def get_args():
"""Construct the argument parser."""
parser = argparse.ArgumentParser(
description='Word embedding evaluation with Gluon.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Embeddings arguments
group = parser.add_argument_group('Embedding arguments')
group.add_argument('--embedding-path', type=str,
help='Path to a .vec in Word2Vec text foramt or '
'.bin binary fastText model file. ')
group.add_argument('--embedding-name', type=str,
help=('Name of embedding type to load. '
'Valid entries: {}'.format(
', '.join(
nlp.embedding.list_sources().keys()))))
group.add_argument('--embedding-source', type=str,
help=('Source from which to initialize the embedding.'
'Pass --list-embedding-sources to get a list of '
'valid sources for a given --embedding-name.'))
group.add_argument(
'--fasttext-load-ngrams',
action='store_true',
help=('Specify load_ngrams=True '
'when loading pretrained fastText embedding.'))
group.add_argument(
'--max-vocab-size', type=int, default=None,
help=('Only retain the X first tokens from the pre-trained embedding. '
'The tokens are ordererd by decreasing frequency.'
'As the analogy task takes the whole vocabulary into account, '
'removing very infrequent words improves performance.'))
group.add_argument('--list-embedding-sources', action='store_true')
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size', type=int, default=1024,
help='Batch size to use on analogy task. '
'Decrease batch size if evaluation crashes.')
group.add_argument('--gpu', type=int,
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-hybridize', action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir', type=str, default='logs',
help='Directory to store logs.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
validate_args(args)
evaluation.validate_args(args)
return args
def get_args():
"""Construct the argument parser."""
parser = argparse.ArgumentParser(
description='Word embedding evaluation with Gluon.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Embeddings arguments
group = parser.add_argument_group('Embedding arguments')
group.add_argument('--embedding-path', type=str,
help='Path to a .vec in Word2Vec text foramt or '
'.bin binary fastText model file. ')
group.add_argument('--embedding-name', type=str,
help=('Name of embedding type to load. '
'Valid entries: {}'.format(
', '.join(
nlp.embedding.list_sources().keys()))))
group.add_argument('--embedding-source', type=str,
help=('Source from which to initialize the embedding.'
'Pass --list-embedding-sources to get a list of '
'valid sources for a given --embedding-name.'))
group.add_argument(
'--fasttext-load-ngrams',
action='store_true',
help=('Specify load_ngrams=True '
'when loading pretrained fastText embedding.'))
group.add_argument(
'--analogy-max-vocab-size', type=int, default=None,
help=('Only retain the X first tokens from the pre-trained embedding. '
'The tokens are ordered by decreasing frequency.'
'As the analogy task takes the whole vocabulary into account, '
'removing very infrequent words improves performance.'))
group.add_argument('--list-embedding-sources', action='store_true')
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size', type=int, default=1024,
help='Batch size to use on analogy task. '
'Decrease batch size if evaluation crashes.')
group.add_argument('--gpu', type=int,
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-hybridize', action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir', type=str, default='logs',
help='Directory to store logs.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
validate_args(args)
evaluation.validate_args(args)
return args
def get_args():
"""Construct the argument parser."""
parser = argparse.ArgumentParser(
description='Word embedding evaluation with Gluon.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Embeddings arguments
group = parser.add_argument_group('Embedding arguments')
group.add_argument('path',
type=str,
help='Path to pretrained TokenEmbedding file.')
group.add_argument(
'--max-vocab-size',
type=int,
default=None,
help=('Only retain the X first tokens from the pretrained embedding. '
'The tokens are ordererd by decreasing frequency.'
'As the analogy task takes the whole vocabulary into account, '
'removing very infrequent words improves performance.'))
group.add_argument('--list-embedding-sources', action='store_true')
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size',
type=int,
default=32,
help='Batch size to use on analogy task.'
'Decrease batch size if evaluation crashes.')
group.add_argument('--gpu',
type=int,
nargs='+',
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-hybridize',
action='store_true',
help='Disable hybridization of gluon HybridBlocks '
'used for evaluation.')
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir',
type=str,
default='logs',
help='Directory to store logs.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
evaluation.validate_args(args)
return args
def parse_args():
"""Parse command line arguments."""
parser = argparse.ArgumentParser(
description='Text Classification with FastText',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--input', type=str, help='Input file location')
group.add_argument('--validation',
type=str,
help='Validation file Location ')
group.add_argument('--output',
type=str,
help='Location to save trained model')
group.add_argument('--ngrams',
type=int,
default=1,
help='NGrams used for training')
group.add_argument('--batch-size',
type=int,
default=16,
help='Batch size for training.')
group.add_argument('--epochs', type=int, default=10, help='Epoch limit')
group.add_argument('--gpu',
type=int,
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-hybridize',
action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
# Model
group = parser.add_argument_group('Model arguments')
group.add_argument('--emsize',
type=int,
default=100,
help='Size of embedding vectors.')
