为什么Tensorflow张量索引在规范化流"prob“方法中失败

Question

我正在Tensorflow中构建一个规范化的流(分布和双射子链的连接)。以下是Bijector链的代码：

class Flow( tfb.Bijector ):

    def __init__( self, theta, a, **kwargs ):
        tfb.Bijector.__init__( self, forward_min_event_ndims = 0, **kwargs )
        bijectors = [ tfb.Tanh() ]
        self.chain = tfb.Chain( bijectors = bijectors )

    def _forward( self, z ):
        return self.chain( z )

    def _inverse( self, x ):
        result = self.chain.inverse( x ) 
        return result

    def _forward_log_det_jacobian( self, z ):
        return self.chain._forward_log_det_jacobian( z, event_ndims = 2 )

下面是我尝试测试它的方法，特别是测试基本分发和流的prob方法：

Z = tf.convert_to_tensor( [ [ [ 0.1, 0.2 ], [ 0.3, 0.4 ], [ 0.5, 0.6 ] ], 
                            [ [ 0.8, 0.7 ], [ 0.6, 0.5 ], [ 0.4, 0.3 ] ],
                            [ [ 0.4, 0.7 ], [ 0.2, 0.1 ], [ 0.8, 0.0 ] ] ] )
print( "Z", Z )
nf = Flow( 1., 2. )  # ### theta, a 
bd = tfd.MultivariateNormalDiag( loc=[0.,0.], scale_diag=[1.,1.] )
td = tfd.TransformedDistribution( bd, nf )
td.log_prob( Z )

最后一条语句失败，堆栈跟踪如下：

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
<ipython-input-29-9f91e9e1871a> in <module>()
     24 bd = tfd.MultivariateNormalDiag( loc=[0.,0], scale_diag=[1.,1.] )
     25 td = tfd.TransformedDistribution( bd, nf )
---> 26 td.prob( Z )

12 frames
/usr/local/lib/python3.7/dist-packages/tensorflow_probability/python/distributions/distribution.py in prob(self, value, name, **kwargs)
   1322         values of type `self.dtype`.
   1323     """
-> 1324     return self._call_prob(value, name, **kwargs)
   1325 
   1326   def _call_unnormalized_log_prob(self, value, name, **kwargs):

/usr/local/lib/python3.7/dist-packages/tensorflow_probability/python/distributions/distribution.py in _call_prob(self, value, name, **kwargs)
   1304     with self._name_and_control_scope(name, value, kwargs):
   1305       if hasattr(self, '_prob'):
-> 1306         return self._prob(value, **kwargs)
   1307       if hasattr(self, '_log_prob'):
   1308         return tf.exp(self._log_prob(value, **kwargs))

/usr/local/lib/python3.7/dist-packages/tensorflow_probability/python/distributions/transformed_distribution.py in _prob(self, y, **kwargs)
    371         )
    372     ildj = self.bijector.inverse_log_det_jacobian(
--> 373         y, event_ndims=event_ndims, **bijector_kwargs)
    374     if self.bijector._is_injective:  # pylint: disable=protected-access
    375       base_prob = self.distribution.prob(x, **distribution_kwargs)

/usr/local/lib/python3.7/dist-packages/tensorflow_probability/python/bijectors/bijector.py in inverse_log_det_jacobian(self, y, event_ndims, name, **kwargs)
   1318       ValueError: if the value of `event_ndims` is not valid for this bijector.
   1319     """
-> 1320     return self._call_inverse_log_det_jacobian(y, event_ndims, name, **kwargs)
   1321 
   1322   def _call_forward_log_det_jacobian(self, x, event_ndims, name, **kwargs):

/usr/local/lib/python3.7/dist-packages/tensorflow_probability/python/bijectors/bijector.py in _call_inverse_log_det_jacobian(self, y, event_ndims, name, **kwargs)
   1274               'is implemented. One or the other is required.')
   1275 
-> 1276         return self._reduce_jacobian_det_over_shape(ildj, reduce_shape)
   1277 
   1278   def inverse_log_det_jacobian(self,

