def decoder(params):

    data_path = params['data_path']
    preprocessed_v2 = params['preprocessed_v2']
    preprocessed_data = params['preprocessed_data']
    decode_path = params['decode_path']
    model_path = params['model_path']
    result_path = params['result_path']
    result_kp20k = params['result_kp20k']
    file_name = params['file_name']
    weights = params['weights']

    encoder_length = params['encoder_length']
    decoder_length = params['decoder_length']
    embedding_dim = params['embedding_dim']
    birnn_dim = params['birnn_dim']
    rnn_dim = params['rnn_dim']
    vocab_size = params['vocab_size']
    num_samples = params['num_samples']
    batch_size = params['batch_size']
    epoch = params['epoch']
    '''
	Reading vocabulary dictionaries

	'''
    indices_words_connector = DataConnector(preprocessed_v2,
                                            'all_indices_words.pkl',
                                            data=None)
    indices_words_connector.read_pickle()
    indices_words = indices_words_connector.read_file

    words_indices_connector = DataConnector(preprocessed_v2,
                                            'all_words_indices.pkl',
                                            data=None)
    words_indices_connector.read_pickle()
    words_indices = words_indices_connector.read_file

    ## merge all set into one test set for trained model

    outputs_conn = DataConnector(data_path, 'output_tokens.npy', data=None)
    outputs_conn.read_numpys()
    outputs = outputs_conn.read_file

    y_test_true = outputs

    print("Ground truth of keyphrases shape: %s" %
          str(y_test_true.shape))  # input for encoder
    sys.stdout.flush()

    # non-paired data set

    X_connector = DataConnector(preprocessed_data, 'X_pad.npy', data=None)
    X_connector.read_numpys()
    X_in = X_connector.read_file

    print("\n Non-paired test set: \n")
    sys.stdout.flush()
    print("X (input for encoder) shape: %s" %
          str(X_in.shape))  # input for encoder
    sys.stdout.flush()

    sampled_softmax = AttentionSampledSoftmax(encoder_length=encoder_length,
                                              decoder_length=decoder_length,
                                              embedding_dim=embedding_dim,
                                              birnn_dim=birnn_dim,
                                              rnn_dim=rnn_dim,
                                              vocab_size=vocab_size,
                                              num_samples=num_samples,
                                              filepath=result_kp20k,
                                              filename=file_name,
                                              batch_train_iter=None,
                                              batch_val_iter=None,
                                              batch_size=None,
                                              steps_epoch=None,
                                              val_steps=None,
                                              epochs=None)

    sampled_softmax.train_att_sampled_softmax()
    '''
	Model for retrieving softmax probability
	Return: softmax probability of prediction layer
	'''

    sampled_softmax.predict_att_sampled_softmax(weights)
    encoder_model = sampled_softmax.encoder_model
    # 1. Prediction model after being trained on sampled softmax setting
    predict_softmax_model = sampled_softmax.prediction_model
    '''
	Inference stage
	Model: layers from prediction model and decoder model
	Inference (text generation) approach: 
	1. One best search decoding (Greedy search): 
	   Return one best (top) probable word sequence, from joint probability of words within decoder time steps (decoder sequence length)
	2. N-Beam search decoding: 
	   Return N-top best most probable word sequences, by utilizing beam tree search per time steps and joint probability within decoder time steps (decoder sequence length)

	'''
    '''
	Decoder model for inference stage
	Return: generated keyphrases
	'''

    decoder_model = sampled_softmax.create_decoder_model()

    # transform tokenized y_true (ground truth of keyphrases) into full sentences / keyphrases
    keyphrases_transform = TrueKeyphrases(y_test_true)
    keyphrases_transform.get_true_keyphrases()
    keyphrases_transform.get_stat_keyphrases()
    y_true = keyphrases_transform.y_true
    max_kp_num = keyphrases_transform.max_kp_num
    mean_kp_num = keyphrases_transform.mean_kp_num
    std_kp_num = keyphrases_transform.std_kp_num

    print("Maximum number of key phrases per document in corpus: %s" %
          max_kp_num)
    sys.stdout.flush()
    print("Average number of key phrases per document in corpus: %s" %
          mean_kp_num)
    sys.stdout.flush()
    print(
        "Standard Deviation of number of key phrases per document in corpus: %s"
        % std_kp_num)
    sys.stdout.flush()

