def test_predict(self):
lines = sys.stdin.readlines()
reader = IOBReader(lines)
extractors = []
params = {
'epochs': 100,
'learning_rate': 0.01,
'window_size': 5,
'name_model': 'model_we.ckpt',
'word_embeddings_file': 'data/vectors.txt'
}
classifier = WordEmbeddingsClassifier(reader, extractors, WordEmbeddingsEstimator, **params)
predicted = classifier.predict()
labels_idx_rev = {v:k for k,v in reader.vocabulary[reader.getPosition('LABEL')].items()}
i = 0
for line in lines:
line = line.strip()
if line:
print '%s\t%s\t%s' % (line.split()[0], line.split()[1], labels_idx_rev[predicted[i]])
i += 1
else:
print
def main():
parser = argparse.ArgumentParser(description='Named Entity Recognition with TensorFlow')
subparsers = parser.add_subparsers()
parser_train = subparsers.add_parser('train')
parser_train.set_defaults(which='train')
parser_train.add_argument('-e', '--epochs', help='epochs number', type=int, required=True)
parser_train.add_argument('-l', '--learning-rate', help='learning rate', type=float, required=True)
parser_train.add_argument('-o', '--optimizer', help='optimizer', type=str, required=True, choices=OPTIMIZERS.keys())
parser_tag = subparsers.add_parser('tag')
parser_tag.set_defaults(which='tag')
parser_collect_data = subparsers.add_parser('collect')
parser_collect_data.set_defaults(which='collect')
parser_collect_data.add_argument('-d', '--directory', help='directory', type=str, required=True)
parser_collect_data.add_argument('-i', '--input-file', help='input file', type=str, required=False)
parser_score = subparsers.add_parser('score')
parser_score.set_defaults(which='score')
parser_score.add_argument('-p', '--predicted', help='predicted file', type=str, required=True)
parser_score.add_argument('-g', '--gold', help='gold file', type=str, required=True)
# common arguments
for p in (parser_train, parser_tag):
p.add_argument('-m', '--model', help='model-file', type=str, required=True)
p.add_argument('-r', '--reader-file', help='reader file', type=str, required=True)
p.add_argument('-w', '--word-embeddings', help='word embeddings', type=str, required=False)
p.add_argument('-et', '--word-embeddings-type', help='word embeddings type', type=str, required=False)
p.add_argument('-i', '--input-file', help='input file', type=str, required=False)
p.add_argument('-t', '--type', help='estimator type', type=str, required=True, choices=ESTIMATORS.keys())
p.add_argument('-nl', '--num-layers', help='number layers for multi rnn estimator', type=int, required=False)
p.add_argument('-f', '--feats-conf', help='add the feats in the conf number', type=int, required=False)
args = parser.parse_args()
try:
infile = open(args.input_file) if args.input_file is not None else sys.stdin
except:
pass
print args
if args.which == 'collect':
with infile as f:
for line in f:
domain, agro, categories = line.strip().split('\t')
# TODO: skipping multi-categories
if ',' in categories:
continue
cfile = os.path.join(args.directory, domain[0], domain[1], domain[2], domain, 'content.txt')
try:
content = open(cfile).read()
except:
print >> sys.stderr, '%s not found in %s' % (domain, cfile)
continue
words = ' '.join([word.strip() for word in content.split()])
if words:
print '%s\t%s\t%s' % (domain, categories, words)
elif args.which == 'train':
max_size = 500000
word_embeddings = Word2VecReader(open(args.word_embeddings)).read()
reader = WebContentReader(infile, separator='\t')
examples, labels = reader.read()
label_extractor = LabelExtractor(reader.getPosition('LABEL'), one_hot=False)
X, y = reader.map2idx(examples, labels, [], label_extractor, word_embeddings)
_X = []
for x, _ in X: # we dont want feats from extractors
if len(x) < max_size:
x = np.lib.pad(x, (0, max_size - len(x)), 'constant', constant_values=0)
else:
x = x[:max_size]
_X.append(x)
X = np.array(_X)
y = np.array(y)
d = {}
for yy in y:
try:
d[yy] += 1
except:
d[yy] = 1
print d
X = preprocessing.StandardScaler().fit_transform(X)
classifier = skflow.TensorFlowDNNClassifier(hidden_units=[100, 200, 150, 100], n_classes=len(label_extractor.vocabulary))
#classifier = skflow.TensorFlowLinearRegressor()
#classifier.fit(X, y, logdir='log')
classifier.fit(X, y)
#### test
lines = open('data/fine-it-test').readlines()
reader = WebContentReader(lines, separator='\t')
examples, labels = reader.read()
label_extractor = LabelExtractor(reader.getPosition('LABEL'), one_hot=False)
X, y = reader.map2idx(examples, labels, [], label_extractor, word_embeddings)
