parser.add_argument('--transformer',
help='Are we using a Transformer (default is LSTM) LM',
type=str2bool,
default=False)
parser.add_argument("--device",
type=str,
default="cuda" if torch.cuda.is_available() else "cpu",
help="Device (cuda or cpu)")
args = parser.parse_known_args()[0]
embed_type = 'learned-positional' if args.transformer else 'default'
feature_desc = {
'word': {
'vectorizer':
baseline.Token1DVectorizer(mxlen=-1, transform_fn=baseline.lowercase),
'embed': {
'embed_file': args.embeddings,
'embed_type': embed_type,
'unif': 0.05
}
}
}
class DictionaryDatasetWrapper(Dataset):
def __init__(self, x, y):
self.tensor_dataset = TensorDataset(x, y)
def __getitem__(self, index):
# stuff
pool_field = 'cmotsz' if args.model_type == 'default' else 'rnnsz'
model_params = {
'model_type': args.model_type,
'filtsz': args.filts,
pool_field: args.poolsz
}
if args.stacksz is not None:
model_params['hsz'] = args.stacksz
feature_desc = {
'word': {
'vectorizer': bl.Token1DVectorizer(mxlen=args.mxlen, transform_fn=bl.lowercase),
'embed': {'file': args.embeddings, 'type': 'default', 'unif': 0.25}
}
}
# Create a reader that is using our vectorizers to parse a TSV file
# with rows like:
# <label>\t<sentence>\n
vectorizers = {k: v['vectorizer'] for k, v in feature_desc.items()}
reader = bl.TSVSeqLabelReader(vectorizers,
clean_fn=bl.TSVSeqLabelReader.do_clean)
train_file = args.train
valid_file = args.valid
test_file = args.test
parser.add_argument('--mxlen', help='Maximum post length (number of words) during training', type=int, default=100)
parser.add_argument('--train', help='Training file', default='../data/stsa.binary.phrases.train')
parser.add_argument('--valid', help='Validation file', default='../data/stsa.binary.dev')
parser.add_argument('--test', help='Testing file', default='../data/stsa.binary.test')
parser.add_argument('--embeddings', help='Pretrained embeddings file', default='/data/embeddings/GoogleNews-vectors-negative300.bin')
parser.add_argument('--ll', help='Log level', type=str, default='info')
parser.add_argument('--lr', help='Learning rate', type=float, default=0.001)
parser.add_argument("--device", type=str,
default="cuda" if torch.cuda.is_available() else "cpu",
help="Device (cuda or cpu)")
args = parser.parse_known_args()[0]
feature_desc = {
'word': {
'vectorizer': baseline.Token1DVectorizer(mxlen=100, transform_fn=baseline.lowercase),
'embed': {'file': args.embeddings, 'type': 'default', 'unif': 0.25}
}
}
# Create a reader that is using our vectorizers to parse a TSV file
# with rows like:
# <label>\t<sentence>\n
class DictionaryDatasetWrapper(Dataset):
def __init__(self, x, x_lengths, y):
self.tensor_dataset = TensorDataset(x, x_lengths, y)
def __getitem__(self, index):
# stuff
x, x_length, y = self.tensor_dataset[index]
return {'word': x.to(args.device), "lengths": x_length.to(args.device)}, y.to(args.device)
import baseline as bl
import argparse
import os
import numpy as np
"""Take a file of TSVs where format is `label<\t>content` and convert to NPZ file of vectors from pretrained embeddings
"""
BP = '../data'
TRAIN = 'stsa.binary.phrases.train'.format(BP)
VALID = 'stsa.binary.dev'
TEST = 'stsa.binary.test'
LABELS = os.path.join(BP, 'stsa.binary.labels')
W2V_GN_300 = '/data/embeddings/GoogleNews-vectors-negative300.bin'
VECTORIZERS = {'word': bl.Token1DVectorizer(mxlen=40)}
def output_file(input_file):
return input_file + '.npz'
def convert_input(file, embeddings, batchsz=50):
batch_x = []
batch_y = []
dsz = embeddings.get_dsz()
ts = reader.load(file, vocabs={'word': embeddings.vocab}, batchsz=batchsz)
pg = bl.create_progress_bar(len(ts))
for batch in pg(ts):
x = batch['word']
B, T = x.shape
flat_x = x.reshape(B * T)