    # round up function for computing beam width
    def roundup(x):
        return x if x % 5 == 0 else x + 5 - x % 5

    beam_width = int(roundup(mean_kp_num + (3 * std_kp_num)))
    num_hypotheses = beam_width
    print("\nBeam width: %s\n" % beam_width)
    sys.stdout.flush()

    y_dummy_test = np.zeros((len(X_in), decoder_length + 1, 1))
    s0_test = np.zeros((len(X_in), rnn_dim))
    att0_test = np.zeros((len(X_in), encoder_length, 1))

    print(str(datetime.now()))
    sys.stdout.flush()

    inference_mode = Decoding(encoder_model=encoder_model,
                              decoder_model=decoder_model,
                              indices_words=indices_words,
                              words_indices=words_indices,
                              enc_in_seq=None,
                              labels=None,
                              states=None,
                              attentions=None,
                              decoder_length=decoder_length,
                              rnn_dim=rnn_dim,
                              beam_width=beam_width,
                              num_hypotheses=num_hypotheses,
                              filepath=decode_path,
                              filename=file_name)

    t0_1 = time.time()
    print("Start beam decoding...")
    sys.stdout.flush()

    beam_keyphrases = inference_mode.beam_decoder(X_in[:500],
                                                  y_dummy_test[:500],
                                                  s0_test[:500],
                                                  att0_test[:500])

    beam_decode_connector = DataConnector(decode_path,
                                          'beam_kp-%s.npy' % (file_name),
                                          beam_keyphrases)
    beam_decode_connector.save_numpys()

    t1_1 = time.time()
    print("Beam decoding is done in %.3fsec" % (t1_1 - t0_1))
    sys.stdout.flush()
예제 #2
0
def decoder(params):

    data_path = params['data_path']
    glove_embed = params['glove_embedding']
    oov_embed = params['oov_embedding']
    preprocessed_data = params['preprocessed_data']
    model_path = params['model_path']
    result_path = params['result_path']
    decode_path = params['decode_path']
    file_name = params['file_name']
    weights = params['weights']

    encoder_length = params['encoder_length']
    decoder_length = params['decoder_length']
    max_sents = params['max_sents']
    embedding_dim = params['embedding_dim']
    birnn_dim = params['birnn_dim']
    rnn_dim = params['rnn_dim']
    vocab_size = params['vocab_size']
    num_samples = params['num_samples']
    batch_size = params['batch_size']
    epoch = params['epoch']
    '''
	Reading vocabulary dictionaries

	'''
    indices_words_connector = DataConnector(preprocessed_data,
                                            'all_idxword_vocabulary_sent.pkl',
                                            data=None)
    indices_words_connector.read_pickle()
    indices_words = indices_words_connector.read_file

    words_indices_connector = DataConnector(preprocessed_data,
                                            'all_wordidx_vocabulary_sent.pkl',
                                            data=None)
    words_indices_connector.read_pickle()
    words_indices = words_indices_connector.read_file

    y_test_true_connector = DataConnector(data_path,
                                          'test_sent_output_tokens.npy',
                                          data=None)
    y_test_true_connector.read_numpys()
    y_test_true = y_test_true_connector.read_file

    # non-paired data set

    X_test_connector = DataConnector(preprocessed_data,
                                     'X_test_pad_sent.npy',
                                     data=None)
    X_test_connector.read_numpys()
    X_test = X_test_connector.read_file
    '''
	Decoder model for inference stage
	Return: generated keyphrases
	'''

    glove_embedding_conn = DataConnector(preprocessed_data,
                                         glove_embed,
                                         data=None)
    glove_embedding_conn.read_pickle()
    pretrained_embedding = glove_embedding_conn.read_file

    print("pretrained_embedding shape: %s" % str(pretrained_embedding.shape))
    print("pretrained_embedding [0][:10]: %s" %
          str(pretrained_embedding[0, :10]))
    print("pretrained_embedding [1][:10]: %s" %
          str(pretrained_embedding[1, :10]))

    oov_embedding_conn = DataConnector(preprocessed_data, oov_embed, data=None)
    oov_embedding_conn.read_pickle()
    oov_embedding = oov_embedding_conn.read_file

    print("oov_embedding shape: %s" % str(oov_embedding.shape))
    print("oov_embedding [0][:10]: %s" % str(oov_embedding[0, :10]))
    print("oov_embedding [1][:10]: %s" % str(oov_embedding[1, :10]))
    print("oov_embedding [2][:10]: %s" % str(oov_embedding[2, :10]))

