mindmeld.models.taggers.lstm module

class mindmeld.models.taggers.lstm.LstmModel(**parameters)

Bases: mindmeld.models.taggers.taggers.Tagger

This class encapsulates the bi-directional LSTM model and provides the correct interface for use by the tagger model

apply_convolution(input_tensor, batch_size, char_window_size)

Constructs a convolution network of a specific window size

Parameters:

• input_tensor (tensor) -- The input tensor to the network
• batch_size (int) -- The batch size of the training data
• char_window_size (int) -- The character window size of each stride

Returns:

Convolved output tensor

Return type:

(Tensor)

construct_feed_dictionary(batch_examples, batch_char, batch_gaz, batch_seq_len, batch_labels=None)

Constructs the feed dictionary that is used to feed data into the tensors

Parameters:

• batch_examples (ndarray) -- A batch of examples
• batch_char (ndarray) -- A batch of character features
• batch_gaz (ndarray) -- A batch of gazetteer features
• batch_seq_len (ndarray) -- A batch of sequence length of each query
• batch_labels (ndarray) -- A batch of labels

Returns:

The feed dictionary

construct_tf_variables()

Constructs the variables and operations in the TensorFlow session graph

dump(path)

Saves the Tensorflow model

Parameters:path (str) -- the folder path for the entity model folder Returns:entity model folder Return type:path (str)

extract_features(examples, config, resources, y=None, fit=True)

Transforms a list of examples into features that are then used by the deep learning model.

Parameters:

• examples (list of mindmeld.core.Query) -- a list of queries
• config (ModelConfig) -- The ModelConfig which may contain information used for feature extraction
• resources (dict) -- Resources which may be used for this model's feature extraction
• y (list) -- A list of label sequences

Returns:

features for the LSTM network

Return type:

(sequence_embeddings, encoded_labels, groups)

fit(X, y)

Trains the model. X and y are the format of what is returned by extract_features. There is no restriction on their type or content. X should be the fully processed data with extracted features that are ready to be used to train the model. y should be a list of classes as encoded by the label_encoder

Parameters:

• X (list) -- Generally a list of feature vectors, one for each training example
• y (list) -- A list of classification labels (encoded by the label_encoder, NOT MindMeld entity objects)

Returns:

self

get_params(deep=True)

Gets a dictionary of all of the current model parameters and their values

Parameters:deep (bool) -- Not used, needed for sklearn compatibility Returns:A dictionary of the model parameter names as keys and their set values Return type:(dict)

load(path)

Loads the Tensorflow model

Parameters:path (str) -- the folder path for the entity model folder

predict(X, dynamic_resource=None)

Predicts the labels from a feature matrix X. Again X is the format of what is returned by extract_features.

Parameters:X (list) -- A list of feature vectors, one for each example Returns:a list of predicted labels (in an encoded format) Return type:(list of classification labels)

set_params(**parameters)

Initialize params for the LSTM. The keys in the parameters dictionary are as follows:

Parameters:

• parameters (dict) -- The keys in the parameters dictionary are as follows:
• number_of_epochs -- The number of epochs to run (int)
• batch_size -- The batch size for mini-batch training (int)
• token_lstm_hidden_state_dimension -- The hidden state dimension of the LSTM cell (int)
• learning_rate -- The learning rate of the optimizer (int)
• optimizer -- The optimizer used to train the network is the number of entities in the dataset (str)
• display_epoch -- The number of epochs after which the network displays common stats like accuracy (int)
• padding_length -- The length of each query, which is fixed, so some queries will be cut short in length representing the word embedding, the row index is the word's index (int)
• token_embedding_dimension -- The embedding dimension of the word (int)
• token_pretrained_embedding_filepath -- The pretrained embedding file-path (str)
• dense_keep_prob -- The dropout rate of the dense layers (float)
• lstm_input_keep_prob -- The dropout rate of the inputs to the LSTM cell (float)
• lstm_output_keep_prob -- The dropout rate of the outputs of the LSTM cell (float)
• gaz_encoding_dimension -- The gazetteer encoding dimension (int)

setup_model(config)

"Not implemented.

unload()

is_serializable