def init_trainer(self, args):
if args.gpuid:
print('Running with GPU {}.'.format(args.gpuid[0]))
cuda.set_device(args.gpuid[0])
else:
print('Running with CPU.')
if args.random_seed:
random.seed(args.random_seed + os.getpid())
np.random.seed(args.random_seed + os.getpid())
schema = Schema(args.schema_path)
scenario_db = ScenarioDB.from_dict(schema,
read_json(args.scenarios_path),
Scenario)
valid_scenario_db = ScenarioDB.from_dict(
schema, read_json(args.valid_scenarios_path), Scenario)
# if len(args.agent_checkpoints) == 0
# assert len(args.agent_checkpoints) <= len(args.agents)
if len(args.agent_checkpoints) < len(args.agents):
ckpt = [None] * 2
else:
ckpt = args.agent_checkpoints
systems = [
get_system(name, args, schema, False, ckpt[i])
for i, name in enumerate(args.agents)
]
rl_agent = 0
system = systems[rl_agent]
model = system.env.model
loss = None
# optim = build_optim(args, [model, system.env.critic], None)
optim = {
'model': build_optim(args, model, None),
'critic': build_optim(args, system.env.critic, None)
}
optim['critic']._set_rate(0.05)
scenarios = {
'train': scenario_db.scenarios_list,
'dev': valid_scenario_db.scenarios_list
}
from neural.a2c_trainer import RLTrainer as A2CTrainer
trainer = A2CTrainer(systems,
scenarios,
loss,
optim,
rl_agent,
reward_func=args.reward,
cuda=(len(args.gpuid) > 0),
args=args)
self.args = args
self.trainer = trainer
self.systems = systems
def __init__(self, chats, surveys=None, worker_ids=None):
self.chats = []
for f in chats:
self.chats.extend(read_json(f))
self.uuid_to_chat = {chat['uuid']: chat for chat in self.chats}
if surveys:
# This is a list because we might have multiple batches of surveys
self.surveys = [read_json(survey) for survey in surveys]
if worker_ids:
self.worker_ids = {}
for f in worker_ids:
self.worker_ids.update(read_json(f))
else:
self.worker_ids = None
def init(path):
global stats_path, STATS
stats_path = path
try:
STATS = read_json(stats_path)
except Exception:
STATS = {}
def _read_transcripts(self, transcripts_paths, max_examples):
transcripts = []
for transcripts_path in transcripts_paths:
transcripts.extend(read_json(transcripts_path))
if max_examples is not None:
transcripts = transcripts[:max_examples]
return transcripts
def __init__(self, schema, price_tracker, retriever, model_path, mappings, timed_session=False):
super(NeuralRankerSystem, self).__init__()
self.schema = schema
self.price_tracker = price_tracker
self.timed_session = timed_session
# Load arguments
args_path = os.path.join(model_path, 'config.json')
config = read_json(args_path)
# TODO: handle this properly
config['batch_size'] = 1
config['pretrained_wordvec'] = None
args = argparse.Namespace(**config)
mappings_path = os.path.join(mappings, 'vocab.pkl')
mappings = read_pickle(mappings_path)
vocab = mappings['vocab']
logstats.add_args('model_args', args)
model = build_model(schema, mappings, None, args)
# Tensorflow config
if args.gpu == 0:
print 'GPU is disabled'
config = tf.ConfigProto(device_count = {'GPU': 0})
else:
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction = 0.5, allow_growth=True)
config = tf.ConfigProto(device_count = {'GPU': 1}, gpu_options=gpu_options)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
# NOTE: need to close the session when done
tf_session = tf.Session(config=config)
tf_session.run(tf.global_variables_initializer())
# Load TF model parameters
ckpt = tf.train.get_checkpoint_state(model_path+'-best')
assert ckpt, 'No checkpoint found'
assert ckpt.model_checkpoint_path, 'No model path found in checkpoint'
saver = tf.train.Saver()
saver.restore(tf_session, ckpt.model_checkpoint_path)
preprocessor = Preprocessor(schema, price_tracker, 'canonical', 'canonical', 'canonical')
