def __init__(self, vocab_size: int = 10000, wordvec_size: int = 100, hidden_size: int = 100) -> None: V, D, H = vocab_size, wordvec_size, hidden_size rn = np.random.randn # Initialize of weights embed_W = (rn(V, D) / 100).astype('f') lstm_Wx = (rn(D, 4 * H) / np.sqrt(D).astype('f')) lstm_Wh = (rn(D, 4 * H) / np.sqrt(H).astype('f')) lstm_b = np.zeros(4 * H).astype('f') affine_W = (rn(H, V) / np.sqrt(H)).astype('f') affine_b = np.zeros(V).astype('f') # Generating layers self.layers = [ TimeEmbedding(embed_W), TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True), TimeAffine(affine_W, affine_b) ] self.loss_layer = TimeSoftmaxWithLoss() self.lstm_layer = self.layers[1] # Conclude all of weights and grads as a list self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocabulary_size, wordvec_size, hidden_size): V, D, H = vocabulary_size, wordvec_size, hidden_size rn = np.random.randn # Initialize weights embed_W = (rn(V, D) / 100).astype('f') rnn_Wx = (rn(D, H) / np.sqrt(D)).astype('f') rnn_Wh = (rn(H, H) / np.sqrt(H)).astype('f') rnn_b = np.zeros(H).astype('f') affine_W = (rn(H, V) / np.sqrt(H)).astype('f') affine_b = np.zeros(V).astype('f') # generate layers self.layers = [ TimeEmbedding(embed_W), TimeRNN(rnn_Wx, rnn_Wh, rnn_b, stateful=True), TimeAffine(affine_W, affine_b) ] self.loss_layer = TimeSoftmaxWithLoss() self.rnn_layer = self.layers[1] # list all weights and gradiants self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocab_size, wordvec_size, hidden_size): V, D, H = vocab_size, wordvec_size, hidden_size rn = np.random.randn # 重みの初期化 embed_W = (rn(V, D) / 100).astype('f') rnn_Wx = (rn(D, H) / np.sqrt(D)).astype('f') rnn_Wh = (rn(H, H) / np.sqrt(H)).astype('f') rnn_b = np.zeros(H).astype('f') affine_W = (rn(H, V) / np.sqrt(H)).astype('f') affine_b = np.zeros(V).astype('f') # レイヤの生成 self.layers = [ TimeEmbedding(embed_W), TimeRNN(rnn_Wx, rnn_Wh, rnn_b, stateful=True), TimeAffine(affine_W, affine_b) ] self.loss_layer = TimeSoftmaxWithLoss() self.rnn_layer = self.layers[1] # 全ての重みと勾配をリストにまとめる self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocab_size=10000, wordvec_size=100, hidden_size=100): V, D, H = vocab_size, wordvec_size, hidden_size rn = np.random.randn # initializing weights embed_W = (rn(V, D) / 100).astype('f') lstm_Wx = (rn(D, 4 * H) / np.sqrt(D)).astype('f') lstm_Wh = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_b = np.zeros(4 * H).astype('f') affine_W = (rn(H, V) / np.sqrt(H)).astype('f') affine_b = np.zeros(V).astype('f') # generate each layers self.layers = [ TimeEmbedding(embed_W), TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True), TimeAffine(affine_W, affine_b) ] self.loss_layer = TimeSoftmaxWithLoss() self.lstm_layer = self.layers[1] # gather all weights and gradients self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocab_size, wordvec_size, hidden_size): V, D, H = vocab_size, wordvec_size, hidden_size rn = np.random.randn # Initialize of weights embed_W = (rn(V, D) / 100).astype("f") rnn_Wx = (rn(D, H) / np.sqrt(D)).astype("f") rnn_Wh = (rn(H, H) / np.sqrt(H)).astype("f") rnn_b = np.zeros(H).astype("f") affine_W = (rn(H, V) / np.sqrt(H)).astype("f") affine_b = np.zeros(V).astype("f") # Making layers self.layers = [ TimeEmbedding(embed_W), TimeRNN(rnn_Wx, rnn_Wh, rnn_b, stateful=True), TimeAffine(affine_W, affine_b), ] self.loss_layer = TimeSoftmaxWithLoss() self.rnn_layer = self.layers[1] # Conclude all of weights & grads self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocab_size, wordvec_size, hidden_size): args = vocab_size, wordvec_size, hidden_size self.encoder = AttentionEncoder(*args) self.decoder = AttentionDecoder(*args) self.softmax = TimeSoftmaxWithLoss() self.params = self.encoder.params + self.decoder.params self.grads = self.encoder.grads + self.decoder.grads
def __init__(self, vocab_size, wordvec_size, hidden_size): V, D, H = vocab_size, wordvec_size, hidden_size self.encoder = Encoder(V, D, H) self.decoder = PeekyDecoder(V, D, H) self.softmax = TimeSoftmaxWithLoss() self.params = self.encoder.params + self.decoder.params self.grads = self.encoder.grads + self.decoder.grads
def __init__(self, src_vocab_size, tgt_vocab_size, wordvec_size, hidden_size): Vs, Vt, D, H = src_vocab_size, tgt_vocab_size, wordvec_size, hidden_size self.encoder = Encoder(Vs, D, H) self.decoder = Decoder(Vt, D, H) self.softmax = TimeSoftmaxWithLoss() self.params = self.encoder.params + self.decoder.params self.grads = self.encoder.grads + self.decoder.grads
