jack.yaml

config: './conf/jack.yaml'

description: >
  Train and Evaluate a Machine Reader

# Reader model to use, see jack/readers/implementations.py for options
reader: null

# name of this reader, serves as kind of a namespace for created parameters, defaults to name of the model
name: null

# seed will not be inherited (because of sacred) so you need to set it for all child configs
seed: 1337

# Run in debug mode, in which case the training file is also used for testing
debug: False

# If in debug mode, how many examples should be used (default 10)
debug_examples: 10

# Batch size for training data, default 128
batch_size: 128

# training file
train: null

# number of examples to train on (default is all)
num_train_examples: null

# dev file
dev: null

# Batch size for development data, defaults to batch_size when set to null
dev_batch_size: null

# number of examples to validate on (default is all)
num_dev_examples: null

# test file
test: null

# Size of the hidden representations, default 128
repr_dim: 128

# Use fixed vocab of pretrained embeddings
vocab_from_embeddings: False

# use task-specific embeddings if size bigger than 0
repr_dim_task_embedding: 0

# [word2vec], [glove] or [memory_map_dir] format of embeddings to be loaded
embedding_format: null

# embeddings to be loaded
embedding_file: null

# Optimizer, default Adam
optimizer: 'adam'

# Learning rate, default 0.001
learning_rate: 0.001

# Minimum learning rate, default 0.001
min_learning_rate: 0.0001

# Learning rate decay, default 1 (disabled)
learning_rate_decay: 1.0

# L2 regularization weight, default 0.0
l2: 0.0

# Gradients clipped between [-clip_value, clip_value] (default 0.0; no clipping)
clip_value: 0.0

# Probability for dropout on output (set to 0.0 for no dropout)
dropout: 0.0

# Number of epochs to train for, default 5
epochs: 5

# Number of batches before validation on devset. If null then after each epoch.
validation_interval: null

# Folder for tensorboard logs
tensorboard_folder: null

# Filename to log the metrics of the EvalHooks
write_metrics_to: null

# Use also character based embeddings in readers which support it
with_char_embeddings: True

# maximum number of supports to consider (as scored by tf idf against question in extractive QA for instance)
max_num_support: null

# Directory to write reader to
save_dir: './saved_reader'

# interval for logging eta, training loss, etc
log_interval: 100

# lowercase texts
lowercase: True

# path to output directory
output_dir: './out/'

# loader for the dataset, ['jack', 'squad', 'snli'] are supported. For everything else convert to jtr format first.
loader: 'jack'

# ADVANCED: if we want to (partially) initialize from a pretrained reader we can use the following option
# Note: make sure that the variable names and shapes will match
load_dir: null

# cache preprocessed data on file in $JACK_TEMP/SOME_UUID/cache, if false caches in RAM
file_cache: False