textint_map = TextIntMap(vocab, preprocessor)
int_markers = SpecialSymbols(*[mappings['vocab'].to_ind(m) for m in markers])
model_config = {'retrieve': True}
batcher = DialogueBatcherFactory.get_dialogue_batcher(model_config, int_markers=int_markers, slot_filling=False, kb_pad=mappings['kb_vocab'].to_ind(markers.PAD))
StreamingDialogue.textint_map = textint_map
StreamingDialogue.num_context = args.num_context
StreamingDialogue.mappings = mappings
Env = namedtuple('Env', ['ranker', 'retriever', 'tf_session', 'preprocessor', 'mappings', 'textint_map', 'batcher'])
self.env = Env(model, retriever, tf_session, preprocessor, mappings, textint_map, batcher)
def __init__(self, schema, lexicon, model_path, fact_check, decoding, timed_session=False, consecutive_entity=True, realizer=None):
super(NeuralSystem, self).__init__()
self.schema = schema
self.lexicon = lexicon
self.timed_session = timed_session
self.consecutive_entity = consecutive_entity
# Load arguments
args_path = os.path.join(model_path, 'config.json')
config = read_json(args_path)
config['batch_size'] = 1
config['gpu'] = 0 # Don't need GPU for batch_size=1
config['decoding'] = decoding
args = argparse.Namespace(**config)
mappings_path = os.path.join(model_path, 'vocab.pkl')
mappings = read_pickle(mappings_path)
vocab = mappings['vocab']
# TODO: different models have the same key now
args.dropout = 0
logstats.add_args('model_args', args)
model = build_model(schema, mappings, args)
# Tensorflow config
if args.gpu == 0:
print 'GPU is disabled'
config = tf.ConfigProto(device_count = {'GPU': 0})
else:
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction = 0.5, allow_growth=True)
config = tf.ConfigProto(device_count = {'GPU': 1}, gpu_options=gpu_options)
# NOTE: need to close the session when done
tf_session = tf.Session(config=config)
tf.initialize_all_variables().run(session=tf_session)
# Load TF model parameters
ckpt = tf.train.get_checkpoint_state(model_path+'-best')
assert ckpt, 'No checkpoint found'
assert ckpt.model_checkpoint_path, 'No model path found in checkpoint'
saver = tf.train.Saver()
saver.restore(tf_session, ckpt.model_checkpoint_path)
self.model_name = args.model
if self.model_name == 'attn-copy-encdec':
args.entity_target_form = 'graph'
copy = True
else:
copy = False
preprocessor = Preprocessor(schema, lexicon, args.entity_encoding_form, args.entity_decoding_form, args.entity_target_form, args.prepend)
textint_map = TextIntMap(vocab, mappings['entity'], preprocessor)
Env = namedtuple('Env', ['model', 'tf_session', 'preprocessor', 'vocab', 'copy', 'textint_map', 'stop_symbol', 'remove_symbols', 'max_len', 'evaluator', 'prepend', 'consecutive_entity', 'realizer'])
self.env = Env(model, tf_session, preprocessor, mappings['vocab'], copy, textint_map, stop_symbol=vocab.to_ind(markers.EOS), remove_symbols=map(vocab.to_ind, (markers.EOS, markers.PAD)), max_len=20, evaluator=FactEvaluator() if fact_check else None, prepend=args.prepend, consecutive_entity=self.consecutive_entity, realizer=realizer)
def get_evaluated_qids(cls, db_path):
db = read_json(db_path)
qids = set()
for hit_id, hit_info in db.iteritems():
for assignment_id, result in hit_info.iteritems():
answers = result['answers']
for answer in answers:
if answer['qid'] == 'comment':
continue
qids.add(answer['qid'])
return qids
def init(path, verbose=False):
global stats_path, STATS
stats_path = path
try:
STATS = read_json(stats_path)
if verbose:
print("Stats file loaded from {}".format(stats_path))
except Exception:
STATS = {}
if verbose:
print("New stats file created, will be stored in {}".format(
stats_path))
def read_examples(paths, max_examples, Scenario):
'''
Read a maximum of |max_examples| examples from |paths|.
