lalka-lm/script.py

import numpy as np
import torch
from sklearn.model_selection import train_test_split
import nltk
from nltk.tokenize import word_tokenize

nltk.download('punkt')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_a = "train/train.tsv"
lalka_path_train = 'train/train_train.tsv'
lalka_path_valid = 'train/train_test.tsv'
corpora_train = open(lalka_path_train).read()
corpora_train_tokenized = list(word_tokenize(corpora_train))
corpora_train_tokenized = [token.lower() for token in corpora_train_tokenized]

vocab_itos = sorted(set(corpora_train_tokenized))
vocab_itos = vocab_itos[:15005]
vocab_itos[15001] = "<UNK>"
vocab_itos[15002] = "<BOS>"
vocab_itos[15003] = "<EOS>"
vocab_itos[15004] = "<PAD>"

BATCH_SIZE = 128
EPOCHS = 15


history_ppl_train = []
history_ppl_valid = []

vocab_stoi = dict()
for i, token in enumerate(vocab_itos):
    vocab_stoi[token] = i

NGRAMS = 5


def set_ppl(dataset_id_list):
    lm.eval()

    batches = 0
    loss_sum = 0

    for i in range(0, len(dataset_id_list) - BATCH_SIZE + 1, BATCH_SIZE):
        X = dataset_id_list[i:i + BATCH_SIZE, :NGRAMS - 1]
        Y = dataset_id_list[i:i + BATCH_SIZE, NGRAMS - 1]
        predictions = lm(X)

        loss = criterion(predictions, Y)

        loss_sum += loss.item()
        batches += 1

    return np.exp(loss_sum / batches)


def open_files(path_a, path_b, path_c):
    with open(path_a, "r") as path:
        lines = path.readlines()
    train, test = train_test_split(lines, test_size=0.2)
    with open(path_b, "w") as out_train_file:
        for i in train:
            out_train_file.write(i)
    with open(path_c, "w") as out_test_file:
        for i in test:
            out_test_file.write(i)


def get_samples(dataset):
    samples = []
    for i in range(len(dataset) - NGRAMS):
        samples.append(dataset[i:i + NGRAMS])
    return samples


def get_token_id(dataset):
    token_id_list = [vocab_stoi['<PAD>']] * (NGRAMS - 1) + [vocab_stoi['<BOS>']]
    for token in dataset:
        try:
            token_id_list.append(vocab_stoi[token])
        except KeyError:
            token_id_list.append(vocab_stoi['<UNK>'])
    token_id_list.append(vocab_stoi['<EOS>'])
    return token_id_list


open_files(train_a, lalka_path_train, lalka_path_valid)
train_id_list = get_token_id(corpora_train_tokenized)
train_id_list = get_samples(train_id_list)
train_id_list = torch.tensor(train_id_list, device=device)
corpora_valid = open(lalka_path_valid).read()
corpora_valid_tokenized = list(word_tokenize(corpora_valid))
corpora_valid_tokenized = [token.lower() for token in corpora_valid_tokenized]
valid_id_list = get_token_id(corpora_valid_tokenized)
valid_id_list = torch.tensor(get_samples(valid_id_list), dtype=torch.long, device=device)


class GRU(torch.nn.Module):

    def __init__(self):
        super(GRU, self).__init__()
        self.emb = torch.nn.Embedding(len(vocab_itos), 100)
        self.rec = torch.nn.GRU(100, 256, 1, batch_first=True)
        self.fc1 = torch.nn.Linear(256, len(vocab_itos))

    def forward(self, x):
        emb = self.emb(x)
        output, h_n = self.rec(emb)
        hidden = h_n.squeeze(0)
        out = self.fc1(hidden)
        return out


lm = GRU().to(device)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(lm.parameters(), lr=0.0001)

for epoch in range(EPOCHS):

    batches = 0
    loss_sum = 0
    acc_score = 0
    lm.train()
    for i in range(0, len(train_id_list) - BATCH_SIZE + 1, BATCH_SIZE):
        X = train_id_list[i:i + BATCH_SIZE, :NGRAMS - 1]
        Y = train_id_list[i:i + BATCH_SIZE, NGRAMS - 1]
        predictions = lm(X)
        loss = criterion(predictions, Y)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        loss_sum += loss.item()
        batches += 1

    ppl_train = set_ppl(train_id_list)
    ppl_valid = set_ppl(valid_id_list)

    history_ppl_train.append(ppl_train)
    history_ppl_valid.append(ppl_valid)


tokenized = list(word_tokenize('Gości innych nie widział oprócz spółleśników'))
tokenized = [token.lower() for token in tokenized]

id_list = []
for word in tokenized:
    if word in vocab_stoi:
        id_list.append(vocab_stoi[word])
    else:
        id_list.append(vocab_stoi['<UNK>'])

lm.eval()

id_list = torch.tensor(id_list, dtype=torch.long, device=device)

preds = lm(id_list.unsqueeze(0))

vocab_itos[torch.argmax(torch.softmax(preds, 1), 1).item()]

tokenized = list(word_tokenize('Lalka'))
tokenized = [token.lower() for token in tokenized]

id_list = []
for word in tokenized:
    if word in vocab_stoi:
        id_list.append(vocab_stoi[word])
    else:
        id_list.append(vocab_stoi['<UNK>'])
id_list = torch.tensor([id_list], dtype=torch.long, device=device)

candidates_number = 10
for i in range(30):
    preds = lm(id_list)
    candidates = torch.topk(torch.softmax(preds, 1), candidates_number)[1][0].cpu().numpy()
    candidate = 15001
    while candidate > 15000:
        candidate = candidates[np.random.randint(candidates_number)]
    id_list = torch.cat((id_list, torch.tensor([[candidate]], device=device)), 1)

with open("dev-0/in.tsv", "r") as dev_path:
    nr_of_dev_lines = len(dev_path.readlines())
with open("test-A/in.tsv", "r") as test_a_path:
    nr_of_test_a_lines = len(test_a_path.readlines())
with open("dev-0/out.tsv", "w") as out_dev_file:
    for i in range(nr_of_dev_lines):
        preds = lm(id_list)
        candidates = torch.topk(torch.softmax(preds, 1), candidates_number)[1][0].cpu().numpy()
        candidate = 15001
        while candidate > 15000:
            candidate = candidates[np.random.randint(candidates_number)]
        id_list = torch.cat((id_list, torch.tensor([[candidate]], device=device)), 1)
        out_dev_file.write(vocab_itos[candidate] + '\n')
with open("test-A/out.tsv", "w") as out_test_file:
    for i in range(nr_of_dev_lines):
        preds = lm(id_list)
        candidates = torch.topk(torch.softmax(preds, 1), candidates_number)[1][0].cpu().numpy()
        candidate = 15001
        while candidate > 15000:
            candidate = candidates[np.random.randint(candidates_number)]
        id_list = torch.cat((id_list, torch.tensor([[candidate]], device=device)), 1)
        out_test_file.write(vocab_itos[candidate] + '\n')