hf roberta base epoch1 (fix)

This commit is contained in:
Jakub Pokrywka 2021-12-14 12:30:15 +01:00
parent f986c74861
commit dc6eb48ec2
5 changed files with 19048 additions and 19037 deletions

File diff suppressed because it is too large Load Diff

View File

@ -1,26 +1,38 @@
import pickle
from datasets import load_dataset
from datasets import load_dataset, Dataset
from transformers import AutoTokenizer
from tqdm import tqdm
from sklearn.preprocessing import MinMaxScaler
import numpy as np
dataset = load_dataset('csv', sep='\t', data_files={'train': ['./train_huggingface_format.csv'], 'test': ['./dev-0_huggingface_format.csv']})
test_dataset_A = load_dataset('csv', sep='\t', data_files='test-A_huggingface_format.csv')
#dataset = load_dataset('csv', sep='\t', data_files={'train': ['./train_huggingface_format.csv'], 'test': ['./dev-0_huggingface_format.csv']})
#test_dataset_A = load_dataset('csv', sep='\t', data_files='test-A_huggingface_format.csv')
#
tokenizer = AutoTokenizer.from_pretrained('roberta-base')
#
def tokenize_function(examples):
t = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=512)
return t
#
#test_tokenized_datasets_A = test_dataset_A.map(tokenize_function, batched=True)
#tokenized_datasets = dataset.map(tokenize_function, batched=True)
test_tokenized_datasets_A = test_dataset_A.map(tokenize_function, batched=True)
tokenized_datasets = dataset.map(tokenize_function, batched=True)
def get_dataset_dict(dataset):
with open(dataset) as f_in:
next(f_in)
d = dict()
d['year'] = list()
d['text'] = list()
for l in f_in:
y,t = l.rstrip().split('\t')
d['year'].append(y)
d['text'].append(t)
return d
train_dataset = tokenized_datasets["train"].shuffle(seed=42)
eval_dataset_full = tokenized_datasets["test"]
eval_dataset_small = tokenized_datasets["test"].select(range(2000))
test_dataset_A = test_tokenized_datasets_A["train"]
train_dataset = Dataset.from_dict(get_dataset_dict('train_huggingface_format.csv')).map(tokenize_function, batched=True).shuffle(seed=42)
eval_dataset_full = Dataset.from_dict(get_dataset_dict('dev-0_huggingface_format.csv')).map(tokenize_function, batched=True).shuffle(seed=42)
eval_dataset_small = eval_dataset_full.select(range(2000))
test_dataset_A = Dataset.from_dict(get_dataset_dict('test-A_huggingface_format.csv')).map(tokenize_function, batched=True).shuffle(seed=42)
scalers = dict()
@ -34,7 +46,7 @@ def add_scaled(example):
train_dataset = train_dataset.map(add_scaled)
eval_dataset_full = eval_dataset_full.map(add_scaled)
eval_dataset_small = eval_dataset_small.map(add_scaled)
#test_dataset_A = test_dataset_A.map(add_scaled)
test_dataset_A = test_dataset_A.map(add_scaled)
with open('train_dataset.pickle','wb') as f_p:

View File

@ -10,7 +10,6 @@ from tqdm.auto import tqdm
BATCH_SIZE = 4
with open('train_dataset.pickle','rb') as f_p:
train_dataset = pickle.load(f_p)
@ -29,14 +28,14 @@ model = AutoModelForSequenceClassification.from_pretrained('roberta-base', num_l
optimizer = AdamW(model.parameters(), lr=1e-6)
num_epochs = 3
num_epochs = 1
num_training_steps = num_epochs * len(train_dataloader)
lr_scheduler = get_scheduler(
"linear",
optimizer=optimizer,
num_warmup_steps=0,
num_training_steps=num_training_steps
)
#lr_scheduler = get_scheduler(
# "linear",
# optimizer=optimizer,
# num_warmup_steps=0,
# num_training_steps=num_training_steps
#)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
@ -71,7 +70,7 @@ def eval():
outputs = model(**batch)
loss = outputs.loss
eval_loss += loss.item()
print(f'eval loss: {eval_loss / i }')
print(f'epoch {epoch} eval loss: {eval_loss / i }')
model.train()
@ -84,14 +83,14 @@ for epoch in range(num_epochs):
loss.backward()
optimizer.step()
lr_scheduler.step()
#lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
train_loss += loss.item()
#import pdb; pdb.set_trace()
if i % 5000 == 0 and i > 1 :
print(f'train loss: {train_loss / 5000 }', end = '\t\t')
print(f' epoch {epoch} train loss: {train_loss / 5000 }', end = '\t\t')
train_loss = 0.0
eval()

View File

@ -20,7 +20,7 @@ with open('eval_dataset_full.pickle','rb') as f_p:
eval_dataset_full = pickle.load(f_p)
device = 'cuda'
model = AutoModelForSequenceClassification.from_pretrained('./roberta_year_prediction')
model = AutoModelForSequenceClassification.from_pretrained('./roberta_year_prediction/epoch_0')
model.eval()
model.to(device)

File diff suppressed because it is too large Load Diff