def __init__(self):
super(TradeDST, self).__init__()
# load config
common_config_path = os.path.join(get_config_path(),
TradeDST.common_config_name)
common_config = json.load(open(common_config_path))
model_config_path = os.path.join(get_config_path(),
TradeDST.model_config_name)
model_config = json.load(open(model_config_path))
model_config.update(common_config)
self.model_config = model_config
self.model_config['data_path'] = os.path.join(
get_data_path(), 'crosswoz/dst_trade_data')
self.model_config['n_gpus'] = 0 if self.model_config[
'device'] == 'cpu' else torch.cuda.device_count()
self.model_config['device'] = torch.device(self.model_config['device'])
if model_config['load_embedding']:
model_config['hidden_size'] = 300
# download data
for model_key, url in TradeDST.model_urls.items():
dst = os.path.join(self.model_config['data_path'], model_key)
if model_key.endswith('pth'):
file_name = 'trained_model_path'
elif model_key.endswith('pkl'):
file_name = model_key.rsplit('-', maxsplit=1)[0]
else:
file_name = model_key.split('.')[0] # ontology
self.model_config[file_name] = dst
if not os.path.exists(dst) or not self.model_config['use_cache']:
download_from_url(url, dst)
# load date & model
ontology = json.load(
open(self.model_config['ontology'], 'r', encoding='utf8'))
self.all_slots = get_slot_information(ontology)
self.gate2id = {'ptr': 0, 'none': 1}
self.id2gate = {id_: gate for gate, id_ in self.gate2id.items()}
self.lang = pickle.load(open(self.model_config['lang'], 'rb'))
self.mem_lang = pickle.load(open(self.model_config['mem-lang'], 'rb'))
model = Trade(
lang=self.lang,
vocab_size=len(self.lang.index2word),
hidden_size=self.model_config['hidden_size'],
dropout=self.model_config['dropout'],
num_encoder_layers=self.model_config['num_encoder_layers'],
num_decoder_layers=self.model_config['num_decoder_layers'],
pad_id=self.model_config['pad_id'],
slots=self.all_slots,
num_gates=len(self.gate2id),
unk_mask=self.model_config['unk_mask'])
model.load_state_dict(
torch.load(self.model_config['trained_model_path']))
self.model = model.to(self.model_config['device']).eval()
print(f'>>> {self.model_config["trained_model_path"]} loaded ...')
self.state = default_state()
print('>>> State initialized ...')
def load_config() -> dict:
"""Load config from common config and inference config from xbot/config/dst/bert .
Returns:
config dict
"""
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(),
BertDST.common_config_name)
infer_config_path = os.path.join(get_config_path(),
BertDST.infer_config_name)
common_config = json.load(open(common_config_path))
infer_config = json.load(open(infer_config_path))
infer_config.update(common_config)
infer_config['device'] = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
infer_config['data_path'] = os.path.join(get_data_path(),
'crosswoz/dst_bert_data')
infer_config['output_dir'] = os.path.join(root_path,
infer_config['output_dir'])
if not os.path.exists(infer_config['data_path']):
os.makedirs(infer_config['data_path'])
if not os.path.exists(infer_config['output_dir']):
os.makedirs(infer_config['output_dir'])
return infer_config
def main():
model_config_name = 'policy/mle/train.json'
common_config_name = 'policy/mle/common.json'
data_urls = {
'sys_da_voc.json': 'http://qiw2jpwfc.hn-bkt.clouddn.com/usr_da_voc.json',
'usr_da_voc.json': 'http://qiw2jpwfc.hn-bkt.clouddn.com/usr_da_voc.json'
}
# load config
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
model_config_path = os.path.join(get_config_path(), model_config_name)
common_config = json.load(open(common_config_path))
model_config = json.load(open(model_config_path))
model_config.update(common_config)
model_config['n_gpus'] = torch.cuda.device_count()
model_config['batch_size'] = max(1, model_config['n_gpus']) * model_config['batch_size']
model_config['device'] = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model_config['data_path'] = os.path.join(get_data_path(), 'crosswoz/policy_mle_data')
model_config['raw_data_path'] = os.path.join(get_data_path(), 'crosswoz/raw')
model_config['output_dir'] = os.path.join(root_path, model_config['output_dir'])
if model_config['load_model_name']:
model_config['model_path'] = os.path.join(model_config['output_dir'], model_config['load_model_name'])
else:
model_config['model_path'] = ''
if not os.path.exists(model_config['data_path']):
