Tensorflow 2.3: AttributeError: 'Tensor' object has no attribute 'numpy'

Question

I wanted to load the text file borrowed from here , where each line represent a json string like following:

{"overall": 2.0, "verified": true, "reviewTime": "02 4, 2014", "reviewerID": "A1M117A53LEI8", "asin": "7508492919", "reviewerName": "Sharon Williams", "reviewText": "DON'T CARE FOR IT.  GAVE IT AS A GIFT AND THEY WERE OKAY WITH IT.  JUST NOT WHAT I EXPECTED.", "summary": "CASE", "unixReviewTime": 1391472000}

I would like to extract only reviewText and overall feature from the dataset using tensorflow but facing following error.

AttributeError: in user code:

    <ipython-input-4-419019a35c5e>:9 None  *
        line_dataset = line_dataset.map(lambda row: transform(row))
    <ipython-input-4-419019a35c5e>:2 transform  *
        str_example = example.numpy().decode("utf-8")

    AttributeError: 'Tensor' object has no attribute 'numpy'

My code snippet looks like following:

def transform(example):
    str_example = example.numpy().decode("utf-8")
    json_example = json.loads(str_example)
    overall = json_example.get('overall', None)
    text = json_example.get('reviewText', None)
    return (overall, text)

line_dataset = tf.data.TextLineDataset(filenames = [file_path])
line_dataset = line_dataset.map(lambda row: transform(row))
for example in line_dataset.take(5):
    print(example)

I am using tensorflow 2.3.0.

Answer 1

The input pipeline of a dataset is always traced into a graph (as if you used @tf.function ) to make it faster, which means, among other things, that you cannot use .numpy() . You can however use tf.numpy_function to access the data as a NumPy array within the graph:

def transform(example):
    # example will now by a NumPy array
    str_example = example.decode("utf-8")
    json_example = json.loads(str_example)
    overall = json_example.get('overall', None)
    text = json_example.get('reviewText', None)
    return (overall, text)

line_dataset = tf.data.TextLineDataset(filenames = [file_path])
line_dataset = line_dataset.map(
    lambda row: tf.numpy_function(transform, row, (tf.float32, tf.string)))
for example in line_dataset.take(5):
    print(example)

Answer 2

A bit wordy, but try it like this:

def transform(example):     
    str_example = example.numpy().decode("utf-8")     
    json_example = json.loads(str_example)     
    overall = json_example.get('overall', None)     
    text = json_example.get('reviewText', None)     
    return (overall, text)  

line_dataset = tf.data.TextLineDataset(filenames = [file_path]) 
line_dataset = line_dataset.map(
    lambda input:     
        tf.py_function(transform, [input], (tf.float32, tf.string))
)  
for example in line_dataset.take(5):     
    print(example)

This particular snippet works for any python function, not only the for numpy functions. So, if you need functions like print , input and so on, you can use this. You don't have to know all the details, but if you are interested, please ask me. :)