forked from kubapok/en-ner-conll-2003
149 lines
3.9 KiB
Python
149 lines
3.9 KiB
Python
from numpy.lib.shape_base import split
|
|
import pandas as pd
|
|
import numpy as np
|
|
import gensim
|
|
import torch
|
|
import pandas as pd
|
|
from sklearn.model_selection import train_test_split
|
|
from collections import Counter
|
|
from torchtext.vocab import Vocab
|
|
from TorchCRF import CRF
|
|
from tqdm import tqdm
|
|
|
|
EPOCHS = 5
|
|
BATCH = 1
|
|
|
|
# Functions from jupyter
|
|
|
|
|
|
def build_vocab(dataset):
|
|
counter = Counter()
|
|
for document in dataset:
|
|
counter.update(document)
|
|
return Vocab(counter)
|
|
|
|
|
|
def data_process(dt, vocab):
|
|
return [torch.tensor([vocab['<bos>']] + [vocab[token] for token in document] + [vocab['<eos>']], dtype=torch.long) for document in dt]
|
|
|
|
|
|
def get_scores(y_true, y_pred):
|
|
acc_score = 0
|
|
tp = 0
|
|
fp = 0
|
|
selected_items = 0
|
|
relevant_items = 0
|
|
|
|
for p, t in zip(y_pred, y_true):
|
|
if p == t:
|
|
acc_score += 1
|
|
|
|
if p > 0 and p == t:
|
|
tp += 1
|
|
|
|
if p > 0:
|
|
selected_items += 1
|
|
|
|
if t > 0:
|
|
relevant_items += 1
|
|
|
|
if selected_items == 0:
|
|
precision = 1.0
|
|
else:
|
|
precision = tp / selected_items
|
|
|
|
if relevant_items == 0:
|
|
recall = 1.0
|
|
else:
|
|
recall = tp / relevant_items
|
|
|
|
if precision + recall == 0.0:
|
|
f1 = 0.0
|
|
else:
|
|
f1 = 2 * precision * recall / (precision + recall)
|
|
|
|
return precision, recall, f1
|
|
|
|
|
|
def eval_model(dataset_tokens, dataset_labels, model):
|
|
Y_true = []
|
|
Y_pred = []
|
|
for i in tqdm(range(len(dataset_labels))):
|
|
batch_tokens = dataset_tokens[i].unsqueeze(0)
|
|
tags = list(dataset_labels[i].numpy())
|
|
Y_true += tags
|
|
|
|
Y_batch_pred_weights = model(batch_tokens).squeeze(0)
|
|
Y_batch_pred = torch.argmax(Y_batch_pred_weights, 1)
|
|
Y_pred += list(Y_batch_pred.numpy())
|
|
|
|
return get_scores(Y_true, Y_pred)
|
|
|
|
|
|
class GRU(torch.nn.Module):
|
|
def __init__(self, vocab_len):
|
|
super(GRU, self).__init__()
|
|
self.emb = torch.nn.Embedding(vocab_len, 100)
|
|
self.rec = torch.nn.GRU(100, 256, 2, batch_first=True, dropout=0.2)
|
|
self.fc1 = torch.nn.Linear(256, 9)
|
|
|
|
def forward(self, x):
|
|
emb = torch.relu(self.emb(x))
|
|
gru_output, h_n = self.rec(emb)
|
|
out_weights = self.fc1(gru_output)
|
|
return out_weights
|
|
|
|
# Load data
|
|
|
|
|
|
def load_data():
|
|
train = pd.read_csv('train/train.tsv', sep='\t',
|
|
names=['labels', 'document'])
|
|
|
|
Y_train = [y.split(sep=" ") for y in train['labels'].values]
|
|
X_train = [x.split(sep=" ") for x in train['document'].values]
|
|
|
|
dev = pd.read_csv('dev-0/in.tsv', sep='\t', names=['document'])
|
|
exp = pd.read_csv('dev-0/expected.tsv', sep='\t', names=['labels'])
|
|
X_dev = [x.split(sep=" ") for x in dev['document'].values]
|
|
Y_dev = [y.split(sep=" ") for y in exp['labels'].values]
|
|
|
|
test = pd.read_csv('test-A/in.tsv', sep='\t', names=['document'])
|
|
X_test = test['document'].values
|
|
|
|
return X_train, Y_train, X_dev, Y_dev, X_test
|
|
|
|
|
|
def train(model, crf, train_tokens, labels_tokens):
|
|
for i in range(EPOCHS):
|
|
crf.train()
|
|
model.train()
|
|
for i in tqdm(range(len(labels_tokens))):
|
|
batch_tokens = train_tokens[i].unsqueeze(0)
|
|
tags = labels_tokens[i].unsqueeze(1)
|
|
|
|
predicted_tags = model(batch_tokens).squeeze(0).unsqueeze(1)
|
|
|
|
optimizer.zero_grad()
|
|
loss = -crf(predicted_tags, tags)
|
|
|
|
loss.backward()
|
|
optimizer.step()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
X_train, Y_train, X_dev, Y_dev, X_test = load_data()
|
|
vocab_x = build_vocab(X_train)
|
|
vocab_y = build_vocab(Y_train)
|
|
train_tokens = data_process(X_train, vocab_x)
|
|
labels_tokens = data_process(Y_train, vocab_y)
|
|
|
|
# model
|
|
model = GRU(len(vocab_x))
|
|
print(model)
|
|
crf = CRF(9)
|
|
p = list(model.parameters()) + list(crf.parameters())
|
|
optimizer = torch.optim.Adam(p)
|
|
# mask = torch.ByteTensor([1, 1]) # (batch_size. sequence_size)
|
|
train(model, crf, train_tokens, labels_tokens)
|