    sampled_softmax = HierarchyAttSampledSoftmax(encoder_length=encoder_length,
                                                 decoder_length=decoder_length,
                                                 max_sents=max_sents,
                                                 embedding_dim=embedding_dim,
                                                 birnn_dim=birnn_dim,
                                                 rnn_dim=rnn_dim,
                                                 vocab_size=vocab_size,
                                                 num_samples=num_samples,
                                                 filepath=result_path,
                                                 filename=file_name,
                                                 batch_train_iter=None,
                                                 batch_val_iter=None,
                                                 batch_size=None,
                                                 steps_epoch=None,
                                                 val_steps=None,
                                                 epochs=None)

    # skeleton of model architecture
    sampled_softmax.train_hier_att_sampled_softmax(pretrained_embedding,
                                                   oov_embedding)

    sampled_softmax.predict_hier_att(weights)
    # 1. Prediction model after being trained on sampled softmax setting
    predict_softmax_model = sampled_softmax.prediction_model
    encoder_model = sampled_softmax.encoder_model
    '''

	Inference stage
	Model: layers from prediction model and decoder model
	Inference (text generation) approach: 
	1. One best search decoding (Greedy search): 
	   Return one best (top) probable word sequence, from joint probability of words within decoder time steps (decoder sequence length)
	2. N-Beam search decoding: 
	   Return N-top best most probable word sequences, by utilizing beam tree search per time steps and joint probability within decoder time steps (decoder sequence length)

	'''

    decoder_model = sampled_softmax.create_decoder_model()

    # transform tokenized y_true (ground truth of keyphrases) into full sentences / keyphrases
    keyphrases_transform = TrueKeyphrases(y_test_true)
    keyphrases_transform.get_true_keyphrases()
    keyphrases_transform.get_stat_keyphrases()
    y_true = keyphrases_transform.y_true
    max_kp_num = keyphrases_transform.max_kp_num
    mean_kp_num = keyphrases_transform.mean_kp_num
    std_kp_num = keyphrases_transform.std_kp_num

    print("Maximum number of key phrases per document in corpus: %s" %
          max_kp_num)
    sys.stdout.flush()
    print("Average number of key phrases per document in corpus: %s" %
          mean_kp_num)
    sys.stdout.flush()
    print(
        "Standard Deviation of number of key phrases per document in corpus: %s"
        % std_kp_num)
    sys.stdout.flush()

    # round up function for computing beam width
    def roundup(x):
        return x if x % 5 == 0 else x + 5 - x % 5

    beam_width = int(roundup(mean_kp_num + (3 * std_kp_num)))
    print("\nBeam width: %s\n" % beam_width)
    sys.stdout.flush()
    num_hypotheses = beam_width

    y_dummy_test = np.zeros((len(X_test), decoder_length + 1, 1))
    s0_test = np.zeros((len(X_test), rnn_dim))
    att0_test = np.zeros((len(X_test), encoder_length, 1))

    print("y_dummy_test shape: %s" % str(y_dummy_test.shape))
    sys.stdout.flush()
    print("s0_test shape: %s" % str(s0_test.shape))
    sys.stdout.flush()
    print("att0_test shape: %s" % str(att0_test.shape))
    sys.stdout.flush()

    print(str(datetime.now()))
    sys.stdout.flush()

    inference_mode = Decoding(encoder_model=encoder_model,
                              decoder_model=decoder_model,
                              indices_words=indices_words,
                              words_indices=words_indices,
                              enc_in_seq=None,
                              labels=None,
                              states=None,
                              attentions=None,
                              decoder_length=decoder_length,
                              rnn_dim=rnn_dim,
                              beam_width=beam_width,
                              num_hypotheses=num_hypotheses,
                              filepath=decode_path,
                              filename=file_name)

    t0_1 = time.time()
    print("Start beam decoding...")
    sys.stdout.flush()

    beam_keyphrases = inference_mode.beam_decoder(X_test[:500],
                                                  y_dummy_test[:500],
                                                  s0_test[:500],
                                                  att0_test[:500])

    beam_decode_connector = DataConnector(
        decode_path, 'beam_kp-hier-att-%s.npy' % (file_name), beam_keyphrases)
    beam_decode_connector.save_numpys()

    t1_1 = time.time()
    print("Beam decoding is done in %.3fsec" % (t1_1 - t0_1))
    sys.stdout.flush()