'''
examples = []
for path in paths:
print('read_examples: %s' % path)
for raw in read_json(path):
if max_examples >= 0 and len(examples) >= max_examples:
break
examples.append(Example.from_dict(raw, Scenario))
return examples
def __init__(self,
slot_scores_path=None,
stop_words_path=None,
threshold=4.):
if not stop_words_path:
self.stopwords = set(stopwords.words('english'))
else:
with open(stop_words_path, 'r') as fin:
self.stopwords = set(fin.read().split()[:200])
self.stopwords.update(
['.', '...', ',', '?', '!', '"', "n't", "'m", "'d", "'ll"])
if slot_scores_path:
self.slot_scores = read_json(slot_scores_path)
self.threshold = threshold
self.stemmer = PorterStemmer()
def load_candidates(self, paths):
candidates = defaultdict(list)
# When dumped to json, NamedTuple becomes list. Now convert it back.
is_str = lambda x: isinstance(x, basestring)
# x[0] (surface of entity): note that for prices from the offer action,
# surface is float instead of string
to_ent = lambda x: x.encode('utf-8') if is_str(x) else \
Entity(x[0].encode('utf-8') if is_str(x[0]) else x[0], CanonicalEntity(*x[1]))
for path in paths:
print 'Load candidates from', path
results = read_json(path)
for r in results:
# None for encoding turns
if r['candidates'] is None:
candidates[(r['uuid'], r['role'])].append(None)
else:
# Only take the response (list of tokens)
candidates_ = [[to_ent(x) for x in c['response']]
for c in ifilter(lambda x: 'response' in x,
r['candidates'])]
candidates[(r['uuid'], r['role'])].append(candidates_)
return candidates
def read_system_responses(cls, system, path, num_context_utterances, data):
"""Read system responses and update the database.
Args:
system (str): system name
path (str): a JSON file containing system outputs.
|-[]
|-ex_id (str): unique id that identifies a context-reference pair
|-prev_turns (list)
|-reference
|-response
num_context_utterances (int)
data (dict): see from_file.
"""
examples = read_json(path)
for ex in examples:
if not cls.valid_example(ex, num_context_utterances):
continue
qid = ex['ex_id']
context_turns = ex['prev_turns'][-1*num_context_utterances:]
agent_names = cls.get_agent_name(context_turns + [ex['reference']])
context = []
for i, u in enumerate(context_turns):
u = cls.process_utterance(u, role=agent_names[i])
if len(u[1]) > 0:
context.append(u)
reference = cls.process_utterance(ex['reference'], role=agent_names[-1])
response = cls.process_utterance(ex['response'], role=agent_names[-1])
if not (len(reference) and len(response) and len(context)):
continue
if qid not in data:
data[qid] = {
'context': context,
'responses': {}
}
assert system not in data[qid]['responses']
data[qid]['responses'][system] = response
data[qid]['responses']['reference'] = reference
def get_data_generator(args, model_args, mappings, schema):
from preprocess import DataGenerator, Preprocessor
from cocoa.core.scenario_db import ScenarioDB
from cocoa.core.mutualfriends.lexicon import Lexicon
from cocoa.core.dataset import read_dataset
from cocoa.core.util import read_json
import time
scenario_db = ScenarioDB.from_dict(schema, read_json(args.scenarios_path))
dataset = read_dataset(scenario_db, args)
print 'Building lexicon...'
start = time.time()
lexicon = Lexicon(schema, args.learned_lex, stop_words=args.stop_words)
print '%.2f s' % (time.time() - start)
# Dataset
use_kb = False if model_args.model == 'encdec' else True
copy = True if model_args.model == 'attn-copy-encdec' else False
if model_args.model == 'attn-copy-encdec':
model_args.entity_target_form = 'graph'
preprocessor = Preprocessor(schema, lexicon,
model_args.entity_encoding_form,
model_args.entity_decoding_form,
model_args.entity_target_form,
model_args.prepend)
if args.test:
model_args.dropout = 0
data_generator = DataGenerator(None, None, dataset.test_examples,
preprocessor, schema,
model_args.num_items, mappings, use_kb,
copy)
else:
data_generator = DataGenerator(dataset.train_examples,
dataset.test_examples, None,
preprocessor, schema,
model_args.num_items, mappings, use_kb,
copy)
return data_generator
default=10,
help='Number of questions repeated in each HIT for estimating agreement'
)
parser.add_argument('--compare',
action='store_true',
help='Rate by comparing two responses')
