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, use_gpu=False):
# make args that are supposed to be passed in by command line arguments
args = {
"model": "lf2lf",
"word_vec_size": 300,
"dropout": 0.,
"encoder_type": "rnn",
"decoder_type": "rnn",
"context_embedder_type": "mean",
"global_attention": "multibank_general",
"share_embeddings": False,
"share_decoder_embeddings": False,
"enc_layers": 1,
"copy_attn": False,
"dec_layers": 1,
"pretrained_wordvec": "",
"rnn_size": 300,
"rnn_type": "LSTM",
"enc_layers": 1,
"num_context": 2,
"stateful": True,
"sample": True,
"max_length": 10,
"n_best": 1,
"batch_size": 128,
"optim": "adagrad",
"alpha": 0.01,
"temperature": 0.5,
"epochs": 30,
"report_every": 500,
}
if use_gpu:
args.gpuid = 0
# HACK: convert args from dict into object. Ex. args["epochs"]
# becomes args.epochs
args = type("args", (), args)
# load price tracker
with open(self.PRICE_TRACKER_PATH) as f:
price_tracker = pickle.load(f)
# load schema
schema = Schema(self.SCHEMA_PATH)
# load system
self.system = PytorchNeuralSystem(args, schema, price_tracker,
self.MODEL_PATH, False)
# load scenario db
with open(self.DATA_PATH) as f:
raw = json.load(f)
raw = [r["scenario"] for r in raw] # HACK
self.scenario_db = ScenarioDB.from_dict(schema, raw, Scenario)
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
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)
scenario_list = []
price_unit = 1 # Just use the real price
scenario_generator = generate_scenario(schema, base_price, price_unit,
args.discounts, listings)
for i, s in enumerate(scenario_generator):
if i % 100 == 0:
print i
if i < args.skip:
continue
if len(scenario_list) == args.num_scenarios:
break
scenario_list.append(s)
if len(scenario_list) < args.num_scenarios:
print 'Not enough listings: {} scenarios generated.'.format(
len(scenario_list))
scenario_db = ScenarioDB(scenario_list)
write_json(scenario_db.to_dict(), args.scenarios_path)
for i in range(min(10, len(scenario_db.scenarios_list))):
print '---------------------------------------------------------------------------------------------'
print '---------------------------------------------------------------------------------------------'
scenario = scenario_db.scenarios_list[i]
print "Scenario id: %s" % scenario.uuid
for agent in (0, 1):
kb = scenario.kbs[agent]
kb.dump()
num_listings_per_category = defaultdict(int)
for s in scenario_list:
cat = s.kbs[0].facts['item']['Category']
num_listings_per_category[cat] += 1
def scenarios(schema):
scenarios_path = 'data/negotiation/craigslist-scenarios.json'
scenario_db = ScenarioDB.from_dict(schema, read_json(scenarios_path))
return scenario_db