类型错误：参数"weight_decay"的多个值

Question

我正在使用一个AdamW优化器，它使用余弦衰减和预热学习调度程序。我使用TensorFlow插件库提供的AdamW优化器从头开始编写自定义调度器。

class CosineScheduler(tf.keras.optimizers.schedules.LearningRateSchedule):
def __init__(self,
learning_rate_base,
total_steps,
warmup_learning_rate=0.0,
warmup_steps=0):
self.learning_rate_base = learning_rate_base
self.total_steps = total_steps
self.warmup_learning_rate =warmup_learning_rate
self.warmup_steps = warmup_steps

def __call__(self,step):
learning_rate = 0.5 * self.learning_rate_base * (1 + tf.cos(
np.pi * 
(tf.cast(step, tf.float32) - self.warmup_steps)/ float(self.total_steps-self.warmup_steps)))
if self.warmup_steps > 0:
slope = (self.learning_rate_base - self.warmup_learning_rate) / self.warmup_steps
warmup_rate = slope * tf.cast(step, tf.float32) + self.warmup_learning_rate
learning_rate = tf.where(step < self.warmup_steps, warmup_rate, learning_rate)
lr = tf.where(step > self.total_steps, 0.0, learning_rate, name='learning_rate')
wandb.log({"lr": lr})
return lr
learning_rate = CosineScheduler(learning_rate_base=0.001, 
total_steps=23000, 
warmup_learning_rate=0.0, 
warmup_steps=1660)

loss_func = tf.keras.losses.CategoricalCrossentropy(label_smoothing=0.1)
optimizer = tfa.optimizers.AdamW(learning_rate,weight_decay=0.1)

我得到以下错误提示，它说weight_decay有多个参数

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-12-6f9fd0a9c1cb> in <module>
1 loss_func = tf.keras.losses.CategoricalCrossentropy(label_smoothing=0.1)
----> 2 optimizer = tfa.optimizers.AdamW(learning_rate,weight_decay=0.1)
/opt/conda/lib/python3.7/site-packages/typeguard/__init__.py in wrapper(*args, **kwargs)
923 
924     def wrapper(*args, **kwargs):
--> 925         memo = _CallMemo(python_func, _localns, args=args, kwargs=kwargs)
926         check_argument_types(memo)
927         retval = func(*args, **kwargs)
/opt/conda/lib/python3.7/site-packages/typeguard/__init__.py in __init__(self, func, frame_locals, args, kwargs, forward_refs_policy)
126 
127         if args is not None and kwargs is not None:
--> 128             self.arguments = signature.bind(*args, **kwargs).arguments
129         else:
130             assert frame_locals is not None, 'frame must be specified if args or kwargs is None'
/opt/conda/lib/python3.7/inspect.py in bind(*args, **kwargs)
3013         if the passed arguments can not be bound.
3014         """
-> 3015         return args[0]._bind(args[1:], kwargs)
3016 
3017     def bind_partial(*args, **kwargs):
/opt/conda/lib/python3.7/inspect.py in _bind(self, args, kwargs, partial)
2954                         raise TypeError(
2955                             'multiple values for argument {arg!r}'.format(
-> 2956                                 arg=param.name)) from None
2957 
2958                     arguments[param.name] = arg_val
TypeError: multiple values for argument 'weight_decay'

是什么引起的问题，我如何解决这个问题?

Answer 1

问题是weight_decay是tfa.optimizers.AdamW的第一个位置参数。在

optimizer = tfa.optimizers.AdamW(learning_rate,weight_decay=0.1)

传递一个位置参数，传递一个kw参数weight_decay。这会导致错误。根据文档，learning rate是第二个位置参数(尽管是可选的)，而不是第一个。

写

optimizer = tfa.optimizers.AdamW(0.1, learning_rate)

或

optimizer = tfa.optimizers.AdamW(weight_decay=0.1, learning_rate=learning_rate)

或

optimizer = tfa.optimizers.AdamW(learning_rate=learning_rate, weight_decay=0.1)