parser.add_argument('--title',
default='Dialogue response evaluation',
help='Title of the HIT')
add_turk_task_arguments(parser)
add_eval_data_arguments(parser)
args = parser.parse_args()
random.seed(1)
config = read_json(args.aws_config)
mtc = get_mturk_connection(config, debug=args.debug)
if args.debug:
lifetime = 0.1
args.hit_db = 'test.json'
else:
lifetime = 30
if args.compare:
Task = HTMLCompareEvalTask
else:
Task = HTMLEvalTask
task = Task(
mtc=mtc,
title=args.title,
description='Rate a response in a dialogue',
raise ValueError(
"Location of HTML templates should be specified in config with the key templates_dir"
)
if not os.path.exists(templates_dir):
raise ValueError("Specified HTML template location doesn't exist: %s" %
templates_dir)
app = create_app(debug=False, templates_dir=templates_dir)
schema_path = args.schema_path
if not os.path.exists(schema_path):
raise ValueError("No schema file found at %s" % schema_path)
schema = Schema(schema_path)
scenarios = read_json(args.scenarios_path)
if args.num_scenarios is not None:
scenarios = scenarios[:args.num_scenarios]
scenario_db = ScenarioDB.from_dict(schema, scenarios, Scenario)
app.config['scenario_db'] = scenario_db
if 'models' not in params.keys():
params['models'] = {}
if 'quit_after' not in params.keys():
params['quit_after'] = params['status_params']['chat'][
'num_seconds'] + 500
if 'skip_chat_enabled' not in params.keys():
params['skip_chat_enabled'] = False
default=[],
help='Input test examples')
parser.add_argument('--train-max-examples',
type=int,
help='Maximum number of training examples')
parser.add_argument('--test-max-examples',
type=int,
help='Maximum number of test examples')
parser.add_argument('--eval-examples-paths',
nargs='*',
default=[],
help='Path to multi-response evaluation files')
def read_dataset(args, Scenario):
'''
Return the dataset specified by the given args.
'''
train_examples = read_examples(args.train_examples_paths,
args.train_max_examples, Scenario)
test_examples = read_examples(args.test_examples_paths,
args.test_max_examples, Scenario)
dataset = Dataset(train_examples, test_examples)
return dataset
if __name__ == "__main__":
raw = read_json("fb-negotiation/data/transformed_test.json")
for idx, example in enumerate(raw):
print(Example.test_dict(example))
raise ValueError(
"Location of HTML templates should be specified in config with the key templates_dir"
)
if not os.path.exists(templates_dir):
raise ValueError("Specified HTML template location doesn't exist: %s" %
templates_dir)
app = create_app(debug=False, templates_dir=templates_dir)
schema_path = args.schema_path
if not os.path.exists(schema_path):
raise ValueError("No schema file found at %s" % schema_path)
schema = Schema(schema_path)
scenarios = read_json(args.scenarios_path)
if args.num_scenarios is not None:
scenarios = scenarios[:args.num_scenarios]
scenario_db = ScenarioDB.from_dict(schema, scenarios, Scenario)
app.config['scenario_db'] = scenario_db
if 'models' not in params.keys():
params['models'] = {}
if 'quit_after' not in params.keys():
params[
'quit_after'] = params['status_params']['chat']['num_seconds'] + 1
if 'skip_chat_enabled' not in params.keys():
params['skip_chat_enabled'] = False
help='Transciprts paths',
nargs='*',
default=[])
parser.add_argument('--train-frac',
help='Fraction of training examples',
type=float,
default=0.6)
parser.add_argument('--test-frac',
help='Fraction of test examples',
type=float,
default=0.2)
parser.add_argument('--dev-frac',
help='Fraction of dev examples',
type=float,
default=0.2)
parser.add_argument('--output-path', help='Output path for splits')
args = parser.parse_args()
np.random.seed(0)
json_data = ([], [], [])
for path in args.example_paths:
examples = read_json(path)
folds = np.random.choice(
3, len(examples), p=[args.train_frac, args.dev_frac, args.test_frac])
for ex, fold in izip(examples, folds):
json_data[fold].append(ex)
for fold, dataset in izip(('train', 'dev', 'test'), json_data):
if len(dataset) > 0:
write_json(dataset, '%s%s.json' % (args.output_path, fold))
required=True,
help="Fractions of data from different categories")
parser.add_argument('--skip',
default=0,
type=int,
help="Skip the first X scenarios")
cocoa.options.add_scenario_arguments(parser)
