Compare commits

...

2 Commits

Author SHA1 Message Date
Dawid
ded9a74adc added fine tunning 2021-06-22 10:32:27 +02:00
Dawid
59a86d7607 added logistic reg with neural net 2021-05-23 11:45:52 +02:00
11 changed files with 600169 additions and 0 deletions

5272
dev-0/in.tsv Normal file

File diff suppressed because one or more lines are too long

5272
dev-0/out.tsv Normal file

File diff suppressed because it is too large Load Diff

59
fine_tuning.py Normal file
View File

@ -0,0 +1,59 @@
import random
import torch
from transformers import (
AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer
)
class DataWrapper(torch.utils.data.Dataset):
def __init__(self, encodings, labels):
self.encodings = encodings
self.labels = labels
def __getitem__(self, idx):
item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}
item['labels'] = torch.tensor(self.labels[idx])
return item
def __len__(self):
return len(self.labels)
def read_data(file_path):
with open(file_path) as f:
return f.readlines()
def wirte_output(file_path, data):
with open(file_path, 'w') as writer:
for result in trainer.predict(data):
writer.write(f"{str(result)}\n")
print("STEP 1 - READ DATA")
X_train = read_data('train/in.tsv')
y_train = read_data('train/expected.tsv')
X_dev = read_data('dev-0/in.tsv')
X_test = read_data('test-A/in.tsv')
print("STEP 2 - SHUFFLE")
data_train = list(zip(X_train, y_train))
data_train = random.sample(data_train, 15000)
print("STEP 3 - FINE TUNING")
tokenizer = AutoTokenizer.from_pretrained("roberta-base")
train_encodings = tokenizer([text[0] for text in data_train], truncation=True, padding=True)
train_dataset = DataWrapper(train_encodings, [int(text[1]) for text in data_train])
model = AutoModelForSequenceClassification.from_pretrained("roberta-base", num_labels=2)
args = TrainingArguments("model")
device = torch.device("cpu")
# device = torch.device("cuda")
model.to(device)
trainer = Trainer(model=model, args=args, train_dataset=train_dataset)
trainer.train()
print("STEP 4 - WRITE OUTPUT")
wirte_output('train/out.tsv', X_train)
wirte_output('dev-0/out.tsv', X_dev)
wirte_output('test-A/out.tsv', X_test)

BIN
geval Executable file

Binary file not shown.

5
geval_results.txt Normal file
View File

@ -0,0 +1,5 @@
Likelihood 0.0000
Accuracy 0.7517
F1.0 0.6119
Precision 0.6848
Recall 0.5531

94
log_reg.py Normal file
View File

@ -0,0 +1,94 @@
import numpy as np
import pandas as pd
import torch
import csv
import gensim.downloader
import torch
from nltk import word_tokenize
class NeuralNetwork(torch.nn.Module):
def __init__(self, input_size, hidden_size, num_classes):
super(NeuralNetwork, self).__init__()
self.l1 = torch.nn.Linear(input_size, hidden_size)
self.l2 = torch.nn.Linear(hidden_size, num_classes)
def forward(self, x):
x = self.l1(x)
x = torch.relu(x)
x = self.l2(x)
x = torch.sigmoid(x)
return x
print('STEP 1 - LOAD DATA')
names = ['content', 'id', 'label']
train_data_content = pd.read_table('train/in.tsv', error_bad_lines=False, header=None, quoting=csv.QUOTE_NONE, names=names[:2])
train_data_labels = pd.read_table('train/expected.tsv', error_bad_lines=False, header=None, quoting=csv.QUOTE_NONE, names=names[2:])
dev_data = pd.read_table('dev-0/in.tsv', error_bad_lines=False, header=None, quoting=csv.QUOTE_NONE, names=names[:2])
test_data = pd.read_table('test-A/in.tsv', error_bad_lines=False, header=None, quoting=csv.QUOTE_NONE, names=names[:2])
print('STEP 2 - SET PARAMS')
hidden_size = int(input('Hidden units size: ') or '600')
epochs = int(input("Epochs: ") or '5')
batch_size = int(input("Batch size: ") or '15')
print('STEP 3 - PREPROCESSING')
# lowercase all content
X_train = train_data_content['content'].str.lower()
y_train = train_data_labels['label']
X_dev = dev_data['content'].str.lower()
X_test = test_data['content'].str.lower()
# tokenize datasets
X_train = [word_tokenize(content) for content in X_train]
X_dev = [word_tokenize(content) for content in X_dev]
X_test = [word_tokenize(content) for content in X_test]
# use Google word2vec algorithm
word2vec = gensim.downloader.load('word2vec-google-news-300')
X_train = [np.mean([word2vec[word] for word in content if word in word2vec] or [np.zeros(300)], axis=0) for content in X_train]
X_dev = [np.mean([word2vec[word] for word in content if word in word2vec] or [np.zeros(300)], axis=0) for content in X_dev]
X_test = [np.mean([word2vec[word] for word in content if word in word2vec] or [np.zeros(300)], axis=0) for content in X_test]
print('STEP 4 - MODEL TRAINING')
#prepare neural model
model = NeuralNetwork(300, hidden_size, 1)
criterion = torch.nn.BCELoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
for epoch in range(epochs):
model.train()
for i in range(0, y_train.shape[0], batch_size):
X = X_train[i:i+batch_size]
X = torch.tensor(X)
y = y_train[i:i+batch_size]
y = torch.tensor(y.astype(np.float32).to_numpy()).reshape(-1, 1)
outputs = model(X.float())
loss = criterion(outputs, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('STEP 5 - PREDICTION')
y_dev, y_test = [], []
model.eval()
with torch.no_grad():
for i in range(0, len(X_dev), batch_size):
X = X_dev[i:i+batch_size]
X = torch.tensor(X)
outputs = model(X.float())
prediction = (outputs > 0.5)
y_dev += prediction.tolist()
for i in range(0, len(X_test), batch_size):
X = X_test[i:i+batch_size]
X = torch.tensor(X)
outputs = model(X.float())
y = (outputs > 0.5)
y_test += prediction.tolist()
print('STEP 6 - EXPORT RESULTS')
# export results to tsv
y_dev = np.asarray(y_dev, dtype=np.int32)
y_test = np.asarray(y_test, dtype=np.int32)
y_dev.tofile('./dev-0/out.tsv', sep='\n')
y_test.tofile('./test-A/out.tsv', sep='\n')

5
output_geval_fine.txt Normal file
View File

@ -0,0 +1,5 @@
Likelihood 0.0000
Accuracy 0.8253
F1.0 0.7472
Precision 0.7659
Recall 0.7294

5152
test-A/in.tsv Normal file

File diff suppressed because one or more lines are too long

5152
test-A/out.tsv Normal file

File diff suppressed because it is too large Load Diff

289579
train/in.tsv Normal file

File diff suppressed because one or more lines are too long

289579
train/out.tsv Normal file

File diff suppressed because it is too large Load Diff