# Optimization options
group = parser.add_argument_group('Optimization arguments')
group.add_argument('--optimizer', type=str, default='adam')
group.add_argument('--lr', type=float, default=0.05)
group.add_argument('--batch_size', type=float, default=16)
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
evaluation.validate_args(args)
return args
def parse_args():
"""Parse command line arguments."""
parser = argparse.ArgumentParser(
description='Word embedding training with Gluon.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Data options
group = parser.add_argument_group('Data arguments')
group.add_argument('--data',
type=str,
default='text8',
help='Training dataset.')
group.add_argument('--wiki-root',
type=str,
default='text8',
help='Root under which preprocessed wiki dump.')
group.add_argument('--wiki-language',
type=str,
default='text8',
help='Language of wiki dump.')
group.add_argument('--wiki-date', help='Date of wiki dump.')
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size',
type=int,
default=1024,
help='Batch size for training.')
group.add_argument('--epochs', type=int, default=5, help='Epoch limit')
group.add_argument('--gpu',
type=int,
nargs='+',
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-hybridize',
action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
group.add_argument(
'--no-static-alloc',
action='store_true',
help='Disable static memory allocation for HybridBlocks.')
group.add_argument('--no-sparse-grad',
action='store_true',
help='Disable sparse gradient support.')
# Model
group = parser.add_argument_group('Model arguments')
group.add_argument('--emsize',
type=int,
default=300,
help='Size of embedding vectors.')
group.add_argument('--ngrams', type=int, nargs='+', default=[3, 4, 5, 6])
group.add_argument(
'--ngram-buckets',
type=int,
default=2000000,
help='Size of word_context set of the ngram hash function. '
'Set this to 0 for Word2Vec style training.')
group.add_argument('--model',
type=str,
default='skipgram',
help='SkipGram or CBOW.')
group.add_argument('--window',
type=int,
default=5,
help='Context window size.')
group.add_argument('--negative',
type=int,
default=5,
help='Number of negative samples '
'per source-context word pair.')
group.add_argument('--frequent-token-subsampling',
type=float,
default=1E-4,
help='Frequent token subsampling constant.')
# Optimization options
group = parser.add_argument_group('Optimization arguments')
group.add_argument('--optimizer', type=str, default='adagrad')
group.add_argument('--lr', type=float, default=0.1)
group.add_argument('--seed', type=int, default=1, help='random seed')
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir',
type=str,
default='logs',
help='Directory to store logs.')
group.add_argument('--log-interval', type=int, default=100)
group.add_argument('--eval-interval',
type=int,
help='Evaluate every --eval-interval iterations '
'in addition to at the end of every epoch.')
group.add_argument('--no-eval-analogy',
action='store_true',
help='Don\'t evaluate on the analogy task.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
evaluation.validate_args(args)
random.seed(args.seed)
mx.random.seed(args.seed)
np.random.seed(args.seed)
return args
def parse_args():
"""Parse command line arguments."""
parser = argparse.ArgumentParser(
description='GloVe with GluonNLP',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Data options
group = parser.add_argument_group('Data arguments')
group.add_argument(
'cooccurrences', type=str,
help='Path to cooccurrences.npz containing a sparse (COO) '
'representation of the co-occurrence matrix in numpy archive format. '
'Output of ./cooccur')
group.add_argument('vocab', type=str,
help='Vocabulary indices. Output of vocab_count tool.')
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size', type=int, default=512,
help='Batch size for training.')
group.add_argument('--epochs', type=int, default=50, help='Epoch limit')
group.add_argument(
'--gpu', type=int, nargs='+',
help='Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.')
group.add_argument('--no-hybridize', action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
group.add_argument(
'--no-static-alloc', action='store_true',
help='Disable static memory allocation for HybridBlocks.')
# Model
group = parser.add_argument_group('Model arguments')
group.add_argument('--emsize', type=int, default=300,
help='Size of embedding vectors.')
group.add_argument('--x-max', type=int, default=100)
group.add_argument('--alpha', type=float, default=0.75)
# Optimization options
group = parser.add_argument_group('Optimization arguments')
group.add_argument('--adagrad-eps', type=float, default=1,
help='Initial AdaGrad state value.')
group.add_argument('--lr', type=float, default=0.1, help='Learning rate')
group.add_argument('--seed', type=int, default=1, help='Random seed')
group.add_argument('--dropout', type=float, default=0.15)
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir', type=str, default='logs',
help='Directory to store logs.')
group.add_argument('--log-interval', type=int, default=100)
group.add_argument(
'--eval-interval', type=int,
help='Evaluate every --eval-interval iterations '
'in addition to at the end of every epoch.')
group.add_argument('--no-eval-analogy', action='store_true',
help='Don\'t evaluate on the analogy task.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
evaluation.validate_args(args)
random.seed(args.seed)
mx.random.seed(args.seed)
np.random.seed(args.seed)
return args
def parse_args():
"""Parse command line arguments."""
parser = argparse.ArgumentParser(
description='Word embedding training with Gluon.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Data options
group = parser.add_argument_group('Data arguments')
group.add_argument('--data', type=str, default='text8',
help='Training dataset.')
group.add_argument('--wiki-root', type=str, default='text8',
help='Root under which preprocessed wiki dump.')
group.add_argument('--wiki-language', type=str, default='text8',
help='Language of wiki dump.')
group.add_argument('--wiki-date', help='Date of wiki dump.')