args = parser.parse_args()
random.seed(args.random_seed)
np.random.seed(args.random_seed)
schema = Schema(args.schema_path)
listings = [
read_json(
os.path.join(args.scraped_data, 'craigslist_{}.json'.format(c)))
for c in args.categories
]
fractions = np.array([float(x) for x in args.fractions])
fractions = fractions / np.sum(fractions)
# Sample listings
sampled_listings = []
N = sum([len(l) for l in listings])
for listing, fraction in izip(listings, fractions):
n = int(N * fraction)
print listing[0]['category'], len(listing), fraction, n
sampled_listings.append(listing[:n])
listings = [x for l in sampled_listings for x in l]
N = len(listings)
inds = np.random.permutation(N)
def _read_surveys(self, survey_paths):
dialogue_scores = {}
for path in survey_paths:
dialogue_scores.update(read_json(path)[1])
return dialogue_scores
tooltip = mpld3.plugins.PointHTMLTooltip(points,
labels=labels,
voffset=10,
hoffset=10,
css=self.css)
mpld3.plugins.connect(fig, tooltip)
fig_dict = mpld3.fig_to_dict(fig)
plt.close()
return fig_dict
####### TEST #########
if __name__ == '__main__':
from cocoa.core.negotiation.price_tracker import PriceTracker
from cocoa.core.util import read_json
from liwc import LIWC
transcripts = read_json('web_output/combined/transcripts/transcripts.json')
price_tracker = PriceTracker('/scr/hehe/game-dialogue/price_tracker.pkl')
liwc = LIWC.from_pkl('data/liwc.pkl')
dialogue = Dialogue.from_dict(transcripts[0], price_tracker)
dialogue.label_liwc(liwc)
for u in dialogue.iter_utterances():
print(u.text)
print(u.categories)
#dialogue.extract_keywords()
#dialogue.label_speech_acts()
#dialogue.label_stage()
#dialogue.fig_dict()
self).__init__(timed_session,
configs,
policy=policy,
max_chats_per_config=max_chats_per_config,
db=db)
self.lexicon = lexicon
self.templates = templates
def _new_session(self, agent, kb, config):
config = Config(*config)
return RulebasedSession.get_session(agent, kb, self.lexicon, config,
self.templates)
############# TEST ##############
if __name__ == '__main__':
from cocoa.core.util import read_json
configs = read_json('data/rulebased_configs.json')
configs = [tuple(c) for c in configs]
s = BaseConfigurableRulebasedSystem(False, configs)
#print s.configs
#print s.choose_config()
#print s.choose_config()
print s.trials
#s.update_trials([(configs[0], '0', {'margin': 0.1, 'humanlike': 1})])
#s.update_trials([(configs[0], '1', {'margin': 0.1, 'humanlike': 1})])
s.update_trials([(configs[1], '3', {'margin': 0.1})])
#s.update_trials([(configs[1], '2', {'humanlike': 0.1})])
print s.trials
def scenarios(schema):
scenarios_path = 'data/negotiation/craigslist-scenarios.json'
scenario_db = ScenarioDB.from_dict(schema, read_json(scenarios_path))
return scenario_db
def __init__(self,
schema,
price_tracker,
model_path,
mappings_path,
decoding,
index=None,
num_candidates=20,
retriever_context_len=2,
timed_session=False):
super(NeuralSystem, self).__init__()
self.schema = schema
self.price_tracker = price_tracker
self.timed_session = timed_session
# Load arguments
args_path = os.path.join(model_path, 'config.json')
config = read_json(args_path)
config['batch_size'] = 1
config['gpu'] = 0 # Don't need GPU for batch_size=1
config['decoding'] = decoding
config['pretrained_wordvec'] = None
args = argparse.Namespace(**config)
vocab_path = os.path.join(mappings_path, 'vocab.pkl')
mappings = read_pickle(vocab_path)
vocab = mappings['vocab']
# TODO: different models have the same key now
args.dropout = 0
logstats.add_args('model_args', args)
model = build_model(schema, mappings, None, args)
# Tensorflow config
if args.gpu == 0:
print 'GPU is disabled'
config = tf.ConfigProto(device_count={'GPU': 0})
else:
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5,
allow_growth=True)
config = tf.ConfigProto(device_count={'GPU': 1},
gpu_options=gpu_options)
# NOTE: need to close the session when done
tf_session = tf.Session(config=config)
tf.initialize_all_variables().run(session=tf_session)
# Load TF model parameters
ckpt = tf.train.get_checkpoint_state(model_path + '-best')
assert ckpt, 'No checkpoint found'
assert ckpt.model_checkpoint_path, 'No model path found in checkpoint'
saver = tf.train.Saver()
saver.restore(tf_session, ckpt.model_checkpoint_path)