/usr/local/lib/python3.7/dist-packages/tensorflow_probability/python/bijectors/bijector.py in _reduce_jacobian_det_over_shape(self, unreduced, reduce_shape)
   1531     ones = tf.ones(reduce_shape, unreduced.dtype)
   1532     reduce_dims = ps.range(-ps.size(reduce_shape), 0)
-> 1533     return tf.reduce_sum(ones * unreduced, axis=reduce_dims)
   1534 
   1535   def _parameter_control_dependencies(self, is_init):

/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/math_ops.py in binary_op_wrapper(x, y)
   1232         #   r_binary_op_wrapper use different force_same_dtype values.
   1233         x, y = maybe_promote_tensors(x, y, force_same_dtype=False)
-> 1234         return func(x, y, name=name)
   1235       except (TypeError, ValueError) as e:
   1236         # Even if dispatching the op failed, the RHS may be a tensor aware

/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/math_ops.py in _mul_dispatch(x, y, name)
   1573     return sparse_tensor.SparseTensor(y.indices, new_vals, y.dense_shape)
   1574   else:
-> 1575     return multiply(x, y, name=name)
   1576 
   1577 

/usr/local/lib/python3.7/dist-packages/tensorflow/python/util/dispatch.py in wrapper(*args, **kwargs)
    204     """Call target, and fall back on dispatchers if there is a TypeError."""
    205     try:
--> 206       return target(*args, **kwargs)
    207     except (TypeError, ValueError):
    208       # Note: convert_to_eager_tensor currently raises a ValueError, not a

/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/math_ops.py in multiply(x, y, name)
    528   """
    529 
--> 530   return gen_math_ops.mul(x, y, name)
    531 
    532 

/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/gen_math_ops.py in mul(x, y, name)
   6238       return _result
   6239     except _core._NotOkStatusException as e:
-> 6240       _ops.raise_from_not_ok_status(e, name)
   6241     except _core._FallbackException:
   6242       pass

/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/ops.py in raise_from_not_ok_status(e, name)
   6895   message = e.message + (" name: " + name if name is not None else "")
   6896   # pylint: disable=protected-access
-> 6897   six.raise_from(core._status_to_exception(e.code, message), None)
   6898   # pylint: enable=protected-access
   6899 

/usr/local/lib/python3.7/dist-packages/six.py in raise_from(value, from_value)

InvalidArgumentError: required broadcastable shapes at loc(unknown) [Op:Mul]

我无法从堆栈跟踪中找出哪里出了问题。

你能帮上忙吗？

Answer 1

通过尝试随机扰动我的代码，我能够找到解决方案:在对Bijector.__init__的调用中，将事件形状的最小维数从0改为1，并将_forward_log_det_jacobian方法的事件形状的维数从2改为1。

<opinion>为什么错误消息不能更直接地指出此问题？</opinion>

Answer 2

使用此代码我将event_ndims =3从2更改为3，然后定义init( self，theta，a，b，**kwargs )：给出3个变量而不是2，因为数组是3_3_2，所以需要在"tfd.MultivariateNormalDiag(loc=1.，2.，scale_identity_multiplier=1.，2.，3.“中进行一些更改。)

import tensorflow as tf
import tensorflow_probability as tfp
tfb = tfp.bijectors  
tfd = tfp.distributions
Z = tf.convert_to_tensor( [[[ 0.1, 0.2 ], [ 0.3, 0.4 ], [ 0.5, 0.6 ]], [[ 0.8, 0.7 ], [ 0.6, 0.5 ], [ 0.4, 0.3 ]],[[ 0.4, 0.7 ], [ 0.2, 0.1 ], [ 0.8, 0.0 ]]] )

class Flow( tfb.Bijector ):
     def __init__( self, theta, a,b, **kwargs ):
          tfb.Bijector.__init__( self, forward_min_event_ndims = 0, **kwargs )
          bijectors = [ tfb.Tanh() ]
          self.chain = tfb.Chain( bijectors = bijectors )

     def _forward( self, z ):
          return self.chain( z )

     def _inverse( self, x ):
          result = self.chain.inverse( x ) 
          return result

     def _forward_log_det_jacobian( self, z ):
          return self.chain._forward_log_det_jacobian( z, event_ndims = 3 )

 nf = Flow(1,2,3)
 bd = tfd.MultivariateNormalDiag(loc=[1., 2.], scale_identity_multiplier=[1., 2.,3.])
 td = tfd.TransformedDistribution(bd, nf)
 td.log_prob(Z)

希望它能为您工作！:)