def __init__(self, x_max_lengths=None, y_max_length=8):
self.name = 'wikipedia'
self.x_max_length = max(x_max_lengths)
self.x_max_lengths = x_max_lengths
self.y_max_length = y_max_length
# Vocabulary as defined in two dictionaries
# self.vocabulary | token -> idx
# self.reversed_vocabulary | idx -> token
self.vocabulary = self.read_vocabulary()
self.vocab_size = len(self.vocabulary)
self.reversed_vocabulary = dict(
zip(self.vocabulary.values(), self.vocabulary.keys()))
self.test = WikipediaTestData(self,
path=directory('/data/wikipedia_data/') +
'test.txt',
x_max_length=self.x_max_length,
y_max_length=self.y_max_length)
self.train = WikipediaBucketedTrainData(
self,
paths=[
directory('/data/wikipedia_data/') + 'train-33.txt',
directory('/data/wikipedia_data/') + 'train-62.txt',
directory('/data/wikipedia_data/') + 'train-118.txt',
directory('/data/wikipedia_data/') + 'train-504.txt'
],
x_max_lengths=self.x_max_lengths,
y_max_length=self.y_max_length)
def setupMainWindow(self):
self.setWindowTitle("Classifier")
self.documents_directory = utils.directory(config[KEY_DOCUMENTS])
self.documents = self.documents_directory.get_all_files()
self.categories_directory = utils.directory(config[KEY_CATEGORIES])
self.categories = self.categories_directory.get_all_directories()
QtGui.QShortcut(QtGui.QKeySequence("Ctrl+Z"), self, self.undoAction)
def wikipedia(n, should_filter_pretrained=False):
d = WikipediaDataset(x_max_lengths=[33, 62, 118, 504], y_max_length=8)
ep = EmbeddingProcessor(d.vocabulary,
path=directory('/data/compvec_wikipedia/'))
d.store_dataset(path=directory('/data/compvec_wikipedia/'))
if should_filter_pretrained:
filter_pretrained(ep)
runs(n, d, ep, neural=True, random=False)
def single_wordnet(n, should_filter_pretrained=False):
d = WordnetDataset(x_max_length=32, y_max_length=1)
ep = EmbeddingProcessor(d.vocabulary,
path=directory('/data/compvec_wordnet_single/'))
d.store_dataset(path=directory('/data/compvec_wordnet_single/'))
if should_filter_pretrained:
filter_pretrained(ep)
runs(n, d, ep, yc=False, neural=False, random=False)
runs(n, d, ep, yc=True, neural=False, random=False)
def multi_wordnet(n, should_filter_pretrained=False):
d = WordnetDataset(
test_data_path=directory('/data/compvec_wordnet_single/') +
'test_data.gz',
x_max_length=32,
y_max_length=6)
ep = EmbeddingProcessor(d.vocabulary,
path=directory('/data/compvec_wordnet_multi/'))
d.store_dataset(path=directory('/data/compvec_wordnet_multi/'))
if should_filter_pretrained:
filter_pretrained(ep)
runs(n, d, ep, neural=True, random=True)
def __init__(self,
run_group_name,
dataset,
embedding_processor,
yc=True,
pretraining=None,
learning_rate=1e-3,
batch_size=512,
embedding_size=300,
stop_gradients_y_n=False,
dropout_keep_p=0.75,
margin=0.25,
composition='sum',
loss='mse',
refine_after_x_steps=0,
no=None):
self.dataset = dataset
self.embedding_processor = embedding_processor
self.y_composition = yc
self.no = no
self.learning_rate = learning_rate
self.batch_size = batch_size
self.dropout_keep_p = dropout_keep_p
self.pretraining = pretraining
self.refine_after_x_steps = refine_after_x_steps
self.run_dir = directory('/out/run-%s' % run_group_name,
['logs', 'tsne', 'embeddings', 'output'])
self.data_dir = embedding_processor.path
self.x_max_buckets = None
if hasattr(self.dataset, 'x_max_lengths'):
self.x_max_buckets = self.dataset.x_max_lengths
self.graph = tf.Graph()
with self.graph.as_default():
# Model
self.model = Model(embedding_size=embedding_size,
x_max_buckets=self.x_max_buckets,
x_max_length=self.dataset.x_max_length,
y_max_length=self.dataset.y_max_length,
y_composition=self.y_composition,
vocab_size=self.dataset.vocab_size,
margin=margin,
composition=composition,
loss=loss,
stop_gradients_y_n=stop_gradients_y_n)
# Evaluator
self.evaluation = Evaluation(self.model, self.dataset)
# Assign Tensorflow Operations
self.assign_ops()
# Setup Writers for Tensorboard
self.setup_writers()
def deploy_file(path, kwargs, config):
with utils.directory(os.path.dirname(path)):
config.update(kwargs)
if 'FunctionName' not in config:
clip.exit('You must provide a function name', err=True)
# Zip up directory
utils.make_zip(config['FunctionName'])
# Upload!
upload(config['FunctionName'], config)
def deploy_dir(path, kwargs):
with utils.directory(path):
config = LambdaConfig().load_from_cwd().update_config(kwargs)
config.verify()
# Remove ignore paths
for e in config.get('ignore', []) + ['.git/', '.gitignore']:
utils.delete_resource(e)
# Run install command
if 'install' in config:
utils.shell(config.get('install'))
upload(config.get_config())
def deploy_dir(path, kwargs):
with utils.directory(path):
config = LambdaConfig().load_from_cwd().update_config(kwargs)
config.verify()
# Remove ignore paths
for e in config.get('ignore', []) + ['.git/', '.gitignore']:
utils.delete_resource(e)
# Run install command
if 'install' in config:
utils.shell(config.get('install'))
upload(config.get_config())
def favicon():
'''
Serves GET favicon requests
Favicons are the icons that show up in your browser's tab.