_X = []
for x, _ in X: # we dont want feats from extractors
if len(x) < max_size:
x = np.lib.pad(x, (0, max_size - len(x)), 'constant', constant_values=0)
else:
x = x[:max_size]
_X.append(x)
X = np.array(_X)
y = np.array(y)
predicted = classifier.predict(X)
score = metrics.accuracy_score(y, predicted)
labels_idx_rev = {v:k for k,v in reader.vocabulary[reader.getPosition('LABEL')].items()}
i = 0
for line in lines:
line = line.strip()
if line:
print '%s\t%s' % (line.split()[0], labels_idx_rev[predicted[i]])
i += 1
else:
print
print("Accuracy: %f" % score)
elif args.which == 'tag':
lines = sys.stdin.readlines()
reader = reader = WebContentReader(lines, separator='\t')
extractors = []
params = {
'name_model': args.model,
'word_embeddings_file': args.word_embeddings,
'reader_file': args.reader_file,
'num_layers': args.num_layers
}
classifier = WordEmbeddingsClassifier(reader, extractors, ESTIMATORS[args.type], **params)
predicted = classifier.predict()
print >> sys.stderr, 'l predicted', len(predicted), 'l lines', len(lines)
labels_idx_rev = {v:k for k,v in reader.vocabulary[reader.getPosition('LABEL')].items()}
i = 0
for line in lines:
line = line.strip()
if line:
print '%s\t%s' % (line.split()[0], labels_idx_rev[predicted[i]])
i += 1
else:
print
elif args.which == 'score':
gold_dict = {}
for line in open(args.gold):
domain, label = line.strip().split('\t')[:2]
gold_dict[domain] = label
y_true = []
y_pred = []
for line in open(args.predicted):
domain, label = line.strip().split('\t')[:2]
y_pred.append(int(label))
y_true.append(int(gold_dict[domain]))
print f1_score(y_true, y_pred, average='macro')
def main():
parser = argparse.ArgumentParser(description='Named Entity Recognition with TensorFlow')
subparsers = parser.add_subparsers()
parser_train = subparsers.add_parser('train')
parser_train.set_defaults(which='train')
parser_train.add_argument('-e', '--epochs', help='epochs number', type=int, required=True)
parser_train.add_argument('-l', '--learning-rate', help='learning rate', type=float, required=True)
parser_train.add_argument('-o', '--optimizer', help='optimizer', type=str, required=True, choices=OPTIMIZERS.keys())
parser_tag = subparsers.add_parser('tag')
parser_tag.set_defaults(which='tag')
parser_collect_data = subparsers.add_parser('collect')
parser_collect_data.set_defaults(which='collect')
parser_collect_data.add_argument('-d', '--directory', help='directory', type=str, required=True)
parser_collect_data.add_argument('-i', '--input-file', help='input file', type=str, required=False)
parser_score = subparsers.add_parser('score')
parser_score.set_defaults(which='score')
parser_score.add_argument('-p', '--predicted', help='predicted file', type=str, required=True)
parser_score.add_argument('-g', '--gold', help='gold file', type=str, required=True)
# common arguments
for p in (parser_train, parser_tag):
p.add_argument('-m', '--model', help='model-file', type=str, required=True)
p.add_argument('-r', '--reader-file', help='reader file', type=str, required=True)
p.add_argument('-w', '--word-embeddings', help='word embeddings', type=str, required=False)
p.add_argument('-et', '--word-embeddings-type', help='word embeddings type', type=str, required=False)
p.add_argument('-i', '--input-file', help='input file', type=str, required=False)
p.add_argument('-t', '--type', help='estimator type', type=str, required=True, choices=ESTIMATORS.keys())
p.add_argument('-nl', '--num-layers', help='number layers for multi rnn estimator', type=int, required=False)
p.add_argument('-f', '--feats-conf', help='add the feats in the conf number', type=int, required=False)
args = parser.parse_args()
try:
infile = open(args.input_file) if args.input_file is not None else sys.stdin
except:
pass
if args.which == 'collect':
with infile as f:
for line in f:
domain, agro, categories = line.strip().split('\t')
# TODO: skipping multi-categories
if ',' in categories:
continue
cfile = os.path.join(args.directory, domain[0], domain[1], domain[2], domain, 'content.txt')
try:
content = open(cfile).read()
except:
print >> sys.stderr, '%s not found in %s' % (domain, cfile)
continue
words = ' '.join([word.strip() for word in content.split()])
if words:
print '%s\t%s\t%s' % (domain, categories, words)
elif args.which == 'train':
reader = WebContentReader(infile, separator='\t')
params = {
'epochs': args.epochs,
'learning_rate': args.learning_rate,
'name_model': args.model,
'word_embeddings_file': args.word_embeddings,