os.makedirs(model_config['data_path'])
if not os.path.exists(model_config['output_dir']):
os.makedirs(model_config['output_dir'])
# download data
for data_key, url in data_urls.items():
dst = os.path.join(model_config['data_path'], data_key)
file_name = data_key.split('.')[0]
model_config[file_name] = dst
if not os.path.exists(dst):
download_from_url(url, dst)
print(f'>>> Train configs:')
print('\t', model_config)
set_seed(model_config['random_seed'])
agent = Trainer(model_config)
# 训练
if model_config['do_train']:
start_epoch = 0 if not model_config['model_path'] else int(model_config['model_path'].split('-')[2]) + 1
best = float('inf')
for epoch in tqdm(range(start_epoch, model_config['num_epochs']), desc='Epoch'):
agent.imitating(epoch)
best = agent.imit_eval(epoch, best)
agent.calc_metrics()
def main():
model_config_name = 'dst/bert/train.json'
common_config_name = 'dst/bert/common.json'
data_urls = {
'train4bert_dst.json': 'http://xbot.bslience.cn/train4bert_dst.json',
'dev4bert_dst.json': 'http://xbot.bslience.cn/dev4bert_dst.json',
'test4bert_dst.json': 'http://xbot.bslience.cn/test4bert_dst.json',
'cleaned_ontology.json':
'http://xbot.bslience.cn/cleaned_ontology.json',
'config.json': 'http://xbot.bslience.cn/bert-base-chinese/config.json',
'pytorch_model.bin':
'http://xbot.bslience.cn/bert-base-chinese/pytorch_model.bin',
'vocab.txt': 'http://xbot.bslience.cn/bert-base-chinese/vocab.txt'
}
# load config
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
train_config_path = os.path.join(get_config_path(), model_config_name)
common_config = json.load(open(common_config_path))
train_config = json.load(open(train_config_path))
train_config.update(common_config)
train_config['n_gpus'] = torch.cuda.device_count()
train_config['train_batch_size'] = max(
1, train_config['n_gpus']) * train_config['train_batch_size']
train_config['device'] = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
train_config['data_path'] = os.path.join(get_data_path(),
'crosswoz/dst_bert_data')
train_config['output_dir'] = os.path.join(root_path,
train_config['output_dir'])
if not os.path.exists(train_config['data_path']):
os.makedirs(train_config['data_path'])
if not os.path.exists(train_config['output_dir']):
os.makedirs(train_config['output_dir'])
# download data
for data_key, url in data_urls.items():
dst = os.path.join(train_config['data_path'], data_key)
file_name = data_key.split('.')[0]
train_config[file_name] = dst
if not os.path.exists(dst):
download_from_url(url, dst)
# train
trainer = Trainer(train_config)
trainer.train()
trainer.eval_test()
get_recall(train_config['data_path'])
def main():
model_config_name = "dst/bert/train.json"
common_config_name = "dst/bert/common.json"
data_urls = {
"train4bert_dst.json": "http://xbot.bslience.cn/train4bert_dst.json",
"dev4bert_dst.json": "http://xbot.bslience.cn/dev4bert_dst.json",
"test4bert_dst.json": "http://xbot.bslience.cn/test4bert_dst.json",
"cleaned_ontology.json": "http://xbot.bslience.cn/cleaned_ontology.json",
"config.json": "http://xbot.bslience.cn/bert-base-chinese/config.json",
"pytorch_model.bin": "http://xbot.bslience.cn/bert-base-chinese/pytorch_model.bin",
"vocab.txt": "http://xbot.bslience.cn/bert-base-chinese/vocab.txt",
}
# load config
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
train_config_path = os.path.join(get_config_path(), model_config_name)
common_config = json.load(open(common_config_path))
train_config = json.load(open(train_config_path))
train_config.update(common_config)
train_config["n_gpus"] = torch.cuda.device_count()
train_config["train_batch_size"] = (
max(1, train_config["n_gpus"]) * train_config["train_batch_size"]
)
train_config["device"] = torch.device(
"cuda" if torch.cuda.is_available() else "cpu"
)
train_config["data_path"] = os.path.join(get_data_path(), "crosswoz/dst_bert_data")
train_config["output_dir"] = os.path.join(root_path, train_config["output_dir"])
if not os.path.exists(train_config["data_path"]):
os.makedirs(train_config["data_path"])
if not os.path.exists(train_config["output_dir"]):
os.makedirs(train_config["output_dir"])
# download data
for data_key, url in data_urls.items():
dst = os.path.join(train_config["data_path"], data_key)
file_name = data_key.split(".")[0]
train_config[file_name] = dst
if not os.path.exists(dst):
download_from_url(url, dst)
# train
trainer = Trainer(train_config)
