def validate_model(self):
logger.info('Validate GuidAbs model %s' % self.model_path)
fn_touch = path.join(self.model_path,
'finished.validate_guidabs_model')
if path.exists(fn_touch):
return
args = self._build_abs_args()
args.mode = 'validate'
args.bert_data_path = path.join(self.data_path, 'cnndm')
args.model_path = self.model_path
args.log_file = path.join(self.model_path, 'val_abs_bert_cnndm.log')
args.batch_size = args.test_batch_size
init_logger(args.log_file)
acc_top3 = train_abs.validate_abs(args, args.device_id)
# rename top3 models and remove other models
for i, (acc, xent, cp) in enumerate(acc_top3):
fn = path.basename(cp)
tgt_path = path.join(self.model_path, 'top%s.%s' % (i, fn))
os.system('mv %s %s -f' % (cp, tgt_path))
logger.info('Archive validated GuidAbs model %s' % tgt_path)
os.system('rm %s/model_step_*.pt -f' % self.model_path)
os.system('touch %s' % fn_touch)
parser.add_argument("-block_trigram", type=str2bool, nargs='?', const=True, default=True)
args = parser.parse_args()
args.gpu_ranks = [int(i) for i in range(len(args.visible_gpus.split(',')))]
args.world_size = len(args.gpu_ranks)
os.environ["CUDA_VISIBLE_DEVICES"] = args.visible_gpus
init_logger(args.log_file)
device = "cpu" if args.visible_gpus == '-1' else "cuda"
device_id = 0 if device == "cuda" else -1
if (args.task == 'abs'):
if (args.mode == 'train'):
train_abs(args, device_id)
elif (args.mode == 'validate'):
validate_abs(args, device_id)
elif (args.mode == 'lead'):
baseline(args, cal_lead=True)
elif (args.mode == 'oracle'):
baseline(args, cal_oracle=True)
if (args.mode == 'test'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
step = 0
test_abs(args, device_id, cp, step)
elif (args.mode == 'test_text'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
init_logger(args.log_file)
device = "cpu" if args.visible_gpus == '-1' else "cuda"
device_id = 0 if device == "cuda" else -1
if (args.task == 'abs'):
if args.uncased:
tokenizer = BertTokenizer.from_pretrained(
'bert-base-multilingual-uncased', do_lower_case=True)
else:
tokenizer = BertTokenizer.from_pretrained(
'bert-base-multilingual-cased', do_lower_case=False)
if (args.mode == 'train'):
train_abs(args, device_id, tokenizer)
elif (args.mode == 'validate'):
validate_abs(args, device_id, tokenizer)
elif (args.mode == 'lead'):
baseline(args, cal_lead=True)
elif (args.mode == 'oracle'):
baseline(args, cal_oracle=True)
if (args.mode == 'test'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
step = 0
test_abs(args, device_id, cp, step, tokenizer)
elif (args.mode == 'dev'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
def run(args, hpspace):
if (args.task == 'abs'):
if (args.mode == 'train'):
if (args.hyperopt):
t = time.time()
newT = str(t).split(".")
args.model_path = "../models/" + newT[0]
args.log_file = "../logs/abs_bert_abs_" + newT[0]
args.lr_bert = hpspace['lr_bert']
args.lr_dec = hpspace['lr_dec']
args.accum_count = int(hpspace['accum_count'])
args.beta1 = hpspace['beta1']
args.beta2 = hpspace['beta2']
args.visible_gpus = '0'
args.bert_model = '..temp/bert-base-danish-uncased-v2'
args.vocab = '..temp/bert-base-danish-uncased-v2'
train_stats = train_abs(args, device_id)
x = train_stats.x
ppl = train_stats.perplexity
acc = train_stats.acc
print(x)
return {
'loss': x,
'eval_time': time.time(),
'status': STATUS_OK,
'other_stuff': {
'ppl': ppl,
'acc': acc
}
}
elif (args.mode == 'validate'):
validate_abs(args, device_id)
elif (args.mode == 'lead'):
baseline(args, cal_lead=True)
elif (args.mode == 'oracle'):
baseline(args, cal_oracle=True)
if (args.mode == 'test'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
step = 0
test_abs(args, device_id, cp, step)
elif (args.mode == 'test_text'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
step = 0
test_text_abs(args, device_id, cp, step)
elif (args.task == 'ext'):
if (args.mode == 'train'):
train_ext(args, device_id)
elif (args.mode == 'validate'):
validate_ext(args, device_id)
if (args.mode == 'test'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
step = 0
test_ext(args, device_id, cp, step)
elif (args.mode == 'test_text'):
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
step = 0
test_text_abs(args, device_id, cp, step)
if (args.mode == "sent_label"):
step = 0
sent_label_ext(args, device_id)
DEVICE = "cpu" if args.visible_gpus == '-1' else "cuda"
DEVICE_ID = 0 if DEVICE == "cuda" else -1
# Create directories
os.makedirs(args.model_path, exist_ok=True)
os.makedirs(args.log_path, exist_ok=True)
os.makedirs(args.result_path, exist_ok=True)
# Train/valid/test
if args.task == 'abs':
if args.mode == 'train':
init_logger(os.path.join(args.log_path, 'train.log'))
train_abs(args, DEVICE_ID)
elif args.mode == 'validate':
init_logger(os.path.join(args.log_path, 'valid.log'))
validate_abs(args, DEVICE_ID)
elif args.mode == 'test':
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
print("Not correct model name (EX: model_step_1200.pt)")
init_logger(
os.path.join(args.log_path, 'test.' + str(step) + '.log'))
test_abs(args, DEVICE_ID, cp, step)
elif args.mode == 'test_text':
cp = args.test_from
try:
step = int(cp.split('.')[-2].split('_')[-1])
except:
print("Not correct model name (EX: model_step_1200.pt)")
