問題文
問題の概要
RNN を双方向しました。具体的には bidirectional=True にし、続く層の hidden_size を 2 倍にしています。なお実装時には『現場で使える！PyTorch開発入門 深層学習モデルの作成とアプリケーションへの実装』（翔泳社）のサンプルコードを一部流用しました。
# ref: https://www.shoeisha.co.jp/book/detail/9784798157184 import re from collections import defaultdict import joblib import pandas as pd import torch from gensim.models import KeyedVectors from torch import nn, optim from torch.utils.data import DataLoader, Dataset from tqdm import tqdm def cleanText(text): remove_marks_regex = re.compile("[,\.\(\)\[\]\*:;]|<.*?>") shift_marks_regex = re.compile("([?!])") # !?以外の記号の削除 text = remove_marks_regex.sub("", text) # !?と単語の間にスペースを挿入 text = shift_marks_regex.sub(r" \1 ", text) return text def list2tensor(token_idxes, max_len=20, padding=True): if len(token_idxes) > max_len: token_idxes = token_idxes[:max_len] n_tokens = len(token_idxes) if padding: token_idxes = token_idxes + [0] * (max_len - len(token_idxes)) return torch.tensor(token_idxes, dtype=torch.int64), n_tokens class RNN(nn.Module): def __init__(self, num_embeddings, embedding_dim=300, hidden_size=300, output_size=1, num_layers=1, dropout=0.2): super().__init__() # self.emb = nn.Embedding(num_embeddings, embedding_dim, # padding_idx=0) model = KeyedVectors.load_word2vec_format('ch07/GoogleNews-vectors-negative300.bin', binary=True) weights = torch.FloatTensor(model.vectors) self.emb = nn.Embedding.from_pretrained(weights) self.lstm = nn.LSTM(embedding_dim, hidden_size, num_layers, batch_first=True, dropout=dropout, bidirectional=True) self.linear = nn.Linear(hidden_size * 2, output_size) def forward(self, x, h0=None, n_tokens=None): # IDをEmbeddingで多次元のベクトルに変換する # xは(batch_size, step_size) # -> (batch_size, step_size, embedding_dim) x = self.emb(x) # 初期状態h0と共にRNNにxを渡す # xは(batch_size, step_size, embedding_dim) # -> (batch_size, step_size, hidden_dim) x, h = self.lstm(x, h0) # 最後のステップのみ取り出す # xは(batch_size, step_size, hidden_dim) # -> (batch_size, 1) if n_tokens is not None: # 入力のもともとの長さがある場合はそれを使用する x = x[list(range(len(x))), n_tokens - 1, :] else: # なければ単純に最後を使用する x = x[:, -1, :] # 取り出した最後のステップを線形層に入れる x = self.linear(x) # 余分な次元を削除する # (batch_size, 1) -> (batch_size, ) # x = x.squeeze() return x class TITLEDataset(Dataset): def __init__(self, section='train'): X_train = pd.read_table(f'ch06/{section}.txt', header=None) use_cols = ['TITLE', 'CATEGORY'] X_train.columns = use_cols d = defaultdict(int) for text in X_train['TITLE']: text = cleanText(text) for word in text.split(): d[word] += 1 dc = sorted(d.items(), key=lambda x: x[1], reverse=True) words = [] idx = [] for i, a in enumerate(dc, 1): words.append(a[0]) if a[1] < 2: idx.append(0) else: idx.append(i) self.word2token = dict(zip(words, idx)) self.data = (X_train['TITLE'].apply(lambda x: list2tensor( [self.word2token[word] if word in self.word2token.keys() else 0 for word in cleanText(x).split()]))) y_train = pd.read_table(f'ch06/{section}.txt', header=None)[1].values self.labels = y_train @property def vocab_size(self): return len(self.word2token) def __len__(self): return len(self.labels) def __getitem__(self, idx): data, n_tokens = self.data[idx] label = self.labels[idx] return data, label, n_tokens def eval_net(net, data_loader, device='cpu'): net.eval() ys = [] ypreds = [] for x, y, nt in data_loader: x = x.to(device) y = y.to(device) nt = nt.to(device) with torch.no_grad(): y_pred = net(x, n_tokens=nt) # print(f'test loss: {loss_fn(y_pred, y.long()).item()}') _, y_pred = torch.max(y_pred, 1) ys.append(y) ypreds.append(y_pred) ys = torch.cat(ys) ypreds = torch.cat(ypreds) print(f'test acc: {(ys == ypreds).sum().item() / len(ys)}') return if __name__ == "__main__": device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') batch_size = 640 train_data = TITLEDataset(section='train') train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True, num_workers=4) test_data = TITLEDataset(section='test') test_loader = DataLoader(test_data, batch_size=batch_size, shuffle=False, num_workers=4) net = RNN(train_data.vocab_size + 1, num_layers=2, output_size=4) net = net.to(device) loss_fn = nn.CrossEntropyLoss() optimizer = optim.SGD(net.parameters(), lr=0.01) for epoch in tqdm(range(10)): losses = [] net.train() for x, y, nt in train_loader: x = x.to(device) y = y.to(device) nt = nt.to(device) y_pred = net(x, n_tokens=nt) loss = loss_fn(y_pred, y.long()) net.zero_grad() loss.backward() optimizer.step() losses.append(loss.item()) _, y_pred_train = torch.max(y_pred, 1) # print(f'train loss: {loss.item()}') # print(f'train acc: {(y_pred_train == y).sum().item() / len(y)}') eval_net(net, test_loader, device)
