IBM/pytorchpipe

# Part of pipeline containing components constituting the "Question Categorizer" pipeline.

# Inputs:
#   * tokenized_questions
#   * category_ids (targets)

# Outputs:
#   * predicted_categories (predictions)
#   * predicted_category_names

# "Question Categorizer"
# 0.51: pipe_qc_global_publisher
# 0.52: pipe_qc_question_embeddings
# 0.53: pipe_qc_lstm
# 0.54: pipe_qc_classifier

# Plus additional components for displaying/statistics:
# 0.55: pipe_qc_category_decoder
# 0.56: pipe_qc_category_accuracy

# Loaded checkpoint: 20190508_000705
checkpoint: &checkpoint ~/image-clef-2019/experiments/q_categorization/20190508_000705/checkpoints/vqa_med_question_categorization_rnn_ffn_best.pt

pipeline:
  ################# PIPE: QUESTION CATEGORIZATION #################

  # Add global variables - the ones related to only question categorization.
  pipe_qc_global_publisher:
    priority: 0.51
    type: GlobalVariablePublisher
    # Add input_size to globals.
    keys: [pipe_qc_question_encoder_output_size]
    values: [100]

  # Model 1: question embeddings
  pipe_qc_question_embeddings:
    priority: 0.52
    type: SentenceEmbeddings
    # LOAD AND FREEZE #
    load: 
      file: *checkpoint
      model: question_embeddings
    freeze: True
    ###################
    embeddings_size: 50
    pretrained_embeddings_file: glove.6B.50d.txt
    data_folder: ~/data/vqa-med
    word_mappings_file: questions.all.word.mappings.csv
    streams:
      inputs: tokenized_questions
      outputs: pipe_qc_embedded_questions
    globals:
      embeddings_size: pipe_qc_embeddings_size
  
  # Model 2: question RNN
  pipe_qc_lstm:
    priority: 0.53
    type: RecurrentNeuralNetwork
    cell_type: LSTM
    # LOAD AND FREEZE #
    load: 
      file: *checkpoint
      model: lstm
    freeze: True
    ###################
    prediction_mode: Last
    initial_state: Trainable
    use_logsoftmax: False
    streams:
      inputs: pipe_qc_embedded_questions
      predictions: pipe_qc_question_activations
    globals:
      input_size: pipe_qc_embeddings_size
      prediction_size: pipe_qc_question_encoder_output_size

  # Model 3: FFN question category
  pipe_qc_classifier:
    priority: 0.54
    type: FeedForwardNetwork
    # LOAD AND FREEZE #
    load: 
      file: *checkpoint
      model: classifier
    freeze: True
    ###################
    hidden: [50]
    dropout_rate: 0.7
    streams:
      inputs: pipe_qc_question_activations
      predictions: predicted_categories
    globals:
      input_size: pipe_qc_question_encoder_output_size # Set by global publisher
      prediction_size: num_categories # C1,C2,C3,C4, BINARY, UNK

  pipe_qc_category_decoder:
    priority: 0.55
    type: WordDecoder
    # Use the same word mappings as label indexer.
    import_word_mappings_from_globals: True
    streams:
      inputs: predicted_categories
      outputs: predicted_category_names
    globals:
      vocabulary_size: num_categories
      word_mappings: category_word_mappings

  pipe_qc_category_accuracy:
    priority: 0.56
    type: AccuracyStatistics
    streams:
      targets: category_ids
      predictions: predicted_categories
    statistics:
      accuracy: categorization_accuracy