tensorflow连接错误:一个神秘的字符串格式化问题

Question

我试图在偏差+权重上实现L1规范.为此,我尝试将它们连接在一起并采取一种方法.

也就是说,我有偏见b1(形状[1,1]:)和重量W1(形状:) [1, xlen].所以我试着天真地沿着第一维连接:

    self.W1 = tf.Variable(tf.truncated_normal([1, self.xlen], stddev=0.1), name="weight")
    self.b1 = tf.Variable(tf.constant(0.1, shape=[1, 1]), name="bias")
    ...
 l1_penalty = tf.reduce_mean(tf.abs(tf.concat(1, (self.W1,self.b1) ) )) 

但是我得到了:

---> 53         l1_penalty = tf.reduce_mean(tf.abs(tf.concat(1, (self.W1,self.b1) ) ))
     54 
     55         tot_loss = l2_loss + self.ALPHA * l1_penalty

/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/array_ops.py in concat(concat_dim, values, name)
    304   # TODO(mrry): Change to return values?
    305   if len(values) == 1:  # Degenerate case of one tensor.
--> 306     return identity(values[0], name=name)
    307   return gen_array_ops._concat(concat_dim=concat_dim,
    308                                values=values,

/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/gen_array_ops.py in identity(input, name)
    313     A `Tensor`. Has the same type as `input`.
    314   """
--> 315   return _op_def_lib.apply_op("Identity", input=input, name=name)
    316 
    317 

/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/op_def_library.py in apply_op(self, op_type_name, g, name, **keywords)
    419             values = ops.convert_to_tensor(
    420                 values, name=input_arg.name, dtype=dtype,
--> 421                 as_ref=input_arg.is_ref)
    422           except ValueError:
    423             # What type does convert_to_tensor think it has?

/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/ops.py in convert_to_tensor(value, dtype, name, as_ref)
    522     for base_type, conversion_func in funcs_at_priority:
    523       if isinstance(value, base_type):
--> 524         ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
    525         if not isinstance(ret, Tensor):
    526           raise RuntimeError(

/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/constant_op.py in _constant_tensor_conversion_function(v, dtype, name, as_ref)
    176                                          as_ref=False):
    177   _ = as_ref
--> 178   return constant(v, dtype=dtype, name=name)
    179 
    180 

/usr/local/lib/python3.4/dist-packages/tensorflow/python/ops/constant_op.py in constant(value, dtype, shape, name)
    159   tensor_value = attr_value_pb2.AttrValue()
    160   tensor_value.tensor.CopyFrom(
--> 161       tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape))
    162   dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)
    163   const_tensor = g.create_op(

/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/tensor_util.py in make_tensor_proto(values, dtype, shape)
    382   if numpy_dtype == dtypes.string and not isinstance(values, np.ndarray):
    383     proto_values = _FlattenToStrings(values)
--> 384     tensor_proto.string_val.extend([compat.as_bytes(x) for x in proto_values])
    385     return tensor_proto
    386 

/usr/local/lib/python3.4/dist-packages/tensorflow/python/framework/tensor_util.py in <listcomp>(.0)
    382   if numpy_dtype == dtypes.string and not isinstance(values, np.ndarray):
    383     proto_values = _FlattenToStrings(values)
--> 384     tensor_proto.string_val.extend([compat.as_bytes(x) for x in proto_values])
    385     return tensor_proto
    386 

/usr/local/lib/python3.4/dist-packages/tensorflow/python/util/compat.py in as_bytes(bytes_or_text)
     41     return bytes_or_text
     42   else:
---> 43     raise TypeError('Expected binary or unicode string, got %r' % bytes_or_text)
     44 
     45 

TypeError: not all arguments converted during string formatting

添加名称参数没有帮助.

作为替代方案,tf.reduce_sum与我合作,但问题是tf.concat什么？

Answer 1

这个问题似乎是由这条tf.concat()实现线上的一个错误造成的.目前 - 与大多数其他可变参数TensorFlow运算符不同 - tf.concat()只接受张量列表,并且不能正确处理元组.如果您将代码更改为以下代码,它应该工作:

 l1_penalty = tf.reduce_mean(tf.abs(tf.concat(1, [self.W1, self.b1]))) 

在此期间,我将修复此错误并将其提交到上游,以便在下一版本中处理.