def eval(self, executor, eval_program=None):
if eval_program is None:
eval_program = Program()
block = eval_program.current_block()
with program_guard(main_program=eval_program):
total_distance = _clone_var_(block, self.total_distance)
seq_num = _clone_var_(block, self.seq_num)
instance_error = _clone_var_(block, self.instance_error)
seq_num = layers.cast(x=seq_num, dtype='float32')
instance_error = layers.cast(x=instance_error, dtype='float32')
avg_distance = layers.elementwise_div(x=total_distance, y=seq_num)
avg_instance_error = layers.elementwise_div(
x=instance_error, y=seq_num)
result = executor.run(
eval_program, fetch_list=[avg_distance, avg_instance_error])
return np.array(result[0]), np.array(result[1])
def eval(self, executor, eval_program=None):
if eval_program is None:
eval_program = Program()
block = eval_program.current_block()
with program_guard(main_program=eval_program):
total_distance = _clone_var_(block, self.total_distance)
seq_num = _clone_var_(block, self.seq_num)
instance_error = _clone_var_(block, self.instance_error)
seq_num = layers.cast(x=seq_num, dtype='float32')
instance_error = layers.cast(x=instance_error, dtype='float32')
avg_distance = layers.elementwise_div(x=total_distance, y=seq_num)
avg_instance_error = layers.elementwise_div(x=instance_error,
y=seq_num)
result = executor.run(
eval_program, fetch_list=[avg_distance, avg_instance_error])
return np.array(result[0]), np.array(result[1])
def eval(self, executor, eval_program=None):
if eval_program is None:
eval_program = Program()
block = eval_program.current_block()
kwargs = {'main_program': eval_program}
total = _clone_var_(block, self.total)
correct = _clone_var_(block, self.correct)
total = layers.cast(total, dtype='float32', **kwargs)
correct = layers.cast(correct, dtype='float32', **kwargs)
out = layers.elementwise_div(x=correct, y=total, **kwargs)
return np.array(executor.run(eval_program, fetch_list=[out])[0])
def eval(self, executor, eval_program=None):
if eval_program is None:
eval_program = Program()
block = eval_program.current_block()
kwargs = {'main_program': eval_program}
total = _clone_var_(block, self.total)
correct = _clone_var_(block, self.correct)
total = layers.cast(total, dtype='float32', **kwargs)
correct = layers.cast(correct, dtype='float32', **kwargs)
out = layers.elementwise_div(x=correct, y=total, **kwargs)
return np.array(executor.run(eval_program, fetch_list=[out])[0])
def _add_average_apply_op(self, block, param_grad):
param = block.clone_variable(param_grad[0])
grad = block.clone_variable(param_grad[1])
sum_1 = block.clone_variable(self._get_accumulator('sum_1', param))
sum_2 = block.clone_variable(self._get_accumulator('sum_2', param))
sum_3 = block.clone_variable(self._get_accumulator('sum_3', param))
num_accumulates = block.clone_variable(
self._get_accumulator('num_accumulates', param))
old_num_accumulates = block.clone_variable(
self._get_accumulator('old_num_accumulates', param))
num_updates = block.clone_variable(
self._get_accumulator('num_updates', param))
# backup param value to grad
layers.assign(input=param, output=grad)
# param = (sum_1 + sum_2 + sum_3) / (num_accumulates + old_num_accumulates)
tmp = layers.sum(x=[num_accumulates, old_num_accumulates])
sum = layers.sum(x=[sum_1, sum_2, sum_3])
tmp = layers.cast(x=tmp, dtype='float32')
sum = layers.cast(x=sum, dtype='float32')
layers.elementwise_div(x=sum, y=tmp, out=param)
def create_operators(self, param, grad):
group_scale_name = self.group_name + "_scale"
if group_scale_name not in self.context:
group_norm_var = layers.sums(input=self.context[self.group_name])
layers.sqrt(x=group_norm_var, out=group_norm_var)
clip_var = self.context[self.group_name + "_clip"]
group_scale_var = layers.elementwise_div(x=clip_var,
y=layers.elementwise_max(
x=clip_var,
y=group_norm_var))
assert group_scale_var.shape == (1L, )
self.context[group_scale_name] = group_scale_var
new_grad = layers.elementwise_mul(x=grad,
y=self.context[group_scale_name])
return param, new_grad
def create_operators(self, param, grad):
group_scale_name = self.group_name + "_scale"
if group_scale_name not in self.context:
group_norm_var = layers.sums(input=self.context[self.group_name])
layers.sqrt(x=group_norm_var, out=group_norm_var)
clip_var = self.context[self.group_name + "_clip"]
group_scale_var = layers.elementwise_div(
x=clip_var,
y=layers.elementwise_max(
x=clip_var, y=group_norm_var))
assert group_scale_var.shape == (1L, )
self.context[group_scale_name] = group_scale_var
new_grad = layers.elementwise_mul(
x=grad, y=self.context[group_scale_name])
return param, new_grad