def test_it(MODEL_PATH='roberta-base'):
models = []
for t in os.listdir('type'):
for model_file in os.listdir(os.path.join('type', t)):
model = TweetModel(MODEL_PATH=t)
# model.cuda()
model.cpu()
model.load_state_dict(
torch.load(os.path.join(os.path.join('type', t), model_file)))
model.eval()
models.append(model)
test_df = pd.read_csv('data/test.csv')
test_df['text'] = test_df['text'].astype(str)
test_loader = get_test_loader(test_df, MODEL_PATH=MODEL_PATH)
predictions = []
for data in test_loader:
ids = data['ids'].cuda()
masks = data['masks'].cuda()
tweet = data['tweet']
offsets = data['offsets'].numpy()
sentiment = data['sentiment']
start_logits = []
end_logits = []
# len_logits = []
for model in models:
with torch.no_grad():
model.cuda()
output = model(ids, masks)
start_logits.append(
torch.softmax(output[0], dim=1).cpu().detach().numpy())
end_logits.append(
torch.softmax(output[1], dim=1).cpu().detach().numpy())
# len_logits.append(torch.softmax(output[2], dim=1).cpu().detach().numpy())
model.cpu()
start_logits = np.mean(start_logits, axis=0)
end_logits = np.mean(end_logits, axis=0)
# len_logits = np.mean(len_logits, axis=0)
for i in range(len(ids)):
start_pred = np.argmax(start_logits[i])
end_pred = np.argmax(end_logits[i])
# length = np.argmax(len_logits[i])
# end_pred = start_pred + int(length)
sentiment_val = sentiment[i]
original_tweet = tweet[i]
if start_pred > end_pred:
pred = original_tweet
else:
pred = get_selected_text(tweet[i], start_pred, end_pred,
offsets[i])
if sentiment_val == "neutral" or len(original_tweet.split()) < 2:
pred = original_tweet
predictions.append(pred)
sub_df = pd.read_csv('data/sample_submission.csv')
sub_df['selected_text'] = predictions
sub_df['selected_text'] = sub_df['selected_text'].apply(
lambda x: x.replace('!!!!', '!') if len(x.split()) == 1 else x)
sub_df['selected_text'] = sub_df['selected_text'].apply(
lambda x: x.replace('..', '.') if len(x.split()) == 1 else x)
sub_df['selected_text'] = sub_df['selected_text'].apply(
lambda x: x.replace('...', '.') if len(x.split()) == 1 else x)
sub_df.to_csv('submission.csv', index=False)
sub_df.head()
def main():
seed = 42
seed_everything(seed)
num_epochs = 3
batch_size = 32
skf = StratifiedKFold(n_splits=10, shuffle=True, random_state=seed)
train_df = pd.read_csv('data/train.csv')
train_df['text'] = train_df['text'].astype(str)
train_df['selected_text'] = train_df['selected_text'].astype(str)
for fold, (train_idx,
val_idx) in enumerate(skf.split(train_df, train_df.sentiment),
start=1):
print(f'Fold: {fold}')
model = TweetModel()
optimizer = optim.AdamW(model.parameters(),
lr=3e-5,
betas=(0.9, 0.999))
criterion = loss_fn
dataloaders_dict = get_train_val_loaders(train_df, train_idx, val_idx,
batch_size)
train_model(model, dataloaders_dict, criterion, optimizer, num_epochs,
f'roberta_fold{fold}.pth')
# inference
test_df = pd.read_csv('data/test.csv')
test_df['text'] = test_df['text'].astype(str)
test_loader = get_test_loader(test_df)
predictions = []
models = []
for fold in range(skf.n_splits):
model = TweetModel()
model.cuda()
model.load_state_dict(torch.load(f'roberta_fold{fold+1}.pth'))
model.eval()
models.append(model)
for data in test_loader:
ids = data['ids'].cuda()
masks = data['masks'].cuda()
tweet = data['tweet']
offsets = data['offsets'].numpy()
start_logits = []
end_logits = []
for model in models:
with torch.no_grad():
output = model(ids, masks)
start_logits.append(
torch.softmax(output[0], dim=1).cpu().detach().numpy())
end_logits.append(
torch.softmax(output[1], dim=1).cpu().detach().numpy())
start_logits = np.mean(start_logits, axis=0)
end_logits = np.mean(end_logits, axis=0)
for i in range(len(ids)):
start_pred = np.argmax(start_logits[i])
end_pred = np.argmax(end_logits[i])
if start_pred > end_pred:
pred = tweet[i]
else:
pred = get_selected_text(tweet[i], start_pred, end_pred,
offsets[i])
predictions.append(pred)
#submission
sub_df = pd.read_csv('data/sample_submission.csv')
sub_df['selected_text'] = predictions
sub_df['selected_text'] = sub_df['selected_text'].apply(
lambda x: x.replace('!!!!', '!') if len(x.split()) == 1 else x)
sub_df['selected_text'] = sub_df['selected_text'].apply(
lambda x: x.replace('..', '.') if len(x.split()) == 1 else x)
sub_df['selected_text'] = sub_df['selected_text'].apply(
lambda x: x.replace('...', '.') if len(x.split()) == 1 else x)
sub_df.to_csv('submission.csv', index=False)
sub_df.head()
def predict(df_test):
device = torch.device("cuda")
model_config = transformers.RobertaConfig.from_pretrained(config.ROBERTA_PATH)
model_config.output_hidden_states = True
model1 = TweetModel(conf=model_config)
model1.to(device)
model1.load_state_dict(torch.load("model_0.bin"))
model1.eval()
model2 = TweetModel(conf=model_config)
model2.to(device)
model2.load_state_dict(torch.load("model_1.bin"))
model2.eval()
model3 = TweetModel(conf=model_config)
model3.to(device)
model3.load_state_dict(torch.load("model_2.bin"))
model3.eval()
model4 = TweetModel(conf=model_config)
model4.to(device)
model4.load_state_dict(torch.load("model_3.bin"))
model4.eval()
model5 = TweetModel(conf=model_config)
model5.to(device)
model5.load_state_dict(torch.load("model_4.bin"))
model5.eval()
final_output = []
test_dataset = TweetDataset(
tweet=df_test.text.values,
sentiment=df_test.sentiment.values,
selected_text=df_test.selected_text.values
)
data_loader = torch.utils.data.DataLoader(
test_dataset,
shuffle=False,
batch_size=config.VALID_BATCH_SIZE,
num_workers=1
)
jaccards = utils.AverageMeter()
with torch.no_grad():
tk0 = tqdm(data_loader, total=len(data_loader))
for bi, d in enumerate(tk0):
ids = d["ids"]
token_type_ids = d["token_type_ids"]
mask = d["mask"]
sentiment = d["sentiment"]
orig_selected = d["orig_selected"]
orig_tweet = d["orig_tweet"]
targets_start = d["targets_start"]
targets_end = d["targets_end"]
offsets = d["offsets"].numpy()
ids = ids.to(device, dtype=torch.long)
token_type_ids = token_type_ids.to(device, dtype=torch.long)
mask = mask.to(device, dtype=torch.long)
targets_start = targets_start.to(device, dtype=torch.long)
targets_end = targets_end.to(device, dtype=torch.long)
outputs_start1, outputs_end1 = model1(
ids=ids,
mask=mask,
token_type_ids=token_type_ids
)
outputs_start2, outputs_end2 = model2(
ids=ids,
mask=mask,
token_type_ids=token_type_ids
)
outputs_start3, outputs_end3 = model3(
ids=ids,
mask=mask,
token_type_ids=token_type_ids
)
outputs_start4, outputs_end4 = model4(
ids=ids,
mask=mask,
token_type_ids=token_type_ids
)
outputs_start5, outputs_end5 = model5(
ids=ids,
mask=mask,
token_type_ids=token_type_ids
)
outputs_start = (
outputs_start1
+ outputs_start2
+ outputs_start3
+ outputs_start4
+ outputs_start5
) / 5
outputs_end = (
outputs_end1
+ outputs_end2
+ outputs_end3
+ outputs_end4
+ outputs_end5
) / 5
outputs_start = torch.softmax(outputs_start, dim=1).cpu().detach().numpy()
outputs_end = torch.softmax(outputs_end, dim=1).cpu().detach().numpy()
jaccard_scores = []
for px, tweet in enumerate(orig_tweet):
selected_tweet = orig_selected[px]
tweet_sentiment = sentiment[px]
jaccard_score, output_sentence = calculate_jaccard_score(
original_tweet=tweet,
target_string=selected_tweet,
sentiment_val=tweet_sentiment,
idx_start=np.argmax(outputs_start[px, :]),
idx_end=np.argmax(outputs_end[px, :]),
offsets=offsets[px]
)
jaccard_scores.append(jaccard_score)
final_output.append(output_sentence)
jaccards.update(np.mean(jaccard_scores), ids.size(0))
return final_output, jaccards.avg
def main(args, mode):
config = Config(
train_dir='/mfs/renxiangyuan/tweets/data/train_folds.csv', # 原始数据
# train_dir='/mfs/renxiangyuan/tweets/data/train_folds_extra.csv', # 加入更多sentimen分类数据
model_save_dir=