# Model config tells data generator which batcher to use
model_config = {}
if args.retrieve or args.model in ('ir', 'selector'):
model_config['retrieve'] = True
if args.predict_price:
model_config['price'] = True
self.model_name = args.model
preprocessor = Preprocessor(schema, price_tracker,
args.entity_encoding_form,
args.entity_decoding_form,
args.entity_target_form)
textint_map = TextIntMap(vocab, preprocessor)
int_markers = SpecialSymbols(
*[mappings['vocab'].to_ind(m) for m in markers])
dialogue_batcher = DialogueBatcherFactory.get_dialogue_batcher(
model_config,
int_markers=int_markers,
slot_filling=False,
kb_pad=mappings['kb_vocab'].to_ind(markers.PAD))
# Retriever
if args.model == 'selector':
retriever = Retriever(index,
context_size=retriever_context_len,
num_candidates=num_candidates)
else:
retriever = None
#TODO: class variable is not a good way to do this
Dialogue.mappings = mappings
Dialogue.textint_map = textint_map
Dialogue.preprocessor = preprocessor
Dialogue.num_context = args.num_context
Env = namedtuple('Env', [
'model', 'tf_session', 'preprocessor', 'vocab', 'textint_map',
'stop_symbol', 'remove_symbols', 'max_len', 'dialogue_batcher',
'retriever'
])
self.env = Env(model,
tf_session,
preprocessor,
mappings['vocab'],
textint_map,
stop_symbol=vocab.to_ind(markers.EOS),
remove_symbols=map(vocab.to_ind,
(markers.EOS, markers.PAD)),
max_len=20,
dialogue_batcher=dialogue_batcher,
retriever=retriever)
add_model_arguments(parser)
add_learner_arguments(parser)
args = parser.parse_args()
random.seed(args.random_seed)
if not os.path.isdir(os.path.dirname(args.stats_file)):
os.makedirs(os.path.dirname(args.stats_file))
logstats.init(args.stats_file)
logstats.add_args('config', args)
# Save or load models
if args.init_from:
start = time.time()
print 'Load model (config, vocab, checkpoint) from', args.init_from
config_path = os.path.join(args.init_from, 'config.json')
saved_config = read_json(config_path)
# NOTE: args below can be overwritten
# TODO: separate temperature from decoding arg
saved_config['decoding'] = args.decoding
saved_config['temperature'] = args.temperature
saved_config['batch_size'] = args.batch_size
saved_config['pretrained_wordvec'] = args.pretrained_wordvec
saved_config['ranker'] = args.ranker
model_args = argparse.Namespace(**saved_config)
# Checkpoint
if args.test and args.best:
ckpt = tf.train.get_checkpoint_state(args.init_from + '-best')
else:
def from_json(cls, data_path):
"""Construct from dumped data.
"""
data = read_json(data_path)
return cls(data)
import argparse
from cocoa.core.util import read_json, write_json
from cocoa.core.scenario_db import ScenarioDB
from cocoa.core.schema import Schema
from core.scenario import Scenario
parser = argparse.ArgumentParser()
parser.add_argument('--chats')
parser.add_argument('--scenarios')
parser.add_argument('--max', type=int)
args = parser.parse_args()
chats = read_json(args.chats)
scenarios = []
n = args.max or len(chats)
for chat in chats[:n]:
scenarios.append(Scenario.from_dict(None, chat['scenario']))
scenario_db = ScenarioDB(scenarios)
write_json(scenario_db.to_dict(), args.scenarios)
action='store_true',
help='Whether or not to have verbose prints')
parser.add_argument('--valid-scenarios-path',
help='Output path for the validation scenarios')
cocoa.options.add_scenario_arguments(parser)
options.add_system_arguments(parser)
options.add_rl_arguments(parser)
options.add_model_arguments(parser)
args = parser.parse_args()
if args.random_seed:
random.seed(args.random_seed)
np.random.seed(args.random_seed)
schema = Schema(args.schema_path)
scenario_db = ScenarioDB.from_dict(schema, read_json(args.scenarios_path),
Scenario)
valid_scenario_db = ScenarioDB.from_dict(
schema, read_json(args.valid_scenarios_path), Scenario)
assert len(args.agent_checkpoints) <= len(args.agents)
systems = [
get_system(name, args, schema, False, args.agent_checkpoints[i])
for i, name in enumerate(args.agents)
]
rl_agent = 0
system = systems[rl_agent]
model = system.env.model
loss = make_loss(args, model, system.mappings['tgt_vocab'])
optim = build_optim(args, model, None)
'''
examples: json chats
Use "$xxx$ as ground truth, and record n-gram context before and after the price.