'reader_file': args.reader_file,
'optimizer': OPTIMIZERS[args.optimizer],
'num_layers': args.num_layers
}
classifier = WordEmbeddingsClassifier(reader, [], ESTIMATORS[args.type], **params)
classifier.train()
elif args.which == 'tag':
lines = sys.stdin.readlines()
reader = reader = WebContentReader(lines, separator='\t')
extractors = []
params = {
'name_model': args.model,
'word_embeddings_file': args.word_embeddings,
'reader_file': args.reader_file,
'num_layers': args.num_layers
}
classifier = WordEmbeddingsClassifier(reader, extractors, ESTIMATORS[args.type], **params)
predicted = classifier.predict()
print >> sys.stderr, 'l predicted', len(predicted), 'l lines', len(lines)
labels_idx_rev = {v:k for k,v in reader.vocabulary[reader.getPosition('LABEL')].items()}
i = 0
for line in lines:
line = line.strip()
if line:
print '%s\t%s' % (line.split()[0], labels_idx_rev[predicted[i]])
i += 1
else:
print
elif args.which == 'score':
gold_dict = {}
for line in open(args.gold):
domain, label = line.strip().split('\t')[:2]
gold_dict[domain] = label
y_true = []
y_pred = []
for line in open(args.predicted):
domain, label = line.strip().split('\t')[:2]
y_pred.append(int(label))
y_true.append(int(gold_dict[domain]))
print f1_score(y_true, y_pred, average='macro')
def main():
parser = argparse.ArgumentParser(description='Named Entity Recognition with TensorFlow')
subparsers = parser.add_subparsers()
parser_train = subparsers.add_parser('train')
parser_train.set_defaults(which='train')
parser_train.add_argument('-e', '--epochs', help='epochs number', type=int, required=True)
parser_train.add_argument('-l', '--learning-rate', help='learning rate', type=float, required=True)
parser_train.add_argument('-o', '--optimizer', help='optimizer', type=str, required=True, choices=OPTIMIZERS.keys())
parser_tag = subparsers.add_parser('tag')
parser_tag.set_defaults(which='tag')
# common arguments
for p in (parser_train, parser_tag):
p.add_argument('-m', '--model', help='model-file', type=str, required=True)
p.add_argument('-r', '--reader-file', help='reader file', type=str, required=True)
p.add_argument('-w', '--word-embeddings', help='word embeddings', type=str, required=False)
p.add_argument('-et', '--word-embeddings-type', help='word embeddings type', type=str, required=False)
p.add_argument('-i', '--input-file', help='input file', type=str, required=False)
p.add_argument('-t', '--type', help='estimator type', type=str, required=True, choices=ESTIMATORS.keys())
p.add_argument('-wi', '--window', help='context window size', type=int, required=True)
p.add_argument('-nl', '--num-layers', help='number layers for multi rnn estimator', type=int, required=False)
p.add_argument('-f', '--feats-conf', help='add the feats in the conf number', type=int, required=False)
args = parser.parse_args()
infile = args.input_file if args.input_file is not None else sys.stdin
if args.which == 'train':
reader = IOBReader(infile, separator='\t')
extractors = []
if args.feats_conf is not None and args.feats_conf != 0:
extractors = [
FieldExtractor(reader.getPosition('FORM')),
FieldExtractor(reader.getPosition('POS')),
CapitalExtractor(reader.getPosition('FORM'))
]
params = {
'epochs': args.epochs,
'learning_rate': args.learning_rate,
'window_size': args.window,
'name_model': args.model,
'word_embeddings_file': args.word_embeddings,
'reader_file': args.reader_file,
'optimizer': OPTIMIZERS[args.optimizer],
'num_layers': args.num_layers
}
classifier = WordEmbeddingsClassifier(reader, extractors, ESTIMATORS[args.type], **params)
classifier.train()
elif args.which == 'tag':
lines = sys.stdin.readlines()
reader = IOBReader(lines)
extractors = []
if args.feats_conf is not None and args.feats_conf != 0:
extractors = [
FieldExtractor(reader.getPosition('FORM')),
FieldExtractor(reader.getPosition('POS')),
CapitalExtractor(reader.getPosition('FORM'))
]
params = {
'window_size': args.window,
'name_model': args.model,
'word_embeddings_file': args.word_embeddings,
'reader_file': args.reader_file,
'num_layers': args.num_layers
}
classifier = WordEmbeddingsClassifier(reader, extractors, ESTIMATORS[args.type], **params)
predicted = classifier.predict()
print >> sys.stderr, len(predicted), len(lines)
labels_idx_rev = {v:k for k,v in reader.vocabulary[reader.getPosition('LABEL')].items()}
i = 0
for line in lines:
line = line.strip()
if line:
print '%s\t%s\t%s' % (line.split()[0], line.split()[1], labels_idx_rev[predicted[i]])
i += 1
else:
print