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size', type=int, default=1024,
help='Batch size for training.')
group.add_argument('--epochs', type=int, default=5, help='Epoch limit')
group.add_argument(
'--gpu', type=int, nargs='+',
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-prefetch-batch', action='store_true',
help='Disable multi-threaded nogil batch prefetching.')
group.add_argument('--num-prefetch-epoch', type=int, default=3,
help='Start data pipeline for next N epochs when beginning current epoch.')
group.add_argument('--no-hybridize', action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
# Model
group = parser.add_argument_group('Model arguments')
group.add_argument('--emsize', type=int, default=300,
help='Size of embedding vectors.')
group.add_argument('--ngrams', type=int, nargs='+', default=[3, 4, 5, 6])
group.add_argument(
'--ngram-buckets', type=int, default=2000000,
help='Size of word_context set of the ngram hash function. '
'Set this to 0 for Word2Vec style training.')
group.add_argument('--model', type=str, default='skipgram',
help='SkipGram or CBOW.')
group.add_argument('--window', type=int, default=5,
help='Context window size.')
group.add_argument(
'--negative', type=int, default=5, help='Number of negative samples '
'per source-context word pair.')
group.add_argument('--frequent-token-subsampling', type=float,
default=1E-4,
help='Frequent token subsampling constant.')
group.add_argument(
'--max-vocab-size', type=int,
help='Limit the number of words considered. '
'OOV words will be ignored.')
# Optimization options
group = parser.add_argument_group('Optimization arguments')
group.add_argument('--optimizer', type=str, default='groupadagrad')
group.add_argument('--lr', type=float, default=0.1)
group.add_argument('--seed', type=int, default=1, help='random seed')
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir', type=str, default='logs',
help='Directory to store logs.')
group.add_argument('--log-interval', type=int, default=100)
group.add_argument(
'--eval-interval', type=int,
help='Evaluate every --eval-interval iterations '
'in addition to at the end of every epoch.')
group.add_argument('--no-eval-analogy', action='store_true',
help='Don\'t evaluate on the analogy task.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
evaluation.validate_args(args)
random.seed(args.seed)
mx.random.seed(args.seed)
np.random.seed(args.seed)
return args
def parse_args():
"""Parse command line arguments."""
parser = argparse.ArgumentParser(
description='Word embedding training with Gluon.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Computation options
group = parser.add_argument_group('Computation arguments')
group.add_argument('--batch-size',
type=int,
default=1024,
help='Batch size for training.')
group.add_argument('--epochs', type=int, default=5, help='Epoch limit')
group.add_argument('--gpu',
type=int,
nargs='+',
help=('Number (index) of GPU to run on, e.g. 0. '
'If not specified, uses CPU.'))
group.add_argument('--no-hybridize',
action='store_true',
help='Disable hybridization of gluon HybridBlocks.')
group.add_argument(
'--no-static-alloc',
action='store_true',
help='Disable static memory allocation for HybridBlocks.')
group.add_argument('--no-sparse-grad',
action='store_true',
help='Disable sparse gradient support.')
# Model
group = parser.add_argument_group('Model arguments')
group.add_argument('--emsize',
type=int,
default=300,
help='Size of embedding vectors.')
group.add_argument('--ngrams', type=int, nargs='+', default=[3, 4, 5, 6])
group.add_argument(
'--ngram-buckets',
type=int,
default=500000,
help='Size of word_context set of the ngram hash function.')
group.add_argument('--model',
type=str,
default='skipgram',
help='SkipGram or CBOW.')
group.add_argument('--window',
type=int,
default=5,
help='Context window size.')
group.add_argument('--negative',
type=int,
default=5,
help='Number of negative samples.')
# Optimization options
group = parser.add_argument_group('Optimization arguments')
group.add_argument('--optimizer', type=str, default='adagrad')
group.add_argument('--lr', type=float, default=0.05)
group.add_argument('--optimizer-subwords', type=str, default='adagrad')
group.add_argument('--lr-subwords', type=float, default=0.01)
# Logging
group = parser.add_argument_group('Logging arguments')
group.add_argument('--logdir',
type=str,
default='logs',
help='Directory to store logs.')
group.add_argument('--eval-interval',
type=int,
default=50000,
help='Evaluate every --eval-interval iterations '
'in addition to at the end of every epoch.')
group.add_argument('--no-eval-analogy',
action='store_true',
help='Don\'t evaluate on the analogy task.')
# Evaluation options
evaluation.add_parameters(parser)
args = parser.parse_args()
evaluation.validate_args(args)
return args