'''
context = {'left': defaultdict(int), 'right': defaultdict(int)}
for ex in examples:
for event in ex['events']:
if event['action'] == 'message':
tokens = tokenize(event['data'])
tokens = ['<s>'] + tokens + ['</s>']
for i, token in enumerate(tokens):
if token[0] == '$' or token[-1] == '$':
context['left'][tokens[i - 1]] += 1
context['right'][tokens[i + 1]] += 1
if output_path:
write_pickle(context, output_path)
return context
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--train-examples-path',
help='Path to training json file')
parser.add_argument('--output', help='Path to output model')
args = parser.parse_args()
examples = read_json(args.train_examples_path)
PriceTracker.train(examples, args.output)
def get_data_generator(args, model_args, mappings, schema):
from cocoa.core.scenario_db import ScenarioDB
from cocoa.core.dataset import read_dataset, EvalExample
from cocoa.core.util import read_json
from core.scenario import Scenario
from core.price_tracker import PriceTracker
from core.slot_detector import SlotDetector
from retriever import Retriever
from preprocess import DataGenerator, LMDataGenerator, EvalDataGenerator, Preprocessor
import os.path
# TODO: move this to dataset
if args.eval:
dataset = []
for path in args.eval_examples_paths:
dataset.extend(
[EvalExample.from_dict(schema, e) for e in read_json(path)])
else:
dataset = read_dataset(args, Scenario)
lexicon = PriceTracker(model_args.price_tracker_model)
slot_detector = SlotDetector(slot_scores_path=model_args.slot_scores)
# Model config tells data generator which batcher to use
model_config = {}
if args.retrieve or model_args.model in ('ir', 'selector'):
model_config['retrieve'] = True
if args.predict_price:
model_config['price'] = True
# For retrieval-based models only: whether to add ground truth response in the candidates
if model_args.model in ('selector', 'ir'):
if 'loss' in args.eval_modes and 'generation' in args.eval_modes:
print '"loss" requires ground truth reponse to be added to the candidate set. Please evaluate "loss" and "generation" separately.'
raise ValueError
if (not args.test) or args.eval_modes == ['loss']:
add_ground_truth = True
else:
add_ground_truth = False
print 'Ground truth response {} be added to the candidate set.'.format(
'will' if add_ground_truth else 'will not')
else:
add_ground_truth = False
# TODO: hacky
if args.model == 'lm':
DataGenerator = LMDataGenerator
if args.retrieve or args.model in ('selector', 'ir'):
retriever = Retriever(args.index,
context_size=args.retriever_context_len,
num_candidates=args.num_candidates)
else:
retriever = None
preprocessor = Preprocessor(schema,
lexicon,
model_args.entity_encoding_form,
model_args.entity_decoding_form,
model_args.entity_target_form,
slot_filling=model_args.slot_filling,
slot_detector=slot_detector)
trie_path = os.path.join(model_args.mappings, 'trie.pkl')
if args.eval:
data_generator = EvalDataGenerator(dataset, preprocessor, mappings,
model_args.num_context)
else:
if args.test:
model_args.dropout = 0
train, dev, test = None, None, dataset.test_examples
else:
train, dev, test = dataset.train_examples, dataset.test_examples, None
data_generator = DataGenerator(train,
dev,
test,
preprocessor,
schema,
mappings,
retriever=retriever,
cache=args.cache,
ignore_cache=args.ignore_cache,
candidates_path=args.candidates_path,
num_context=model_args.num_context,
trie_path=trie_path,
batch_size=args.batch_size,
model_config=model_config,
add_ground_truth=add_ground_truth)
return data_generator
