Tensorflow-Feature Column遇到的坑

在使用特征交叉函数

tf.contrib.layers.crossed_column([user_age, item_age], hash_bucket_size=100),

训练时出错：

InvalidArgumentError (see above for traceback): Dense inputs should be a matrix but received shape [1000] at position 0
[[Node: wide_and_deep_model/wide_model/weighted_sum_from_feature_columns/cross = SparseFeatureCross[N=0, dense_types=[DT_STRING, DT_STRING], hashed_output=true, internal_type=DT_STRING, num_buckets=100, out_type=DT_INT64, sparse_types=[], _device="/job:localhost/replica:0/task:0/cpu:0"](strided_slice_9, strided_slice)]]

分析：

数据维度不匹配，输入特征需要一个2D的tensor，提供的是1D的tensor，在不交叉时可能隐藏了这个问题，交叉时报错。

解决：扩展一个维度

examples_dict = {}for n, header in enumerate(COLUMNS):    examples_dict[header] = example_batch[:, n]

改为

examples_dict = {}for n, header in enumerate(COLUMNS):    examples_dict[header] = tf.expand_dims(example_batch[:, n], 1)