tokenizer = BertTokenizer.from_pretrained(model_config['path'],
do_lower_case=True)
cv_loaders = get_data_loaders_cv(
rv_path='../data/TRAIN/Train_laptop_reviews.csv',
lb_path='../data/TRAIN/Train_laptop_labels.csv',
tokenizer=tokenizer,
batch_size=args.bs,
type='laptop',
folds=FOLDS)
BEST_THRESHS = [0.1] * FOLDS
BEST_F1 = [0] * FOLDS
for cv_idx, (train_loader, val_loader) in enumerate(cv_loaders):
model = OpinioNet.from_pretrained(model_config['path'],
version=model_config['version'],
focal=model_config['focal'])
model.load_state_dict(
torch.load('../models/pretrained_' + model_config['name']))
model.cuda()
optimizer = Adam(model.parameters(), lr=model_config['lr'])
scheduler = GradualWarmupScheduler(optimizer,
total_epoch=10 * len(train_loader))
best_val_f1 = 0
best_val_loss = float('inf')
for e in range(EP):
print('Epoch [%d/%d] train:' % (e, EP))
train_loss, train_f1, train_pr, train_rc = train_epoch(
model, train_loader, optimizer, scheduler, type='laptop')
if __name__ == '__main__':
EP = 100
SAVING_DIR = '../models/'
tokenizer = BertTokenizer.from_pretrained(
'/home/zydq/.torch/models/bert/chinese-bert_chinese_wwm_pytorch',
do_lower_case=True)
train_loader, val_loader = get_data_loaders(
rv_path='../data/TRAIN/Train_reviews.csv',
lb_path='../data/TRAIN/Train_labels.csv',
tokenizer=tokenizer,
batch_size=12,
val_split=0.15)
model = OpinioNet.from_pretrained(
'/home/zydq/.torch/models/bert/chinese-bert_chinese_wwm_pytorch')
model.cuda()
optimizer = Adam(model.parameters(), lr=5e-6)
scheduler = GradualWarmupScheduler(optimizer, total_epoch=2)
best_val_f1 = 0
best_val_loss = float('inf')
for e in range(EP):
print('Epoch [%d/%d] train:' % (e, EP))
train_loss, train_f1, train_pr, train_rc = train_epoch(
model, train_loader, optimizer, scheduler)
print("loss %.5f, f1 %.5f, pr %.5f, rc %.5f" %
(train_loss, train_f1, train_pr, train_rc))
print('Epoch [%d/%d] eval:' % (e, EP))
val_loss, val_f1, val_pr, val_rc = eval_epoch(model, val_loader)
return total_loss, total_f1, total_pr, total_rc
if __name__ == '__main__':
EP = 100
SAVING_DIR = '../models/'
tokenizer = BertTokenizer.from_pretrained('/home/zydq/.torch/models/bert/chinese_roberta_wwm_ext_L-12_H-768_A-12',
do_lower_case=True)
# tokenizer = BertTokenizer.from_pretrained('/home/zydq/.tf/bert/chinese_roberta_wwm_ext_L-12_H-768_A-12',
# do_lower_case=True)
makeup_train_loader, makeup_val_loader, laptop_train_loader, laptop_val_loader, corpus_loader = \
get_data_loaders_round2(tokenizer, batch_size=12)
model = OpinioNet.from_pretrained('/home/zydq/.torch/models/bert/chinese_roberta_wwm_ext_L-12_H-768_A-12')
# model = OpinioNet.from_pretrained('/home/zydq/.tf/bert/chinese_roberta_wwm_ext_L-12_H-768_A-12', from_tf=True)
model.cuda()
optimizer = Adam(model.parameters(), lr=6e-6)
scheduler = GradualWarmupScheduler(optimizer, total_epoch=2*max(len(makeup_train_loader), len(corpus_loader)))
best_val_f1 = 0
best_val_loss = float('inf')
for e in range(EP):
print('Epoch [%d/%d] train:' % (e, EP))