f'/mfs/renxiangyuan/tweets/output/{args.model_type}-5-fold-ak',
# model_save_dir=f'/mfs/renxiangyuan/tweets/output/shuffle/{args.model_type}-5-fold-ak',
model_type=args.model_type,
batch_size=args.bs,
seed=args.seed,
lr=args.lr * 1e-5,
max_seq_length=args.max_seq_length,
num_hidden_layers=args.num_hidden_layers,
cat_n_layers=args.cat_n_layers,
froze_n_layers=args.froze_n_layers,
# conv_head=True,
# eps=args.eps,
shuffle_seed=args.shuffle_seed,
init_seed=args.init_seed,
epochs=args.epochs, # 默认epochs=3
warmup_samples=args.warmup_samples,
# frozen_warmup=False,
warmup_scheduler=args.scheduler,
mask_pad_loss=args.mask_pad_loss,
smooth=args.smooth,
# fp16=False,
io_loss_ratio=args.io_loss_ratio,
io_loss_type=args.io_loss_type,
# multi_sent_loss_ratio=0,
# clean_data=True, # 模型clean_data=False
)
config.print_info()
set_seed(config.seed)
# 训练
if "train" in mode:
os.makedirs(config.MODEL_SAVE_DIR, exist_ok=True)
jaccard_scores = []
for i in args.train_folds:
scores_i = train(fold=i, config=config)
jaccard_scores.append(scores_i)
# if i == 0 and max(scores_i) < 0.705:
# print("Fold 0 Too Weak, Early Stop")
# break
for i, res_i in enumerate(jaccard_scores):
print(i, res_i)
print("mean", np.mean([max(scores) for scores in jaccard_scores]))
for i in range(1, config.EPOCHS):
print(f"\tEpoch{i+1}: ",
np.mean([scores[i] for scores in jaccard_scores]))
config.print_info()
# 测试
if "test" in mode:
model_paths = [
"/mfs/renxiangyuan/tweets/output/shuffle/roberta-squad-5-fold-ak/4e-05lr_32bs_42sd_128len_12layer_1cat_-1froze_11shufflesd/model_0_epoch_2.pth",
"/mfs/renxiangyuan/tweets/output/shuffle/roberta-squad-5-fold-ak/4e-05lr_32bs_42sd_128len_12layer_1cat_-1froze_3shufflesd/model_1_epoch_3.pth",
"/mfs/renxiangyuan/tweets/output/shuffle/roberta-squad-5-fold-ak/4e-05lr_32bs_42sd_128len_12layer_1cat_-1froze_18shufflesd/model_2_epoch_3.pth",
"/mfs/renxiangyuan/tweets/output/shuffle/roberta-squad-5-fold-ak/4e-05lr_32bs_42sd_128len_12layer_1cat_-1froze_13shufflesd/model_3_epoch_2.pth",
"/mfs/renxiangyuan/tweets/output/shuffle/roberta-squad-5-fold-ak/4e-05lr_32bs_42sd_128len_12layer_1cat_-1froze_19shufflesd/model_4_epoch_3.pth",
# "/mfs/renxiangyuan/tweets/output/roberta-base-5-fold-ak/5e-05lr_32bs_42sd_13layer/model_0_epoch_2.pth",
# "/mfs/renxiangyuan/tweets/output/roberta-base-5-fold-ak/5e-05lr_32bs_42sd_13layer/model_1_epoch_2.pth",
# "/mfs/renxiangyuan/tweets/output/roberta-base-5-fold-ak/5e-05lr_32bs_42sd_13layer/model_2_epoch_3.pth",
# "/mfs/renxiangyuan/tweets/output/roberta-base-5-fold-ak/5e-05lr_32bs_42sd_13layer/model_3_epoch_3.pth",
# "/mfs/renxiangyuan/tweets/output/roberta-base-5-fold-ak/5e-05lr_32bs_42sd_13layer/model_4_epoch_3.pth",
]
ensemble_infer(model_paths, config)
# ensemble_infer(model_paths=None, config=config)
# # 评估
if "evaluate" in mode:
device = torch.device("cuda")
model = TweetModel(conf=config.model_config, config=config)
model.to(device)
res = [[] for _ in range(5)]
for fold in range(5):
dfx = pd.read_csv(config.TRAINING_FILE)
df_valid = dfx[dfx.kfold == fold].reset_index(drop=True)
valid_dataset = TweetDataset(
tweet=df_valid.text.values,
sentiment=df_valid.sentiment.values,
selected_text=df_valid.selected_text.values,
config=config,
)
valid_data_loader = DataLoader(valid_dataset,
batch_size=config.VALID_BATCH_SIZE,
num_workers=8)
for ep in range(1, config.EPOCHS):
state_dict_dir = os.path.join(
config.MODEL_SAVE_DIR, f"model_{fold}_epoch_{ep+1}.pth")
print(state_dict_dir)
model.load_state_dict(torch.load(state_dict_dir))
model.eval()
jaccards = eval_fn(valid_data_loader, model, device, config)
print(jaccards)
res[fold].append(jaccards)
for i, res_i in enumerate(res):
print(i, res_i)
print("mean", np.mean([max(scores) for scores in res]))
for i in range(2):
print(f"\tEpoch{i + 1}: ", np.mean([scores[i] for scores in res]))