What's the correct way to use tf.data.Dataset.map?

Question

I have audio files and I'd like to make a tf.DataSet from their audio content (i.e. each audio file in the dataset should be represented as a vector of float values)

Here's my code

def convert_audio_file_to_numpy_array(filepath): sample_rate = sox.file_info.sample_rate(filepath) audio, sr = librosa.load(filepath, sr=sample_rate) array = np.asarray(audio) return array filenames_ds = tf.data.Dataset.from_tensor_slices(input_filepaths) waveforms_ds = filenames_ds.map(convert_audio_file_to_numpy_array, num_parallel_calls=tf.data.AUTOTUNE)

This produces this error: TypeError: stat: path should be string, bytes, os.PathLike or integer, not Tensor

I'm using DataSet's map function following the pattern in this official tutorial (see the call to files_ds.map). In it, the function that map uses takes a filepath.

What am I doing differently to the official tutorial?

The problem is that the function sox.file_info.sample_rate expects a string or a Path, but the values in your dataset are Tensors (of type string). If you want to load things from a string Tensor, you need to use tf.io functions or similar. — Lescurel
– Lescurel, Commented Feb 12, 2021 at 8:01

Lescurel · Accepted Answer · 2021-02-12 08:32:36Z

The problem is that the function def sample_rate(input_filepath: Union[str, Path]) -> float: expects either a string of a pathlib.Path, and you are providing a Tensor. (The elements of your filename_ds are Tensors of type string).

In the tensorflow tutorial, they load the data with a tensorflow functions that expects a Tensor of type string. You should check if you can load your files with tf.audio native functions.

Otherwise, a common workaround is to use a generator with tf.data.Dataset.from_generator, something akin to the following solution:

def generator_func(list_of_path): def convert_audio_file_to_numpy_array(filepath): sample_rate = sox.file_info.sample_rate(filepath) audio, sr = librosa.load(filepath, sr=sample_rate) array = np.asarray(audio) return array for path in list_of_path: yield convert_audio_file_to_numpy_array(path) ds = tf.data.Dataset.from_generator(generator_func, output_types=tf.float32)

Collectives™ on Stack Overflow

What's the correct way to use tf.data.Dataset.map?

1 Answer 1

Comments

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Related