advise.kb.pbtxt

example_reader {
  batch_size: 128
  feature_dimensions: 1536
  max_stmts_per_image: 5
  max_stmt_len: 30
  max_densecaps_per_image: 5
  max_densecap_len: 10
  max_symbols_per_image: 5
  use_single_densecap: true
  shuffle_buffer_size: 20000
  number_of_regions: 10
  image_feature_path: "output/img_features_train.npy"
  region_feature_path: "output/roi_features_train.npy"
  number_of_val_examples: 10000
  statement_vocab_path: "output/action_reason_vocab_200d.txt"
  statement_annot_path: "data/train/QA_Combined_Action_Reason_train.json"
  densecap_vocab_path: "output/densecap_vocab_200d.txt"
  densecap_annot_path: "output/densecap_train.json"
  symbol_annot_path: "output/symbol_train.json"
  symbol_cluster_path: "data/additional/clustered_symbol_list.json"
}
model {
  advise_model {
    use_image_as_proposal: true
    pooling_method: ATT_POOL
    image_encoder {
      scope: "image_encoder"
      num_outputs: 200
      input_dropout_keep_prob: 0.7
      output_dropout_keep_prob: 1.0
      fc_hyperparams: {
        op: FC
        activation: NONE
        regularizer {
          l2_regularizer {
            weight: 1e-6
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          train: true
          scale: true
          center: true
          decay: 0.999
          epsilon: 0.001
        }
      }
    }
    image_decoder {
      scope: "image_decoder"
      num_outputs: 1536
      input_dropout_keep_prob: 0.7
      output_dropout_keep_prob: 1.0
      fc_hyperparams: {
        op: FC
        activation: NONE
        regularizer {
          l2_regularizer {
            weight: 1e-6
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          train: true
          scale: true
          center: true
          decay: 0.999
          epsilon: 0.001
        }
      }
    }
    image_attention_predictor {
      scope: "image_attention_predictor"
      num_outputs: 1
      input_dropout_keep_prob: 0.7
      output_dropout_keep_prob: 1.0
      fc_hyperparams: {
        op: FC
        activation: NONE
        regularizer {
          l2_regularizer {
            weight: 1e-8
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
        batch_norm {
          train: true
          scale: true
          center: true
          decay: 0.999
          epsilon: 0.001
        }
      }
    }
    stmt_encoder {
      bow_encoder {
        scope: "stmt_encoder"
        vocab_size: 25333
        embedding_size: 200
        init_width: 0.08
        weight_decay: 1e-8
        dropout_keep_prob: 0.7
        init_emb_matrix_path: "output/action_reason_vocab_200d.npy"
        repr_method: USE_OUTPUT_AVG
      }
    }
    symbol_encoder {
      bow_encoder {
        scope: "symbol_encoder"
        vocab_size: 54
        embedding_size: 200
        init_width: 0.08
        weight_decay: 1e-8
        dropout_keep_prob: 0.7
        init_emb_matrix_path: "output/symbol_vocab_200d.npy"
        repr_method: USE_OUTPUT_AVG
      }
    }
    densecap_encoder {
      bow_encoder {
        scope: "densecap_encoder"
        vocab_size: 988
        embedding_size: 200
        init_width: 0.08
        weight_decay: 1e-8
        dropout_keep_prob: 0.7
        init_emb_matrix_path: "output/densecap_vocab_200d.npy"
        repr_method: USE_OUTPUT_AVG
      }
    }
    symbol_classifier {
      scope: "symbol_classifier"
      checkpoint_path: "output/symbol_classifier/model.ckpt"
      hidden_layers: 1
      hidden_units: 100
      output_units: 54
      input_dropout_keep_prob: 0.5
      hidden_dropout_keep_prob: 0.5
      hidden_hyperparams: {
        op: FC
        activation: RELU
        regularizer {
          l2_regularizer {
            weight: 1e-8
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
      }
      output_hyperparams: {
        op: FC
        activation: NONE
        regularizer {
          l2_regularizer {
            weight: 1e-8
          }
        }
        initializer {
          truncated_normal_initializer {
            stddev: 0.03
            mean: 0.0
          }
        }
      }
    }
    joint_emb_dropout_keep_prob: 0.5
    triplet_mining {
      mine_hard {
        top_k: 32
      }
    }
    triplet_margin: 0.2
    autoencoder_loss_weight: 0.0
    densecap_loss_weight: 0.1
    symbol_loss_weight: 0.1
    log_prob_loss_weight: 0.0
    use_knowledge_branch: true
    symbol_cluster_path: "data/additional/clustered_symbol_list.json"
  }
}
train_config {
  learning_rate: 0.0001
  optimizer {
    adam {
    }
  }
  exclude_variable: "symbol_classifier"
  exclude_variable: "confidence"
  learning_rate_decay_steps: 50000
  learning_rate_decay_rate: 1.0
  learning_rate_staircase: false
  number_of_steps: 100000
  log_every_n_steps: 10
  save_interval_secs: 60
  save_summaries_secs: 120
}
eval_config {
  eval_min_global_steps: 2000
  number_of_steps: 100000
  eval_interval_secs: 60
}