fb15k237-complex-unigrampool.yaml

############### META

# cuda devices to use
devices: 0
# turn off cuda, i.e. GPU usage
no_cuda: false
# seed for the RNG
seed: 0
# save output to base directory
results_dir: data/experiments/fb237/
# leave empty if left empty (null, None) then a name is automatically generated
experiment_dir: null

############### TASKS

# train and validate
train: true
# evaluate on test
evaluate: false

############### MODEL

# configure the model class
model: UnigramPoolingComplexRelationModel
# configure the model's arguments
model_config:
  dropout: 0.1
  entity_slot_size: 64
  init_std: 0.1
  normalize: batchnorm
  relation_slot_size: 512
  sparse: false

experiment_settings:
  # choose loss out of bce, kl, ce
  loss: bce
  # should entity embeddings be computed from tokens?
  replace_entities_by_tokens: true
  # should relation embeddings be computed from tokens?
  replace_relations_by_tokens: true
  # max token length for subject and relation mentions
  max_lengths_tuple: [10, 10]


############### TRAINING

# max epochs to run
epochs: 100
# global batch size (can be overridden for train, valid and test,; see DATASET)
batch_size: 512
# label smoothing for BCE loss
bce_label_smoothing: 0.0
# learning rate schduler config kwargs dict to tunnel through to pytorch;
# set lr_scheduler to a Pytorch scheduler, none/None else
lr_scheduler_config: null
# gradient clipping
grad_clip: 0


# optimizer config kwargs dict to tunnel through to Pytorch; set optimizer
# to a known Pytorch optimizer and the rest to its arguments
optimization_config:
  optimizer: Adagrad
  epoch: 0
  lr: 0.1
  weight_decay: 1.0e-10

# print stats every this many batch steps
print_freq: 100
# save every this many epochs
save_epoch_freq: 25
# save every this many batch steps
save_freq: -1
# evaluate every this many epochs
eval_epoch_freq: 5
# evaluate every this many batch steps
eval_freq: 2000

############### MODEL SELECTION

# Perform model selection, early stopping, keeping best models according to these
# metrics. The first of the list will be used for early stopping, for the remaining
# metrics the best models will be saved.
# Available metrics: loss, mrr, mr, h1, h3, h10, h50
model_select_metric:
- mrr

# Patience for early stopping for this many epochs no improvement should occur
# before we stop training
patience_epochs: 50
# improvements smaller than this will not be counted
patience_metric_change: 1.0e-05
# depending on the metric (smaller better or larger better), set the minimum or
# maximum value the metric shoulkd achieve after the initial patience_epochs
patience_metric_min_treshold: null
patience_metric_max_treshold: null


############### DATASET


# Configure the default dataset class for all splits, with the following choices:
#
# OneToNMentionRelationDataset:
#
#   Dataset format:
#
#   This dataset class works especially with KGE dataset where the entities
#   can be realized with different alternative mentions.
#
#   The data format contains the following TAB separated fields per line:
#   5 TAB separated columns, COL 4 and COL 5 are lists of space seperated ids:
#
#   COL 1		COL 2           COL 3		COL 4			COL 5
#   entity id	relation id	    entity id	subj entity ids	obj entity ids
#
#   This dataset can also be used for standard KGE tasks, then COL 4 and
#   COL 5 are identical to COL 1 and COL 3
#
#
#   Batch construction:
#
#   The set of answer entities of all examples in the batch are shared accross
#   all examples of the the batch as true or false labels.
#
#   For example, given a batch containing the subject-relation prefix (1, 2) with
#   object answers [3, 4], and relation-object suffix (4, 5) with subject
#   answers [1, 7, 8], then the batch will be constructed as
#
#   potential batch shared entity answers [1, 3, 4, 7, 8]
#
#   sp prefixes: (1, 2) labels: [0, 1, 1, 0, 0]
#
#   po prefixes: (4, 5) labels: [1, 0, 0, 1, 1]
#
#
#   Apart from batch_size, the batch construction is influcenced by the following
#   two parameters:
#
#   Minimum label size for *batch construction*, i.e. when are less batch shared
#   entity answers than min_size_batch_labels, then new randomly sampled entities
#   are added to that set.
#
#   min_size_batch_labels: 4096
#
#
#   Maximium label size during *creation of training examples*, i.e. if a
#   (subject,relation) prefix has more than max_size_prefix_label number of true
#   answers the answers are split up into chunks. The (subject,relation) example
#   is then repeated during training for each chunk.
#
#   max_size_prefix_label: 64

dataset_dir: data/fb15k237/mapped_to_ids

dataset_class: OneToNMentionRelationDataset

# this overrides the default dataset_class
training_dataset_class: OneToNMentionRelationDataset

# training data settings
train_data_config:
  input_file: train.txt
  batch_size: 512 # if batch size undefined here then global batch size is used
  use_batch_shared_entities: False
#  min_size_batch_labels: 4096
#  max_size_prefix_label: 64


# this overrides the default dataset_class
validation_dataset_class: OneToNMentionRelationDataset
# validation data settings
val_data_config:
  input_file: valid.txt
  batch_size: 512 # if batch size undefined here then global batch size is used
  use_batch_shared_entities: False
#  min_size_batch_labels: -1
#  max_size_prefix_label: -1

# this overrides the default dataset_class
test_dataset_class: OneToNMentionRelationDataset
# test data settings
test_data_config:
  input_file: test.txt
  batch_size: 512 # if batch size undefined here then global batch size is used
  use_batch_shared_entities: False
#  min_size_batch_labels: -1
#  max_size_prefix_label: -1


############### EVALUATION

evaluate_on_validation: true
evaluate_scores_file: null
log_predictions: false


############### RESUMING

# set a checkpoint path here, in conjunction with 'train: false', 'evaluate: true'
# 'evaluate_on_validation: false' this is used for testing.
resume: ''
# a whitelist of the parameter names for resuming a model
resume_filter: null
# a blacklist of the parameter names for resuming a model
resume_freeze: null
# load the args from the checkoint instead of this configurations args
resume_load_args: true
# reset optimizer when resuming
reset_optimizer: false
# start epoch for optimzation regime
#start_epoch: 0