我尝试使用Tensorflow的Dataset API从多个CSV文件中读取数据。
以下版本的代码运行正常：
record_defaults = [[""], [0.0], [0.0], [0.0], [0.0], [0.0], [0.]]def decode_csv(line): col1, col2, col3, col4, col5, col6, col7 = tf.decode_csv(line, record_defaults) features = tf.stack([col2, col3, col4, col5, col6]) labels = tf.stack([col7]) return features, labelsfilenames = tf.placeholder(tf.string, shape=[None])dataset5 = tf.data.Dataset.from_tensor_slices(filenames)dataset5 = dataset5.flat_map(lambda filename: tf.data.TextLineDataset(filename).skip(1).map(decode_csv))dataset5 = dataset5.shuffle(buffer_size=1000)dataset5 = dataset5.batch(7)iterator5 = dataset5.make_initializable_iterator()但我想让它更加动态，因为不同项目中的列数（特征数）可能会变化。然而，当我更改代码如下时，它就不工作了。在这个问题上花了大量时间也没有解决…
record_defaults = [[""], [0.0], [0.0], [0.0], [0.0], [0.0], [0.]]def decode_csv(line): csv_columns = tf.decode_csv(line, record_defaults) labels = csv_columns[-1] # 最后一列是标签 del csv_columns[-1] # 删除最后一列 del csv_columns[0] # 删除第一列，因为它不是特征 features = csv_columns return features, labelsfilenames = tf.placeholder(tf.string, shape=[None])dataset5 = tf.data.Dataset.from_tensor_slices(filenames)dataset5 = dataset5.flat_map(lambda filename: tf.data.TextLineDataset(filename).skip(1).map(decode_csv))dataset5 = dataset5.shuffle(buffer_size=1000)dataset5 = dataset5.batch(7)iterator5 = dataset5.make_initializable_iterator()运行上述第二个版本时，我得到了以下错误…也许更有经验的人能立即看出问题所在..?
---------------------------------------------------------------------------AttributeError Traceback (most recent call last)<ipython-input-21-92ea8cc44da0> in <module>() 18 filenames = tf.placeholder(tf.string, shape=[None]) 19 dataset5 = tf.data.Dataset.from_tensor_slices(filenames)---> 20 dataset5 = dataset5.flat_map(lambda filename: tf.data.TextLineDataset(filename).skip(1).map(decode_csv)) 21 dataset5 = dataset5.shuffle(buffer_size=1000) 22 dataset5 = dataset5.batch(7)~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in flat_map(self, map_func) 799 Dataset: A `Dataset`. 800 """--> 801 return FlatMapDataset(self, map_func) 802 803 def interleave(self, map_func, cycle_length, block_length=1):~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in __init__(self, input_dataset, map_func) 1676 1677 self._map_func = tf_map_func-> 1678 self._map_func.add_to_graph(ops.get_default_graph()) 1679 1680 def _as_variant_tensor(self):~/.local/lib/python3.5/site-packages/tensorflow/python/framework/function.py in add_to_graph(self, g) 484 def add_to_graph(self, g): 485 """Adds this function into the graph g."""--> 486 self._create_definition_if_needed() 487 488 # Adds this function into 'g'.~/.local/lib/python3.5/site-packages/tensorflow/python/framework/function.py in _create_definition_if_needed(self) 319 """Creates the function definition if it's not created yet.""" 320 with context.graph_mode():--> 321 self._create_definition_if_needed_impl() 322 323 def _create_definition_if_needed_impl(self):~/.local/lib/python3.5/site-packages/tensorflow/python/framework/function.py in _create_definition_if_needed_impl(self) 336 # Call func and gather the output tensors. 337 with vs.variable_scope("", custom_getter=temp_graph.getvar):--> 338 outputs = self._func(*inputs) 339 340 # There is no way of distinguishing between a function not returning~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in tf_map_func(*args) 1664 dataset = map_func(*nested_args) 1665 else:-> 1666 dataset = map_func(nested_args) 1667 1668 if not isinstance(dataset, Dataset):<ipython-input-21-92ea8cc44da0> in <lambda>(filename) 18 filenames = tf.placeholder(tf.string, shape=[None]) 19 dataset5 = tf.data.Dataset.from_tensor_slices(filenames)---> 20 dataset5 = dataset5.flat_map(lambda filename: tf.data.TextLineDataset(filename).skip(1).map(decode_csv)) 