from dictlearn.config_registry import ConfigRegistry
lm_config_registry = ConfigRegistry()
lm_config_registry.set_root_config({
# data
'data_path': 'onebillionword/',
'dict_path': "",
'vocab_path': "",
'dict_vocab_path': "",
'layout': 'standard',
'num_input_words': 10000,
'def_num_input_words': 0, #0 => num_input_words
'num_output_words': 10000,
'max_length': 100,
'batch_size': 64,
'batch_size_valid': 64,
'max_def_length': 100,
'max_def_per_word': 1000,
'exclude_top_k': 0,
# model
'emb_dim': 500,
'emb_def_dim': 500,
'dim': 500,
'compose_type': 'sum',
'disregard_word_embeddings': False,
'learning_rate': 0.001,
'momentum': 0.9,
'grad_clip_threshold': 5.0,
# embeddings
from dictlearn.config_registry import ConfigRegistry
nli_esim_config_registry = ConfigRegistry()
# Each epoch has ~500k examples
# Params copied from https://github.com/NYU-MLL/multiNLI/blob/master/python/util/parameters.py
nli_esim_config_registry.set_root_config({
'data_path':
'snli',
'layout':
'snli',
# Lookup params
'max_def_per_word':
100000,
'emb_dim':
300,
'bn':
0,
'dim':
300,
'dict_path':
'',
'vocab':
'',
'vocab_text':
'', # Defaults to vocab. Use when original vocab cannot be used for frequency in dict
'encoder':
'bilstm',
# Also used in NYU-MLI
from dictlearn.config_registry import ConfigRegistry
configs_ae = ConfigRegistry()
configs_ae.set_root_config({
# data_path: not useful to use that, it's better to use FUEL_DATA_PATH
# so that we can keep identical configs for different dictionaries
'data_path': '',
# the following param was useful to run a baseline without an encoder
# would be similar to word2vec with only one target word (the defined word)
# this is NOT the baseline in the paper, it is weaker than word2vec
'vocab_keys_path': '',
'layout' : 'dict', # don't change. TODO remove this option
# num_input_words can be set lower than the number of lines in vocab.txt
# this allows to replace rare words with UNK (for example, if set to all the words
# from line 10000 on will be replaced by UNK token if it is set to 10000)
'num_input_words' : 10000,
# same for num_output_words: the loss will ignore words that are ranked
# above the value
'num_output_words': 10000,
# max definition length
'max_length' : 100,
'batch_size' : 32,
'batch_size_valid' : 32,
# model
'encoder': 'lstm', # experimental code with bilstm variants (see seq2seq.py)
'decoder': 'skip-gram', # do not change?
# You should use emb_dim = dim unless you're playing with more experimental
# code.
'emb_dim' : 300,
from dictlearn.config_registry import ConfigRegistry
qa_config_registry = ConfigRegistry()
qa_config_registry.set_root_config({
# data
'data_path': "",
'dict_path': "",
'vocab_path': "",
'dict_vocab_path': "",
'embedding_path': "",
'layout': 'standard',
'num_input_words': 10000,
'def_num_input_words': 0,
'max_length': 100,
'batch_size': 32,
'batch_size_valid': 32,
# retrieval hacks
'max_def_length': 1000,
'with_too_long_defs': 'drop',
'max_def_per_word': 1000,
'with_too_many_defs': 'random',
'exclude_top_k': 0,
# model
'def_reader': 'LSTMReadDefinitions',
'dim': 128,
'emb_dim': 0,
'readout_dims': [],
'coattention': True,
'learning_rate': 0.001,
from dictlearn.config_registry import ConfigRegistry
snli_config_registry = ConfigRegistry()
snli_config_registry.set_root_config({
'data_path': 'snli/',
'layout': 'snli',
# Lookup params
'translate_dim': 300,
'max_def_per_word': 100000,
'bn': True,
'mlp_dim': 600,
'emb_dim': 300, # Used for def and word lookup
'dict_path': '',
# Remove by default embeddings. Our goal ATM is to beat random init
'embedding_path':
'', #/data/lisa/exp/jastrzes/dict_based_learning/data/snli/glove.840B.300d.npy',
'vocab_def': '',
'vocab_text':
'', # If passed will be used for exclude_top_k in Retrieval only
'vocab': '',
'def_dim': 300, # LSTM reader hidden state or translate in MeanPool
'def_emb_dim': -1, # Dimensionality of vectors used in definitions
'compose_type': '',
'disregard_word_embeddings': False,
'exclude_top_k': -1,
'max_def_length': 50,
'with_too_long_defs': 'drop',
'train_emb':
from dictlearn.config_registry import ConfigRegistry
lm_config_registry = ConfigRegistry()
lm_config_registry.set_root_config({
# data
'data_path': "",
'dict_path': "",
'layout': 'standard',
'num_input_words': 10000,
'num_output_words': 10000,
'max_length': 100,
'batch_size': 32,
'batch_size_valid': 32,
'max_def_length': 1000,
'exclude_top_k': -1,
# model
'dim': 128,
'compose_type': 'sum',
'standalone_def_rnn': True,
'disregard_word_embeddings': False,
'learning_rate': 0.001,
'momentum': 0.9,
'grad_clip_threshold': 5.0,
# monitoring and checkpointing
'mon_freq_train': 10,
'mon_freq_valid': 1000,
'save_freq_batches': 1000,
'n_batches': 0,
'monitor_parameters': False