Tensorflow LinearRegressor Feature Cannot have rank 0

Question

I am following the tutorial but failed to build a linear regressor for a dataset generated on top of y=x. Here is the last part of my code, and you can find the complete source code here if you want to reproduce my error:

_CSV_COLUMN_DEFAULTS = [[0],[0]]
_CSV_COLUMNS = ['x', 'y']

def input_fn(data_file):

    def parse_csv(value):
        print('Parsing', data_file)
        columns = tf.decode_csv(value, record_defaults=_CSV_COLUMN_DEFAULTS)
        features = dict(zip(_CSV_COLUMNS, columns))
        labels = features.pop('y')
        return features, labels

    # Extract lines from input files using the Dataset API.
    dataset = tf.data.TextLineDataset(data_file)
    dataset = dataset.map(parse_csv)

    iterator = dataset.make_one_shot_iterator()
    features, labels = iterator.get_next()
    return features, labels

x = tf.feature_column.numeric_column('x')
base_columns = [x]

model_dir = tempfile.mkdtemp()
model = tf.estimator.LinearRegressor(model_dir=model_dir,     feature_columns=base_columns)

model = model.train(input_fn=lambda: input_fn(data_file=file_path))

Somehow this code will fail with error message

ValueError: Feature (key: x) cannot have rank 0. Give: Tensor("IteratorGetNext:0", shape=(), dtype=int32, device=/device:CPU:0)

Due to the nature of tensorflow, I found it a bit hard to inspect where it really went wrong based on the given message.

Answer 1

As far as I can tell, the first dimension of the values is meant to be the batch_size. So when input_fn returns the data, it should return data as a batch.

It works once you return the data as a batch, e.g.:

dataset = tf.data.TextLineDataset(data_file)
dataset = dataset.map(parse_csv)
dataset = dataset.batch(10) # or any other batch size

Answer 2

Feature cannot have rank 0 issue occurs when we dont specify batch_size with input_fn or eval_fn or predict_fn with estimator api, below code will show how shape of tensor changes with batch_size. This code will work with TF2.0, for running this on earlier version enable eager execution (tf.enable_eager_execution()). In the below two code segment notice how shape of output tesnor changes with batch_size and without batch_size.

     ##### content of test.csv ####
        feature1, feature2,label
        234, 235, 24
        345, 345,26
        234, 345, 28
        432, 567, 29
    ########################
    import tensorflow as tf 
    tf.enable_eager_execution()

    CSV_COLUMNS= ['feature1','feature2','label']
    CSV_COLUMN_DEFAULTS = [[0.0], [0.0],[0.0]]
    def parse_csv(value):        
        columns = tf.decode_csv(value, record_defaults=CSV_COLUMN_DEFAULTS)
        features = dict(zip(CSV_COLUMNS, columns))
        labels = features.pop('label')
        return features, labels
###Without batch size 
    dataset = tf.data.TextLineDataset(filenames='./test.csv').skip(count = 1)
    dataset = dataset.map(parse_csv)
    for i in dataset:
        print(i)

# Output tensor shape here is shape=()

    ({'feature1': <tf.Tensor: id=247, shape=(), dtype=float32, numpy=234.0>, 'feature2': <tf.Tensor: id=248, shape=(), dtype=float32, numpy=235.0>}, <tf.Tensor: id=249, shape=(), dtype=float32, numpy=24.0>)
    ({'feature1': <tf.Tensor: id=253, shape=(), dtype=float32, numpy=345.0>, 'feature2': <tf.Tensor: id=254, shape=(), dtype=float32, numpy=345.0>}, <tf.Tensor: id=255, shape=(), dtype=float32, numpy=26.0>)
    ({'feature1': <tf.Tensor: id=259, shape=(), dtype=float32, numpy=234.0>, 'feature2': <tf.Tensor: id=260, shape=(), dtype=float32, numpy=345.0>}, <tf.Tensor: id=261, shape=(), dtype=float32, numpy=28.0>)
    ({'feature1': <tf.Tensor: id=265, shape=(), dtype=float32, numpy=432.0>, 'feature2': <tf.Tensor: id=266, shape=(), dtype=float32, numpy=567.0>}, <tf.Tensor: id=267, shape=(), dtype=float32, numpy=29.0>)

###With batch size

    dataset = tf.data.TextLineDataset(filenames='./test.csv').skip(count = 1)
    dataset = dataset.map(parse_csv).batch(batch_size=2)
    for i in dataset:
        print(i)
# Output tensor shape here is shape=(2,)

    ({'feature1': <tf.Tensor: id=442, shape=(2,), dtype=float32, numpy=array([234., 345.], dtype=float32)>, 'feature2': <tf.Tensor: id=443, shape=(2,), dtype=float32, numpy=array([235., 345.], dtype=float32)>}, <tf.Tensor: id=444, shape=(2,), dtype=float32, numpy=array([24., 26.], dtype=float32)>)
    ({'feature1': <tf.Tensor: id=448, shape=(2,), dtype=float32, numpy=array([234., 432.], dtype=float32)>, 'feature2': <tf.Tensor: id=449, shape=(2,), dtype=float32, numpy=array([345., 567.], dtype=float32)>}, <tf.Tensor: id=450, shape=(2,), dtype=float32, numpy=array([28., 29.], dtype=float32)>)


Using batch_size with dataset would solve the issue "Feature cannot have rank 0".