21 dataset5 = dataset5.shuffle(buffer_size=1000) 22 dataset5 = dataset5.batch(7)~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in map(self, map_func, num_parallel_calls) 784 """ 785 if num_parallel_calls is None:--> 786 return MapDataset(self, map_func) 787 else: 788 return ParallelMapDataset(self, map_func, num_parallel_calls)~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in __init__(self, input_dataset, map_func) 1587 1588 self._map_func = tf_map_func-> 1589 self._map_func.add_to_graph(ops.get_default_graph()) 1590 1591 def _as_variant_tensor(self):~/.local/lib/python3.5/site-packages/tensorflow/python/framework/function.py in add_to_graph(self, g) 484 def add_to_graph(self, g): 485 """Adds this function into the graph g."""--> 486 self._create_definition_if_needed() 487 488 # Adds this function into 'g'.~/.local/lib/python3.5/site-packages/tensorflow/python/framework/function.py in _create_definition_if_needed(self) 319 """Creates the function definition if it's not created yet.""" 320 with context.graph_mode():--> 321 self._create_definition_if_needed_impl() 322 323 def _create_definition_if_needed_impl(self):~/.local/lib/python3.5/site-packages/tensorflow/python/framework/function.py in _create_definition_if_needed_impl(self) 336 # Call func and gather the output tensors. 337 with vs.variable_scope("", custom_getter=temp_graph.getvar):--> 338 outputs = self._func(*inputs) 339 340 # There is no way of distinguishing between a function not returning~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in tf_map_func(*args) 1575 self._output_classes = sparse.get_classes(ret) 1576 self._output_shapes = nest.pack_sequence_as(-> 1577 ret, [t.get_shape() for t in nest.flatten(ret)]) 1578 self._output_types = nest.pack_sequence_as( 1579 ret, [t.dtype for t in nest.flatten(ret)])~/.local/lib/python3.5/site-packages/tensorflow/python/data/ops/dataset_ops.py in <listcomp>(.0) 1575 self._output_classes = sparse.get_classes(ret) 1576 self._output_shapes = nest.pack_sequence_as(-> 1577 ret, [t.get_shape() for t in nest.flatten(ret)]) 1578 self._output_types = nest.pack_sequence_as( 1579 ret, [t.dtype for t in nest.flatten(ret)])AttributeError: 'list' object has no attribute 'get_shape'补充说明：
以下代码同样有效。
feature_names = ['f0','f1','f2','f3','f4','f5']record_defaults = [[""], [0.0], [0.0], [0.0], [0.0], [0.0], [0.]]def decode_csv(line): parsed_line = tf.decode_csv(line, record_defaults) # => tensor label = parsed_line[-1] del parsed_line[-1] features = parsed_line d = dict(zip(feature_names,features)),label return dfilenames = tf.placeholder(tf.string, shape=[None])dataset5 = tf.data.Dataset.from_tensor_slices(filenames)dataset5 = dataset5.flat_map(lambda filename: tf.data.TextLineDataset(filename).skip(1).map(decode_csv))dataset5 = dataset5.shuffle(buffer_size=1000)dataset5 = dataset5.batch(7)iterator5 = dataset5.make_initializable_iterator()但现在decode_csv函数返回的是（特征名，特征值）对的字典。为什么有人会想要从这个函数返回一个字典？这不是让像前向传播等计算的向量化变得非常困难吗？
回答：
已解决。下面是工作版本。为了节省空间，我不复制整个内容。在Excel文件中，第一列不是特征，只是训练样本ID。最后一列仅是标签。使用tf.stack(…)函数堆叠特征解决了这个问题。
feature_names = ['f1','f2','f3','f4','f5']record_defaults = [[""], [0.0], [0.0], [0.0], [0.0], [0.0], [0.0]]def decode_csv(line): parsed_line = tf.decode_csv(line, record_defaults) label = parsed_line[-1] del parsed_line[-1] del parsed_line[0] features = tf.stack(parsed_line) # 新增行 d = features, label return dfilenames = tf.placeholder(tf.string, shape=[None])dataset5 = tf.data.Dataset.from_tensor_slices(filenames)dataset5 = dataset5.flat_map(lambda filename: tf.data.TextLineDataset(filename).skip(1).map(decode_csv))dataset5 = dataset5.shuffle(buffer_size=1000)dataset5 = dataset5.batch(7)iterator5 = dataset5.make_initializable_iterator()