trainer.train()
trainer.eval_test()
get_recall(train_config["data_path"])
def __init__(self):
# path
root_path = get_root_path()
config_file = os.path.join(get_config_path(), IntentWithBertPredictor.default_model_config)
# load config
config = json.load(open(config_file))
device = config['DEVICE']
# load intent vocabulary and dataloader
intent_vocab = json.load(open(os.path.join(get_data_path(),
'crosswoz/nlu_intent_data/intent_vocab.json'),
encoding='utf-8'))
dataloader = Dataloader(intent_vocab=intent_vocab,
pretrained_weights=config['model']['pretrained_weights'])
# load best model
best_model_path = os.path.join(DEFAULT_MODEL_PATH, IntentWithBertPredictor.default_model_name)
if not os.path.exists(best_model_path):
download_from_url(IntentWithBertPredictor.default_model_url,
best_model_path)
model = IntentWithBert(config['model'], device, dataloader.intent_dim)
try:
model.load_state_dict(torch.load(os.path.join(DEFAULT_MODEL_PATH,
IntentWithBertPredictor.default_model_name),
map_location='cpu'))
except Exception as e:
print(e)
# cpu process
model.to("cpu")
model.eval()
self.model = model
self.dataloader = dataloader
print(f"{best_model_path} loaded - {best_model_path}")
def update_config(common_config_name, train_config_name, task_path):
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
train_config_path = os.path.join(get_config_path(), train_config_name)
common_config = json.load(open(common_config_path))
train_config = json.load(open(train_config_path))
train_config.update(common_config)
train_config["n_gpus"] = torch.cuda.device_count()
train_config["train_batch_size"] = (max(1, train_config["n_gpus"]) *
train_config["train_batch_size"])
train_config["device"] = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
train_config["data_path"] = os.path.join(get_data_path(), task_path)
train_config["output_dir"] = os.path.join(root_path,
train_config["output_dir"])
if not os.path.exists(train_config["data_path"]):
os.makedirs(train_config["data_path"])
if not os.path.exists(train_config["output_dir"]):
os.makedirs(train_config["output_dir"])
return train_config
def load_config() -> dict:
"""Load config for inference.
Returns:
config dict
"""
common_config_path = os.path.join(get_config_path(),
BertPolicy.common_config_name)
infer_config_path = os.path.join(get_config_path(),
BertPolicy.inference_config_name)
common_config = load_json(common_config_path)
infer_config = load_json(infer_config_path)
infer_config.update(common_config)
infer_config["device"] = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
infer_config["data_path"] = os.path.join(get_data_path(),
"crosswoz/policy_bert_data")
if not os.path.exists(infer_config["data_path"]):
os.makedirs(infer_config["data_path"])
return infer_config
def __init__(self):
super(MLEPolicy, self).__init__()
# load config
common_config_path = os.path.join(get_config_path(),
MLEPolicy.common_config_name)
common_config = json.load(open(common_config_path))
model_config_path = os.path.join(get_config_path(),
MLEPolicy.model_config_name)
model_config = json.load(open(model_config_path))
model_config.update(common_config)
self.model_config = model_config
self.model_config["data_path"] = os.path.join(
get_data_path(), "crosswoz/policy_mle_data")
self.model_config["n_gpus"] = (0 if self.model_config["device"]
== "cpu" else torch.cuda.device_count())
self.model_config["device"] = torch.device(self.model_config["device"])
# download data
for model_key, url in MLEPolicy.model_urls.items():
dst = os.path.join(self.model_config["data_path"], model_key)
file_name = (model_key.split(".")[0]
if not model_key.endswith("pth") else
"trained_model_path")
self.model_config[file_name] = dst
if not os.path.exists(dst) or not self.model_config["use_cache"]:
download_from_url(url, dst)
self.vector = CrossWozVector(
sys_da_voc_json=self.model_config["sys_da_voc"],
usr_da_voc_json=self.model_config["usr_da_voc"],
)
policy = MultiDiscretePolicy(self.vector.state_dim,
model_config["hidden_size"],
self.vector.sys_da_dim)
policy.load_state_dict(
torch.load(self.model_config["trained_model_path"]))
self.policy = policy.to(self.model_config["device"]).eval()
print(f'>>> {self.model_config["trained_model_path"]} loaded ...')
def load_config() -> dict:
"""Load config from common config and inference config from src/xbot/config/dst/bert .