train_loss, train_lm_loss, train_f1, train_pr, train_rc = train_epoch(model, makeup_train_loader, laptop_train_loader, corpus_loader, optimizer, scheduler)
print("loss %.5f, lm loss %.5f f1 %.5f, pr %.5f, rc %.5f" % (train_loss, train_lm_loss, train_f1, train_pr, train_rc))
print('Epoch [%d/%d] makeup eval:' % (e, EP))
val_loss, val_f1, val_pr, val_rc = eval_epoch(model, makeup_val_loader, type='makeup')
print("makeup_val: loss %.5f, f1 %.5f, pr %.5f, rc %.5f" % (val_loss, val_f1, val_pr, val_rc))
]
tokenizer = BertTokenizer.from_pretrained(
'/home/zydq/.torch/models/bert/chinese-bert_chinese_wwm_pytorch',
do_lower_case=True)
test_dataset = ReviewDataset('../data/TEST/Test_reviews.csv', None,
tokenizer)
test_loader = DataLoader(test_dataset,
12,
collate_fn=test_dataset.batchify,
shuffle=False,
num_workers=5)
ret = None
for name in MODELS:
model_path = osp.join(SAVING_DIR, name)
model = OpinioNet.from_pretrained(
'/home/zydq/.torch/models/bert/chinese-bert_chinese_wwm_pytorch')
model.load_state_dict(torch.load(model_path))
model.cuda()
ret = accum_result(ret, eval_epoch(model, test_loader))
del model
ret = average_result(ret, len(MODELS))
ret = OpinioNet.nms_filter(ret, THRESH)
raw = [s[0][0] for s in test_dataset.samples]
result = gen_submit(ret, raw)
import time
result.to_csv('../submit/ensemble-' + str(round(time.time())) + '.csv',
header=False,
index=False)
print(len(result['id'].unique()), result.shape[0])
tokenizer = BertTokenizer.from_pretrained(model_config['path'],
do_lower_case=True)
test_dataset = ReviewDataset(args.rv, args.lb, tokenizer, 'laptop')
test_loader = DataLoader(test_dataset,
args.bs,
collate_fn=test_dataset.batchify,
shuffle=False,
num_workers=5)
if not raw:
raw = [s[0][0] for s in test_dataset.samples]
if not lb and args.lb:
lb = [s[0][1] for s in test_dataset.samples]
model = OpinioNet.from_pretrained(model_config['path'],
version=model_config['version'],
focal=model_config['focal'])
print(weight_name)
model.load_state_dict(torch.load('../models/' + weight_name))
model.cuda()
ret = accum_result(ret, eval_epoch(model, test_loader, thresh))
del model
ret = average_result(ret, num_model)
ret = OpinioNet.nms_filter(ret, 0.28)
if args.lb:
def f1_score(P, G, S):
pr = S / P
rc = S / G
f1 = 2 * pr * rc / (pr + rc)
type='laptop',
folds=FOLDS))
for model_name, model_config in PRETRAINED_MODELS.items()
])
PRED = []
for cv_idx in range(FOLDS):
cv_model_num = 0
cvret = None
for model_name, model_config in PRETRAINED_MODELS.items():
tokenizer = tokenizers[model_name]
_, val_loader = cv_loaders[model_name][cv_idx]
try:
model = OpinioNet.from_pretrained(
model_config['path'],
version=model_config['version'],
focal=model_config['focal'])
weight_name = model_config['name'] + '_cv' + str(cv_idx)
weight = torch.load('../models/' + weight_name)
except FileNotFoundError:
continue
print(weight_name)
model.load_state_dict(weight)
model.cuda()
try:
thresh = thresh_dict[weight_name]['thresh']
except:
thresh = 0.5
cvret = accum_result(cvret, eval_epoch(model, val_loader, thresh))
cv_model_num += 1
del model