class Encoder:
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
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')
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=False)
self.params = self.embed.params + self.lstm.params
self.grads = self.embed.grads + self.lstm.grads
self.hs = None
def forward(self, xs):
xs = self.embed.forward(xs)
hs = self.lstm.forward(xs)
self.hs = hs
return hs[:, -1, :]
def backward(self, dh):
dhs = np.zeros_like(self.hs)
dhs[:, -1, :] = dh
dout = self.lstm.backward(dhs)
dout = self.embed.backward(dout)
return dout
class Decoder:
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
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')
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True)
self.affine = TimeAffine(affine_W, affine_b)
self.params, self.grads = [], []
for layer in (self.embed, self.lstm, self.affine):
self.params += layer.params
self.grads += layer.grads
def forward(self, xs, h):
self.lstm.set_state(h)
out = self.embed.forward(xs)
out = self.lstm.forward(out)
score = self.affine.forward(out)
return score
def backward(self, dscore):
dout = self.affine.backward(dscore)
dout = self.lstm.backward(dout)
dout = self.embed.backward(dout)
dh = self.lstm.dh
return dh
def generate(self, h, start_id, sample_size):
'''
h: Encoderから受け取る隠れ状態
start_id: 最初に与える文字ID
sample_size: 生成する文字数
'''
sampled = []
sample_id = start_id
self.lstm.set_state(h)
for _ in range(sample_size):
x = np.array(sample_id).reshape((1, 1))
out = self.embed.forward(x)
out = self.lstm.forward(out)
score = self.affine.forward(out)
sample_id = np.argmax(score.flatten())
sampled.append(int(sample_id))
return sampled
class AttentionDecoder:
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
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(2 * H, V) / np.sqrt(H)).astype('f')
affine_b = np.zeros(V).astype('f')
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True)
self.attention = TimeAttention()
self.affine = TimeAffine(affine_W, affine_b)
layers = [self.embed, self.lstm, self.attention, self.affine]
self.params, self.grads = [], []
for layer in layers:
self.params += layer.params
self.grads += layer.grads
def forward(self, xs, enc_hs):
h = enc_hs[:, -1]
self.lstm.set_state(h)
out = self.embed.forward(xs)
dec_hs = self.lstm.forward(out)
c = self.attention.forward(enc_hs, dec_hs)
out = np.concatenate((c, dec_hs), axis=2)
score = self.affine.forward(out)
return score
def backward(self, dscore):
dout = self.affine.backward(dscore)
N, T, H2 = dout.shape
H = H2 // 2
dc, ddec_hs0 = dout[:, :, :H], dout[:, :, H:]
denc_hs, ddec_hs1 = self.attention.backward(dc)
ddec_hs = ddec_hs0 + ddec_hs1
dout = self.lstm.backward(ddec_hs)
dh = self.lstm.dh
denc_hs[:, -1] += dh
self.embed.backward(dout)
return denc_hs
def generate(self, enc_hs, start_id, eos_id, max_len=40):
sampled = []
sample_id = start_id
h = enc_hs[:, -1]
self.lstm.set_state(h)
while sample_id != eos_id:
if len(sampled) > max_len:
break
x = np.array([int(sample_id)]).reshape((1, 1))
out = self.embed.forward(x)
dec_hs = self.lstm.forward(out)
c = self.attention.forward(enc_hs, dec_hs)
out = np.concatenate((c, dec_hs), axis=2)
score = self.affine.forward(out)
sample_id = np.argmax(score.flatten())
sampled.append(sample_id)
return sampled
class PeekyDecoder:
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
embed_W = (rn(V, D) / 100).astype('f')