Returns:
config dict
"""
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), BertDST.common_config_name)
infer_config_path = os.path.join(get_config_path(), BertDST.infer_config_name)
common_config = json.load(open(common_config_path))
infer_config = json.load(open(infer_config_path))
infer_config.update(common_config)
infer_config["device"] = torch.device(
"cuda" if torch.cuda.is_available() else "cpu"
)
infer_config["data_path"] = os.path.join(
get_data_path(), "crosswoz/dst_bert_data"
)
infer_config["output_dir"] = os.path.join(root_path, infer_config["output_dir"])
if not os.path.exists(infer_config["data_path"]):
os.makedirs(infer_config["data_path"])
if not os.path.exists(infer_config["output_dir"]):
os.makedirs(infer_config["output_dir"])
return infer_config
def __init__(self):
# path
root_path = get_root_path()
config_file = os.path.join(
get_config_path(), IntentWithBertPredictor.default_model_config)
# load config
config = json.load(open(config_file))
self.device = config["DEVICE"]
# load intent vocabulary and dataloader
intent_vocab = json.load(
open(
os.path.join(get_data_path(),
"crosswoz/nlu_intent_data/intent_vocab.json"),
encoding="utf-8",
))
dataloader = Dataloader(
intent_vocab=intent_vocab,
pretrained_weights=config["model"]["pretrained_weights"],
)
# load best model
best_model_path = os.path.join(
os.path.join(root_path, DEFAULT_MODEL_PATH),
IntentWithBertPredictor.default_model_name,
)
# best_model_path = os.path.join(DEFAULT_MODEL_PATH, IntentWithBertPredictor.default_model_name)
if not os.path.exists(best_model_path):
download_from_url(IntentWithBertPredictor.default_model_url,
best_model_path)
model = IntentWithBert(config["model"], self.device,
dataloader.intent_dim)
model.load_state_dict(
torch.load(best_model_path, map_location=self.device))
model.to(self.device)
model.eval()
self.model = model
self.dataloader = dataloader
print(f"{best_model_path} loaded - {best_model_path}")
np.random.seed(seed)
torch.manual_seed(seed)
if __name__ == '__main__':
data_urls = {
'intent_train_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_train_data.json',
'intent_val_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_val_data.json',
'intent_test_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_test_data.json'
}
# load config
root_path = get_root_path()
config_path = os.path.join(os.path.join(get_config_path(), 'nlu'),
'crosswoz_all_context_nlu_intent.json')
config = json.load(open(config_path))
data_path = os.path.join(get_data_path(), 'crosswoz/nlu_intent_data/')
output_dir = config['output_dir']
output_dir = os.path.join(root_path, output_dir)
log_dir = config['log_dir']
log_dir = os.path.join(root_path, log_dir)
device = config['DEVICE']
# download data
for data_key, url in data_urls.items():
dst = os.path.join(os.path.join(data_path, data_key))
if not os.path.exists(dst):
download_from_url(url, dst)
torch.manual_seed(seed)
if __name__ == '__main__':
data_urls = {
'intent_train_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_train_data.json',
'intent_val_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_val_data.json',
'intent_test_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_test_data.json'
}
# path
root_path = get_root_path()
config_file = os.path.join(get_config_path(),
IntentWithBertPredictor.default_model_config)
config = json.load(open(config_file))
data_dir = os.path.join(get_data_path(), 'crosswoz/nlu_intent_data/')
output_dir = config['output_dir']
output_dir = os.path.join(root_path, output_dir)