def __init__(self, vocab_size, wordvec_size, head_size, num_heads, num_encoders=3, num_decoders=3): S, D, H = vocab_size, wordvec_size, head_size rn = np.random.randn self.num_encoders = num_encoders self.num_decoders = num_decoders self.params, self.grads = [], [] # Double embed (encoder, decoder) embed_W1 = (rn(S, D) / 100).astype('f') self.e_embed = PositionalEmbedding(embed_W1) self.params += self.e_embed.params self.grads += self.e_embed.grads self.encoders, self.decoders = [], [] for _ in range(num_encoders): te = TransformerEncoder(wordvec_size=D, head_size=H, num_heads=num_heads) self.encoders.append(te) self.params += te.params self.grads += te.grads for _ in range(num_decoders): td = TransformerDecoder(wordvec_size=D, head_size=H, num_heads=num_heads) self.decoders.append(td) self.params += td.params self.grads += td.grads # 편의를 위해 linear 변수에 따로 weight 저장 self.linear = MatMul((rn(D, S) / np.sqrt(D)).astype('f')) self.params += self.linear.params self.grads += self.linear.grads # TimeSoftmaxWithLoss도 params와 grads가 있으나 사용되지 않기때문에 생략 self.softmax = TimeSoftmaxWithLoss(ignore_label=-1)
def __init__(self, vocab_size=10000, word_vec=650, hidden_size=0.5, dropout_ratio=0.5): """Rnnの改良版 LSTMの多層化(2層) Dropoutを使用(深さ方向に使用) 重み共有(EmbeddingレイヤとAffineレイヤで重み共有) """ V, D, H = vocab_size, word_vec, hidden_size rn = np.random.randn # 重みの初期化 embed_W = (rn(V, D) / 100).astype('f') lstm_Wx1 = (rn(D, 4 * H) / np.sqrt(D)).astype('f') lstm_Wh1 = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_b1 = np.zeros(4 * H).astype('f') lstm_Wx2 = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_Wh2 = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_b2 = np.zeros(4 * H).astype('f') affine_b = np.zeros(V).astype('f') # 3つの改善 self.layers = [ TimeEmbedding(embed_W), TimeDropout(dropout_ratio), TimeLSTM(lstm_Wx1, lstm_Wh1, lstm_b1, stateful=True), TimeDropout(dropout_ratio), TimeLSTM(lstm_Wx2, lstm_Wh2, lstm_b2, stateful=True), TimeDropout(dropout_ratio), TimeAffine(embed_W.T, affine_b) # 重み共有 ] self.loss_layer = TimeSoftmaxWithLoss() self.lstm_layers = [self.layers[2], self.layers[4]] self.drop_layers = [self.layers[1], self.layers[3], self.layers[5]] # 全ての重みと勾配をリストにまとめる self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocab_size=10000, wordvec_size=650, hidden_size=650, dropout_ratio=0.5): V, D, H = vocab_size, wordvec_size, hidden_size rn = np.random.randn embed_W = (rn(V, D) / 100).astype(np.float32) lstm_Wx1 = (rn(D, 4 * H) / np.sqrt(D)).astype(np.float32) lstm_Wh1 = (rn(H, 4 * H) / np.sqrt(H)).astype(np.float32) lstm_b1 = np.zeros(4 * H).astype(np.float32) lstm_Wx2 = (rn(D, 4 * H) / np.sqrt(D)).astype(np.float32) lstm_Wh2 = (rn(H, 4 * H) / np.sqrt(H)).astype(np.float32) lstm_b2 = np.zeros(4 * H).astype(np.float32) affine_b = np.zeros(V).astype(np.float32) # 3つの改善 # 1) LSTM層を重ねる # 2) Dropout層の追加 (深さ方向でLSTM層の間に追加) # 3) 重み共有 Time Embedding層とTime Affine層 @ W(V, D) self.layers = [ TimeEmbedding(embed_W), TimeDropout(dropout_ratio), TimeLSTM(lstm_Wx1, lstm_Wh1, lstm_b1, statefull=True), TimeDropout(dropout_ratio), TimeLSTM(lstm_Wx2, lstm_Wh2, lstm_b2, statefull=True), TimeDropout(dropout_ratio), TimeAffine(embed_W.T, affine_b) # embed_W(V, D)とembed_W.T(D, V)を共有 ] self.loss_layer = TimeSoftmaxWithLoss() self.lstm_layers = [self.layers[2], self.layers[4]] self.drop_layers = [self.layers[1], self.layers[3], self.layers[5]] # 重みと勾配をまとめる self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads
def __init__(self, vocab_size=10000, wordvec_size=650, hidden_size=650, dropout_ratio=0.5): V, D, H = vocab_size, wordvec_size, hidden_size rn = np.random.randn # initializing weight embed_W = (rn(V, D) / 100).astype('f') lstm_Wx1 = (rn(D, 4 * H) / np.sqrt(D)).astype('f') lstm_Wh1 = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_b1 = np.zeros(4 * H).astype('f') lstm_Wx2 = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_Wh2 = (rn(H, 4 * H) / np.sqrt(H)).astype('f') lstm_b2 = np.zeros(4 * H).astype('f') affine_b = np.zeros(V).astype('f') # generating layers self.layers = [ TimeEmbedding(embed_W), TimeDropout(dropout_ratio), TimeLSTM(lstm_Wx1, lstm_Wh1, lstm_b1, stateful=True), TimeDropout(dropout_ratio), TimeLSTM(lstm_Wx2, lstm_Wh2, lstm_b2, stateful=True), TimeDropout(dropout_ratio), TimeAffine(embed_W.T, affine_b) ] self.loss_layer = TimeSoftmaxWithLoss() self.lstm_layers = [self.layers[2], self.layers[4]] self.drop_layers = [self.layers[1], self.layers[3], self.layers[5]] # gathering weights and gradients self.params, self.grads = [], [] for layer in self.layers: self.params += layer.params self.grads += layer.grads