For the purposes of this project, a favicon is not needed, so a blank one is used.
Note:
This function is required because browsers will try to access '{base_url}/favicon.ico'
and without this function, the request will automatically go to the date_page() function
where things will get messed up.
'''
return send_from_directory(utils.directory(APP.root_path), 'favicon.ico')
def __init__(self, model, dataset):
self.m = model
self.d = dataset
path = directory('/data/wordnetsingle/') + 'test_data.gz'
self.compveceval_test = WordnetData.from_path(path, self.d.vocabulary,
self.d.x_max_length,
self.d.y_max_length)
self.compveceval = CompVecEvalEvaluation(self.m, self.compveceval_test)
self.senteval = SentEvalEvaluation(self.m, self.d)
self.wordsim = WordSimEvaluation(self.m, self.d)
def __init__(self, vocabulary=None, path=None):
if vocabulary is None:
self.vocabulary = dict()
else:
self.vocabulary = vocabulary
if path is None:
self.path = directory('/data/compositional_wordnet')
else:
self.path = path
self.reversed_vocabulary = dict(
zip(self.vocabulary.values(), self.vocabulary.keys()))
def run(tasks, dir=None, silent=False):
if dir is None:
dir = os.getcwd()
okapi.silent = silent
with utils.directory(dir):
config = utils.load_config()
okapi.log('Running tasks on {}'.format(config.project if hasattr(
config, 'project') else os.path.basename(dir)))
setattr(config, 'ok', okapi)
for task in tasks:
if not hasattr(config, task):
okapi.log('"{}" not a valid task, skipping!'.format(task))
continue
okapi.run(getattr(config, task))
okapi.log('All tasks complete!')
def deploy_dir(path, kwargs):
with utils.directory(path):
config = utils.load_config()
config['config'].update(kwargs)
if 'FunctionName' not in config['config']:
clip.exit('You must provide a function name', err=True)
# Remove ignore paths
for e in config['ignore'] + ['.git/', '.gitignore']:
utils.delete_resource(e)
# Run install command
if 'install' in config:
utils.shell(config['install'])
# Zip up directory
utils.make_zip(config['config']['FunctionName'])
# Upload!
params = config['config']
upload(params['FunctionName'], params)
def read_fallback_embeddings(vocabulary, pretrain=None, evaluation='MR'):
original_embedding_path = original_embedding_file(pretrain)
if original_embedding_path is None:
return dict()
ped = EmbeddingProcessor(vocabulary)
processed_embeddings_path = '%s/%s-%s.vec.gz' % (directory('/data/senteval_embeddings'), pretrain, evaluation)
if not os.path.isfile(processed_embeddings_path):
ped.process_pretrained_embeddings(input_filename=original_embedding_path,
output_filename=processed_embeddings_path)
embeddings = ped.read_embeddings(processed_embeddings_path)
return embeddings
def read_vocabulary(self,
path=None,
vocab_min_frequency=18,
vocab_size_limit=5171164):
"""
Read vocabulary from vocab file.
"""
vocab_ns = dict()
if path is None:
path = directory('/data/wikipedia_data/') + 'vocab.txt'
with open(path) as f:
for i, line in enumerate(f):
if i % 10000 == 0:
sys.stdout.write("\rReading vocabulary… %6.2f%%" %
((100 * i) / float(VOCAB_SIZE), ))
s = line.split()
if len(s) != 2:
continue
t, n = s[0], int(s[1])
if n >= vocab_min_frequency and i < vocab_size_limit - 2:
vocab_ns[t] = n
else:
break # vocabulary file is already sorted
print("\rVocabulary read %d" % (len(vocab_ns) + 2))
# Sort the vocabulary by frequency, frequent words on top
vocab_ns = sorted(vocab_ns.items(),
key=operator.itemgetter(1),
reverse=True)
vocabulary = {token: i for i, (token, count) in enumerate(vocab_ns, 2)}
vocabulary['PAD'] = PAD_SYMBOL
vocabulary['UNK'] = UNK_SYMBOL
return vocabulary
parser.add_argument('data_dir')
parser.add_argument('-a', '--annotator', default='default')
parser.add_argument('-i', '--interface', default='interface')
parser.add_argument('--past_reports', default=False, action='store_true')
parser.add_argument('-m', '--models', action='append')
parser.add_argument('-d', '--device', default='cpu')
parser.add_argument('-r', '--reload', default=False, action='store_true')
parser.add_argument('-p', '--port')
args = parser.parse_args()
if args.data_dir is None:
raise NotImplementedError
startup['annotations_dir'] = join(args.data_dir,
args.annotator + '_annotations')
if not exists(startup['annotations_dir']):
mkdir(startup['annotations_dir'])
np.random.seed(0)
startup['file_generator'] = FileGenerator(args.data_dir,
startup['annotations_dir'],
reload=args.reload)
try:
startup['file'] = next(startup['file_generator'])
except StopIteration:
startup['file'] = None
with directory(args.interface_dir):
exec('import ' + args.interface)
models_to_load = args.models if args.models is not None else []
startup['interface'] = eval(args.interface).FullModelInterface(
models_to_load=models_to_load, device=args.device)
startup['include_past_reports'] = args.past_reports
app.run(debug=True, port=args.port)
def root(self, path):
with utils.directory(path):
yield
def deploy_file(path, kwargs):
with utils.directory(os.path.dirname(path)):
config = LambdaConfig().load_from_front_matter(path).update_config(
kwargs)
config.verify()
upload(config.get_config())
def deploy_file(path, kwargs): with utils.directory(os.path.dirname(path)): config = LambdaConfig().load_from_front_matter(path).update_config(kwargs) config.verify() upload(config.get_config())