log_dir = config['log_dir']
log_dir = os.path.join(root_path, log_dir)
device = config['DEVICE']
# download data
for data_key, url in data_urls.items():
dst = os.path.join(os.path.join(data_dir, data_key))
if not os.path.exists(dst):
download_from_url(url, dst)
np.random.seed(seed)
torch.manual_seed(seed)
if __name__ == "__main__":
data_urls = {
"intent_train_data.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/intent_train_data.json",
"intent_val_data.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/intent_val_data.json",
"intent_test_data.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/intent_test_data.json",
}
# load config
root_path = get_root_path()
config_path = os.path.join(os.path.join(get_config_path(), "nlu"),
"crosswoz_all_context_nlu_intent.json")
config = json.load(open(config_path))
data_path = os.path.join(get_data_path(), "crosswoz/nlu_intent_data/")
output_dir = config["output_dir"]
output_dir = os.path.join(root_path, output_dir)
log_dir = config["log_dir"]
log_dir = os.path.join(root_path, log_dir)
device = config["DEVICE"]
# download data
for data_key, url in data_urls.items():
dst = os.path.join(os.path.join(data_path, data_key))
if not os.path.exists(dst):
download_from_url(url, dst)
def __init__(self):
super(TradeDST, self).__init__()
# load config
common_config_path = os.path.join(get_config_path(),
TradeDST.common_config_name)
common_config = json.load(open(common_config_path))
model_config_path = os.path.join(get_config_path(),
TradeDST.model_config_name)
model_config = json.load(open(model_config_path))
model_config.update(common_config)
self.model_config = model_config
self.model_config["data_path"] = os.path.join(
get_data_path(), "crosswoz/dst_trade_data")
self.model_config["n_gpus"] = (0 if self.model_config["device"]
== "cpu" else torch.cuda.device_count())
self.model_config["device"] = torch.device(self.model_config["device"])
if model_config["load_embedding"]:
model_config["hidden_size"] = 300
# download data
for model_key, url in TradeDST.model_urls.items():
dst = os.path.join(self.model_config["data_path"], model_key)
if model_key.endswith("pth"):
file_name = "trained_model_path"
elif model_key.endswith("pkl"):
file_name = model_key.rsplit("-", maxsplit=1)[0]
else:
file_name = model_key.split(".")[0] # ontology
self.model_config[file_name] = dst
if not os.path.exists(dst) or not self.model_config["use_cache"]:
download_from_url(url, dst)
# load date & model
ontology = json.load(
open(self.model_config["ontology"], "r", encoding="utf8"))
self.all_slots = get_slot_information(ontology)
self.gate2id = {"ptr": 0, "none": 1}
self.id2gate = {id_: gate for gate, id_ in self.gate2id.items()}
self.lang = pickle.load(open(self.model_config["lang"], "rb"))
self.mem_lang = pickle.load(open(self.model_config["mem-lang"], "rb"))
model = Trade(
lang=self.lang,
vocab_size=len(self.lang.index2word),
hidden_size=self.model_config["hidden_size"],
dropout=self.model_config["dropout"],
num_encoder_layers=self.model_config["num_encoder_layers"],
num_decoder_layers=self.model_config["num_decoder_layers"],
pad_id=self.model_config["pad_id"],
slots=self.all_slots,
num_gates=len(self.gate2id),
unk_mask=self.model_config["unk_mask"],
)
model.load_state_dict(
torch.load(self.model_config["trained_model_path"]))
self.model = model.to(self.model_config["device"]).eval()
print(f'>>> {self.model_config["trained_model_path"]} loaded ...')
self.state = default_state()
print(">>> State initialized ...")
