en-ner-conll-2003/seq.py
Maciej Sobkowiak 81fa0ec07f gru added
2021-06-22 02:24:08 +02:00

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)