lstm_Wx = (rn(H + D, 4 * H) / np.sqrt(H + 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 + H, V) / np.sqrt(H + H)).astype('f')
affine_b = np.zeros(V).astype('f')
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, stateful=True)
self.affine = TimeAffine(affine_W, affine_b)
self.params, self.grads = [], []
for layer in (self.embed, self.lstm, self.affine):
self.params += layer.params
self.grads += layer.grads
self.cache = None
def forward(self, xs, h):
N, T = xs.shape
N, H = h.shape
self.lstm.set_state(h)
out = self.embed.forward(xs)
hs = np.repeat(h, T, axis=0).reshape(N, T, H)
out = np.concatenate((hs, out), axis=2)
out = self.lstm.forward(out)
out = np.concatenate((hs, out), axis=2)
score = self.affine.forward(out)
self.cache = H
return score
def backward(self, dscore):
H = self.cache
dout = self.affine.backward(dscore)
dout, dhs0 = dout[:, :, H:], dout[:, :, :H]
dout = self.lstm.backward(dout)
dembed, dhs1 = dout[:, :, H:], dout[:, :, :H]
self.embed.backward(dembed)
dhs = dhs0 + dhs1
dh = self.lstm.dh + np.sum(dhs, axis=1)
return dh
def generate(self, h, start_id, sample_size):
'''
h: Encoderから受け取る隠れ状態
start_id: 最初に与える文字ID
sample_size: 生成する文字数
'''
sampled = []
sample_id = start_id
self.lstm.set_state(h)
for _ in range(sample_size):
x = np.array(sample_id).reshape((1, 1))
out = self.embed.forward(x)
out = self.lstm.forward(out)
score = self.affine.forward(out)
sample_id = np.argmax(score.flatten())
sampled.append(int(sample_id))
return sampled
class PeekyDecoder:
def __init__(self, vocab_size, wordvec_size, hidden_size):
V, D, H = vocab_size, wordvec_size, hidden_size
rn = np.random.randn
embed_W = (rn(V, D) / 100).astype(np.float32)
lstm_Wx = (rn(H + D, 4 * H) / np.sqrt(H + D)).astype(np.float32)
lstm_Wh = (rn(H, 4 * H) / np.sqrt(H)).astype(np.float32)
lstm_b = np.zeros(4 * H, dtype=np.float32)
affine_W = (rn(H + H, V) / np.sqrt(H + H)).astype(np.float32)
affine_b = np.zeros(V, dtype=np.float32)
self.embed = TimeEmbedding(embed_W)
self.lstm = TimeLSTM(lstm_Wx, lstm_Wh, lstm_b, statefull=True)
self.affine = TimeAffine(affine_W, affine_b)
self.params, self.grads = [], []
for layer in (self.embed, self.lstm, self.affine):
self.params += layer.params
self.grads += layer.grads
self.cache = None
def forward(self, xs, h):
N, T = xs.shape
N, H = h.shape
self.lstm.set_state(h)
out = self.embed.forward(xs)
# TimeLSTMへの入力を拡張する
hs = np.repeat(h, T, axis=0).reshape(N, T, H)
out = np.concatenate((hs, out), axis=2) # (N, T, H + D)
# TimeAffineへの入力を拡張する
out = self.lstm.forward(out)
out = np.concatenate((hs, out), axis=2)
score = self.affine.forward(out)
self.cache = H
return score
def backward(self, dscore):
H = self.cache
dout = self.affine.backward(dscore)
dout, dhs0 = dout[:, :, H:], dout[:, :, :H]
dout = self.lstm.backward(dout)
dembed, dhs1 = dout[:, :, H:], dout[:, :, :H]
self.embed.backward(dembed)
dhs = dhs0 + dhs1
dh = self.lstm.dh + np.sum(dhs, axis=1) # 時間方向に集約
return dh
def generate(self, h, start_id, sample_size):
sampled = []
char_id = start_id
self.lstm.set_state(h)
H = h.shape[1]
peeky_h = h.reshape(1, 1, H)
for _ in range(sample_size):
x = np.array([char_id]).reshape((1, 1))
out = self.embed.forward(x)
out = np.concatenate((peeky_h, out), axis=2)
out = self.lstm.forward(out)
out = np.concatenate((peeky_h, out), axis=2)
score = self.affine.forward(out)
char_id = np.argmax(score.flatten())
sampled.append(char_id)
return sampled