def main():
model_config_name = 'dst/trade/train.json'
common_config_name = 'dst/trade/common.json'
data_urls = {
'train_dials.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/train_dials.json',
'dev_dials.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/dev_dials.json',
'test_dials.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/test_dials.json',
'ontology.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/ontology.json',
'sgns.wiki.bigram.bz2':
'http://qiw2jpwfc.hn-bkt.clouddn.com/sgns.wiki.bigram.bz2'
}
# load config
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
model_config_path = os.path.join(get_config_path(), model_config_name)
common_config = json.load(open(common_config_path))
model_config = json.load(open(model_config_path))
model_config.update(common_config)
model_config['n_gpus'] = torch.cuda.device_count()
model_config['batch_size'] = max(
1, model_config['n_gpus']) * model_config['batch_size']
model_config['device'] = torch.device(
'cuda:0' if torch.cuda.is_available() else 'cpu')
if model_config['load_embedding']:
model_config['hidden_size'] = 300
model_config['data_path'] = os.path.join(get_data_path(),
'crosswoz/dst_trade_data')
model_config['output_dir'] = os.path.join(
root_path, model_config['output_dir']) # 可以用来保存模型文件
if model_config['load_model_name']:
model_config['model_path'] = os.path.join(
model_config['output_dir'], model_config['load_model_name'])
else:
model_config['model_path'] = ''
if not os.path.exists(model_config['data_path']):
os.makedirs(model_config['data_path'])
if not os.path.exists(model_config['output_dir']):
os.makedirs(model_config['output_dir'])
# download data
for data_key, url in data_urls.items():
dst = os.path.join(model_config['data_path'], data_key)
if '_' in data_key:
file_name = data_key.split('.')[0]
elif 'wiki.bigram' in data_key:
file_name = 'orig_pretrained_embedding'
else:
file_name = data_key.split('.')[0] # ontology
model_config[file_name] = dst
if not os.path.exists(dst):
download_from_url(url, dst)
avg_best, cnt, acc = 0.0, 0, 0.0
# 数据预处理
train, dev, test, langs, slots, gating_dict = prepare_data_seq(
model_config)
lang = langs[0]
model_config['pretrained_embedding_path'] = os.path.join(
model_config['data_path'], f'emb{len(lang.index2word)}')
print(f'>>> Train configs:')
print('\t', model_config)
# 初始化训练
trainer = Trainer(config=model_config,
langs=langs,
gating_dict=gating_dict,
slots=slots)
# 训练
start_epoch = 0 if not model_config['model_path'] else int(
model_config['model_path'].split('-')[2]) + 1
for epoch in tqdm(range(start_epoch, model_config['num_epochs']),
desc='Epoch'):
progress_bar = tqdm(enumerate(train), total=len(train))
for i, data in progress_bar:
trainer.train_batch(data, slots, reset=(i == 0))
trainer.optimize(int(model_config['grad_clip']))
progress_bar.set_description(trainer.print_loss())
if (epoch + 1) % int(model_config['eval_steps']) == 0:
acc = trainer.evaluate(dev, avg_best, slots, epoch,
model_config['early_stop'])
trainer.scheduler.step(acc)
if acc >= avg_best:
avg_best = acc
cnt = 0
else:
cnt += 1
if cnt == model_config["patience"] or (
acc == 1.0 and model_config['early_stop'] is None):
print("Ran out of patient, early stop...")
break
def main():
model_config_name = "policy/mle/train.json"
common_config_name = "policy/mle/common.json"
data_urls = {
"sys_da_voc.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/usr_da_voc.json",
"usr_da_voc.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/usr_da_voc.json",
}
# load config
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
model_config_path = os.path.join(get_config_path(), model_config_name)
common_config = json.load(open(common_config_path))
model_config = json.load(open(model_config_path))
model_config.update(common_config)
model_config["n_gpus"] = torch.cuda.device_count()
model_config["batch_size"] = (max(1, model_config["n_gpus"]) *
model_config["batch_size"])
model_config["device"] = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
model_config["data_path"] = os.path.join(get_data_path(),
"crosswoz/policy_mle_data")
model_config["raw_data_path"] = os.path.join(get_data_path(),
"crosswoz/raw")
model_config["output_dir"] = os.path.join(root_path,
model_config["output_dir"])
if model_config["load_model_name"]:
model_config["model_path"] = os.path.join(
model_config["output_dir"], model_config["load_model_name"])
else:
model_config["model_path"] = ""
if not os.path.exists(model_config["data_path"]):
os.makedirs(model_config["data_path"])
if not os.path.exists(model_config["output_dir"]):
os.makedirs(model_config["output_dir"])
# download data
for data_key, url in data_urls.items():
dst = os.path.join(model_config["data_path"], data_key)
