def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
t = K.cast(self.iterations, K.floatx()) + 1
lr = K.switch(
t <= self.warmup_steps,
self.lr * (t / self.warmup_steps),
self.lr *
(1.0 - K.minimum(t, self.decay_steps) / self.decay_steps),
)
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
if self.amsgrad:
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
else:
vhats = [K.zeros(1) for _ in params]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
p_t = m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, vhat_t))
else:
p_t = m_t / (K.sqrt(v_t) + self.epsilon)
if self.initial_weight_decay > 0.0:
if self.weight_decay_pattern is None:
p_t += self.weight_decay * p
else:
for pattern in self.weight_decay_pattern:
if pattern in p.name:
p_t += self.weight_decay * p
break
p_t = p - lr_t * p_t
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates
def test_get_token_embedding(self):
model = get_model(
token_num=200,
head_num=3,
transformer_num=2,
attention_activation='gelu',
)
embed = get_token_embedding(model)
self.assertEqual((200, 768), K.int_shape(embed))
def get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
t = K.cast(self.iterations, K.floatx()) + 1
lr = K.switch(
t <= self.warmup_steps,
self.lr * (t / self.warmup_steps),
self.min_lr + (self.lr - self.min_lr) *
(1.0 - K.minimum(t, self.decay_steps) / self.decay_steps),
)
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
(1. - K.pow(self.beta_1, t)))
ms = [
K.zeros(K.int_shape(p), dtype=K.dtype(p), name='m_{}'.format(i))
for i, p in enumerate(params)
]
vs = [
K.zeros(K.int_shape(p), dtype=K.dtype(p), name='v_{}'.format(i))
for i, p in enumerate(params)
]
if self.amsgrad:
vhats = [
K.zeros(K.int_shape(p),
dtype=K.dtype(p),
name='vh_{}'.format(i)) for i, p in enumerate(params)
]
else:
vhats = [
K.zeros(1, dtype=K.dtype(p), name='vh_{}'.format(i))
for i, p in enumerate(params)
]
self.weights = [self.iterations] + ms + vs + vhats
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
if self.amsgrad:
vhat_t = K.maximum(vhat, v_t)
p_t = m_t / (K.sqrt(vhat_t) + self.epsilon)
self.updates.append(K.update(vhat, vhat_t))
else:
p_t = m_t / (K.sqrt(v_t) + self.epsilon)
if self.initial_weight_decay > 0.0:
if self.weight_decay_pattern is None:
p_t += self.weight_decay * p
else:
for pattern in self.weight_decay_pattern:
if pattern in p.name:
p_t += self.weight_decay * p
break
mult_lr_t = lr_t
if self.lr_mult is not None:
key = p.name.split('/')
if key[0] in self.lr_mult:
#print ("going in params : ", p.name, self.lr_mult[key[0]], K.eval(lr_t))
mult_lr_t *= self.lr_mult[key[0]]
p_t = p - mult_lr_t * p_t
self.updates.append(K.update(m, m_t))
self.updates.append(K.update(v, v_t))
new_p = p_t
if getattr(p, 'constraint', None) is not None:
new_p = p.constraint(new_p)
self.updates.append(K.update(p, new_p))
return self.updates