parameters:

X

The input tensor with shape: [batch_size, input_size].

initial_hidden_state

The hidden state tensor at current time step with shape: [batch_size, hidden_size].

initial_cell_state

The cell state tensor at current time step with shape: [batch_size, hidden_size].

W

The weight tensor with shape: [4*hidden_size, input_size].

R

The recurrence weight tensor with shape: [4*hidden_size, hidden_size].

B

The bias tensor for gates with shape: [4*hidden_size].

P

The weight tensor for peepholes with shape: [3*hidden_size] - 3 equals to only iof gates. The order is: input, output, forget gates.

hidden_size

The number of hidden units for recurrent cell.

weights_format

The order of gates in weights tensors. The default format is IFCO since it is used by DNNL.

activations

The vector of activation functions used inside recurrent cell.

activations_alpha

The vector of alpha parameters for activation functions in order respective to activation list.

activations_beta

The vector of beta parameters for activation functions in order respective to activation list.

clip

The value defining clipping range [-clip, clip] on input of activation functions.

input_forget

Controls coupling input and forget gates.

Methods

virtual const NodeTypeInfo& get_type_info() const

Returns the NodeTypeInfo for the node’s class. During transition to type_info, returns a dummy type_info for Node if the class has not been updated yet.

virtual void validate_and_infer_types()

Verifies that attributes and inputs are consistent and computes output shapes and element types. Must be implemented by concrete child classes so that it can be run any number of times.

Throws if the node is invalid.