file_name = data_key.split(".")[0]
model_config[file_name] = dst
if not os.path.exists(dst):
download_from_url(url, dst)
print(f">>> Train configs:")
print("\t", model_config)
set_seed(model_config["random_seed"])
agent = Trainer(model_config)
# 训练
if model_config["do_train"]:
start_epoch = (0 if not model_config["model_path"] else
int(model_config["model_path"].split("-")[2]) + 1)
best = float("inf")
for epoch in tqdm(range(start_epoch, model_config["num_epochs"]),
desc="Epoch"):
agent.imitating(epoch)
best = agent.imit_eval(epoch, best)
agent.calc_metrics()
np.random.seed(seed)
torch.manual_seed(seed)
if __name__ == '__main__':
data_urls = {
'intent_train_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_train_data.json',
'intent_val_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_val_data.json',
'intent_test_data.json':
'http://qiw2jpwfc.hn-bkt.clouddn.com/intent_test_data.json'
}
# load config
root_path = get_root_path()
config_path = os.path.join(get_config_path(),
'crosswoz_all_context_nlu_intent.json')
config = json.load(open(config_path))
data_path = os.path.join(get_data_path(), 'crosswoz/nlu_intent_data/')
output_dir = config['output_dir']
output_dir = os.path.join(root_path, output_dir)
log_dir = config['log_dir']
log_dir = os.path.join(root_path, log_dir)
device = config['DEVICE']
# download data
for data_key, url in data_urls.items():
dst = os.path.join(os.path.join(data_path, data_key))
if not os.path.exists(dst):
download_from_url(url, dst)
def main():
model_config_name = "dst/trade/train.json"
common_config_name = "dst/trade/common.json"
data_urls = {
"train_dials.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/train_dials.json",
"dev_dials.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/dev_dials.json",
"test_dials.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/test_dials.json",
"ontology.json":
"http://qiw2jpwfc.hn-bkt.clouddn.com/ontology.json",
"sgns.wiki.bigram.bz2":
"http://qiw2jpwfc.hn-bkt.clouddn.com/sgns.wiki.bigram.bz2",
}
# load config
root_path = get_root_path()
common_config_path = os.path.join(get_config_path(), common_config_name)
model_config_path = os.path.join(get_config_path(), model_config_name)
common_config = json.load(open(common_config_path))
model_config = json.load(open(model_config_path))
model_config.update(common_config)
model_config["n_gpus"] = torch.cuda.device_count()
model_config["batch_size"] = (max(1, model_config["n_gpus"]) *
model_config["batch_size"])
model_config["device"] = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
if model_config["load_embedding"]:
model_config["hidden_size"] = 300
model_config["data_path"] = os.path.join(get_data_path(),
"crosswoz/dst_trade_data")
model_config["output_dir"] = os.path.join(
root_path, model_config["output_dir"]) # 可以用来保存模型文件
if model_config["load_model_name"]:
model_config["model_path"] = os.path.join(
model_config["output_dir"], model_config["load_model_name"])
else:
model_config["model_path"] = ""
if not os.path.exists(model_config["data_path"]):
os.makedirs(model_config["data_path"])
if not os.path.exists(model_config["output_dir"]):
os.makedirs(model_config["output_dir"])
# download data
for data_key, url in data_urls.items():
dst = os.path.join(model_config["data_path"], data_key)
if "_" in data_key:
file_name = data_key.split(".")[0]
elif "wiki.bigram" in data_key:
file_name = "orig_pretrained_embedding"
else:
file_name = data_key.split(".")[0] # ontology
model_config[file_name] = dst
if not os.path.exists(dst):
download_from_url(url, dst)
avg_best, cnt, acc = 0.0, 0, 0.0
# 数据预处理
train, dev, test, langs, slots, gating_dict = prepare_data_seq(
model_config)
lang = langs[0]
model_config["pretrained_embedding_path"] = os.path.join(
model_config["data_path"], f"emb{len(lang.index2word)}")
print(f">>> Train configs:")
print("\t", model_config)
# 初始化训练
trainer = Trainer(config=model_config,
langs=langs,
gating_dict=gating_dict,
slots=slots)
# 训练
start_epoch = (0 if not model_config["model_path"] else
int(model_config["model_path"].split("-")[2]) + 1)
for epoch in tqdm(range(start_epoch, model_config["num_epochs"]),
desc="Epoch"):
progress_bar = tqdm(enumerate(train), total=len(train))
for i, data in progress_bar:
trainer.train_batch(data, slots, reset=(i == 0))
trainer.optimize(int(model_config["grad_clip"]))
progress_bar.set_description(trainer.print_loss())
if (epoch + 1) % int(model_config["eval_steps"]) == 0:
acc = trainer.evaluate(dev, avg_best, slots, epoch,
model_config["early_stop"])
trainer.scheduler.step(acc)
if acc >= avg_best:
avg_best = acc
cnt = 0
else:
cnt += 1
if cnt == model_config["patience"] or (
acc == 1.0 and model_config["early_stop"] is None):
print("Ran out of patient, early stop...")
break
