def __init__(self):
config = flags.FLAGS
config.out_dir = os.path.join(config.out_base_dir, config.model_name,
str(config.run_id).zfill(2))
config.max_sent_size = config.sent_size_th
config.max_num_sents = config.num_sents_th
config.max_ques_size = config.ques_size_th
config.max_word_size = config.word_size_th
config.max_para_size = config.para_size_th
self.config = config
self.test_data = None
self.data_ready(update=True)
config = self.config
set_dirs(config)
models = get_multi_gpu_models(config)
self.evaluator = ForwardEvaluator(
config,
models[0],
tensor_dict=models[0].tensor_dict if config.vis else None)
self.sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
self.graph_handler = GraphHandler(config, models[0])
self.graph_handler.initialize(self.sess)
self.config = config
def _forward(config):
assert config.load
test_data = read_data(config, config.forward_name, True)
update_config(config, [test_data])
_config_debug(config)
if config.use_glove_for_unk:
word2vec_dict = test_data.shared['lower_word2vec'] if config.lower_word else test_data.shared['word2vec']
new_word2idx_dict = test_data.shared['new_word2idx']
idx2vec_dict = {idx: word2vec_dict[word] for word, idx in new_word2idx_dict.items()}
new_emb_mat = np.array([idx2vec_dict[idx] for idx in range(len(idx2vec_dict))], dtype='float32')
config.new_emb_mat = new_emb_mat
pprint(config.__flags, indent=2)
models = get_multi_gpu_models(config)
model = models[0]
evaluator = ForwardEvaluator(config, model)
graph_handler = GraphHandler(config, model) # controls all tensors and variables in the graph, including loading /saving
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
graph_handler.initialize(sess)
num_batches = math.ceil(test_data.num_examples / config.batch_size)
if 0 < config.test_num_batches < num_batches:
num_batches = config.test_num_batches
e = evaluator.get_evaluation_from_batches(sess, tqdm(test_data.get_batches(config.batch_size, num_batches=num_batches), total=num_batches))
print(e)
if config.dump_answer:
print("dumping answer ...")
graph_handler.dump_answer(e, path=config.answer_path)
if config.dump_eval:
print("dumping eval ...")
graph_handler.dump_eval(e, path=config.eval_path)
class Demo(object):
def __init__(self):
config = flags.FLAGS
config.out_dir = os.path.join(config.out_base_dir, config.model_name, str(config.run_id).zfill(2))
config.max_sent_size = config.sent_size_th
config.max_num_sents = config.num_sents_th
config.max_ques_size = config.ques_size_th
config.max_word_size = config.word_size_th
config.max_para_size = config.para_size_th
self.config = config
self.test_data = None
self.data_ready(update=True)
config = self.config
set_dirs(config)
models = get_multi_gpu_models(config)
self.evaluator = ForwardEvaluator(config, models[0], tensor_dict=models[0].tensor_dict if config.vis else None)
self.sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
self.graph_handler = GraphHandler(config, models[0])
self.graph_handler.initialize(self.sess)
self.config = config
def data_ready(self, data=None, update=False):
config = self.config
config.batch_size = 1
test_data = read_data(self.config, 'demo', True, data=data, data_set=self.test_data)
if update:
update_config(self.config, [test_data])
if config.use_glove_for_unk:
word2vec_dict = test_data.shared['lower_word2vec'] if config.lower_word else test_data.shared['word2vec']
new_word2idx_dict = test_data.shared['new_word2idx']
idx2vec_dict = {idx: word2vec_dict[word] for word, idx in new_word2idx_dict.items()}
new_emb_mat = np.array([idx2vec_dict[idx] for idx in range(len(idx2vec_dict))], dtype='float32')
config.new_emb_mat = new_emb_mat
self.config = config
self.test_data = test_data
def run(self, data):
self.data_ready(data=data)
test_data = self.test_data
config = self.config
e = None
for multi_batch in test_data.get_batches(config.batch_size, num_batches=1, cluster=config.cluster):
ei = self.evaluator.get_evaluation(self.sess, multi_batch)
e = ei if e is None else e + ei
return (e.id2answer_dict[0])
def _server(config):
import bottle
# Pre-load model
assert config.load
server_update_config(config)
pprint(config.__flags, indent=2)
models = get_multi_gpu_models(config)
model = models[0]
evaluator = ForwardEvaluator(config, model)
graph_handler = GraphHandler(
config, model
) # controls all tensors and variables in the graph, including loading /saving
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
graph_handler.initialize(sess)
# Pre-load all GloVe vectors
glove_path = 'glove/glove.6B.100d.txt' # hard code GloVe file
num_lines = 400000
glove_dict = {}
print('Reading all GloVe vectors from %s' % glove_path)
with open(glove_path, 'r', encoding='utf-8') as fh:
for line in tqdm(fh, total=num_lines):
array = line.lstrip().rstrip().split(" ")
word = array[0]
vector = list(map(float, array[1:]))
glove_dict[word] = vector
# Create the app
orig_data_dir = config.data_dir
app = bottle.Bottle()
@app.route('/query', method='post')
def query():
with tempfile.TemporaryDirectory(dir=orig_data_dir) as inter_dir:
# Recieve data, process it
data = bottle.request.json
config.data_dir = inter_dir
with tempfile.NamedTemporaryFile('w',
suffix='.json',
dir=orig_data_dir) as data_file:
json.dump(data, data_file)
data_file.flush()
prepro_args = prepro.get_args([
'--mode', 'single', '--single_path', data_file.name, '-pm',
'--target_dir', inter_dir
])
prepro.prepro(prepro_args, glove_dict=glove_dict)
test_data = read_data(config, config.forward_name, True)
num_batches = math.ceil(test_data.num_examples / config.batch_size)
if 0 < config.eval_num_batches < num_batches:
num_batches = config.eval_num_batches
# Run model on data
e = evaluator.get_evaluation_from_batches(
sess,
tqdm(test_data.get_batches(config.batch_size,
num_batches=num_batches),
total=num_batches))
eval_path = os.path.join(inter_dir, 'eval.pkl.gz')
graph_handler.dump_eval(e, path=eval_path)
# Extract predictions through the ensemble code
data_path = os.path.join(inter_dir, 'data_single.json')
with open(data_path) as f:
data_single_obj = json.load(f)
shared_path = os.path.join(inter_dir, 'shared_single.json')
with open(shared_path) as f:
shared_single_obj = json.load(f)
with tempfile.NamedTemporaryFile('w',
suffix='.json',
dir=orig_data_dir) as target_file:
target_path = target_file.name
ensemble_args = ensemble.get_args([
'--data_path', data_path, '--shared_path', shared_path,
'-o', target_path, eval_path
])
ensemble.ensemble(ensemble_args)
target_file.flush()
with open(target_path, 'r') as f:
pred_obj = json.load(f)
return {
'data_single': data_single_obj,
'eval': e.dict,
'shared_single': shared_single_obj,
'predictions': pred_obj
}
# Run the app
bottle.run(app, host='localhost', port=SERVER_PORT)