TFRecorder

The goal of this project is to ease the creation of tfrecord files for training, evaluation (and optionally testing) a Tensorflow model, hiding boilerplate code under the hood.

Usage

Subclassing Example

• Using the @tfrecordable decorator
• Overriding the load method
• Overriding the split method (optionally)

Instantiating Example subclass

• Overriding the from_csv_row method
• Overriding the to_csv_row method
• Preparing the examples.csv file

Generating the tfrecord files

Generating a tf.data.Dataset

The Example class shall be subclassed to represent the data of the task at hand. Each instance of your Example subclass will then be processed by the factory and converted into one (or various) tf.Example instance(s) and stored in tfrecord files.

Subclassing Example

Using tfrecords files probably means that your data can not fit into memory. Your subclass should therefore be instantiated only with its lightweight attributes, and defer loading of its memory consuming attributes for a later stage (see the load method below).

For instance, consider a classification task where a float array is used as source data and an integer is used as a target label:

import numpy as np
import os

from tfrecorder.helpers.decorator import tfrecordable
from tfrecorder.helpers.marshaller import Example

class ToyExample(Example):

    def __init__(self, name, label):

        super(ToyExample, self).__init__()
        self.name = name
        self._label = label # this is an int, so we assign it when instantiating
        self._data = None # this is a float array, so we defer loading

Using @tfrecordable decorator

The @tfrecordable decorator indicates which attributes of your Example subclass shall be serialized into tfrecords files. It is used in the same fashion as the built-in @property decorator, but is passed the type of the corresponding serializable attribute.

For our classification task, only the _label and _data attributes will be serialized into tfrecords files, but not the name attribute:

    # CAUTION: the attributes must be declared in the same order than expected when parsing in tf.Dataset
    @tfrecordable(dtype=Example.Field.TYPE_INT32)
    def label(self):
        return self._label
    
    @label.setter
    def label(self, val):
        self._label = val
    
    @tfrecordable(dtype=Example.Field.TYPE_ARRAY_FLOAT32)
    def data(self):
        return self._data
        
    @data.setter
    def data(self, val):
        self._data = val

The dtype parameter of the @tfrecordable decorator is used to tell Tensorflow which data type to use when serializing and de-serializing. The supported types are:

TYPE_BOOL	TYPE_ARRAY_INT32	TYPE_INT32	TYPE_FLOAT	TYPE_STRING
	TYPE_ARRAY_FLOAT32	TYPE_INT64	TYPE_DOUBLE

You should use TYPE_ARRAY_INT32 or TYPE_ARRAY_FLOAT32 when serializing a numpy array. Note that if will be stored as tf.int32 or tf.float32 in the tfrecord file.

Overriding the load method

If your Example subclass has some deferred loading of data, it must override the load instance method.

This method will be called just before your example instance is serialized into a tfrecord file, and should implement the logic needed to load your data into memory, and optionally pre-process it if needed. Pass it all parameters required to load and optionally pre-process your data.

Finally, make sure that your data is assigned to its corresponding @tfrecordable attribute.

Once saved to the file, all memory used during serialization will be released.

For our classification task, we could have:

    def load(self, src_data_dirpath=None, normalize=False, **kwargs):
        """
        Loads the data in memory, and possibly pre-process it.
        """
        data = np.load(os.path.join(src_data_dirpath, '%s.npy' % self.name))
            
        if normalize:
            data /= np.max(data)

        self.data = data

Overriding the split method

If the data of your Example subclass shall be split into chunks before serialization, it must override the split instance method, and wrap each chunk into an instance of your Example subclass. Otherwise, don't override the split instance method.

For our classification task, we could write:

    def split(self, chunk_size_in_bins=5, **kwargs):
        """
	    Splits this example data into chunks, and create one ToyExample object per chunk.

	    Returns:
		    examples: list, of ToyExample objects.

	    """
        src_data = self.src_data

        chunked_examples = []
        for i in range(0, src_data.shape[0], chunk_size_in_bins):
            chunked_example = ToyExample('%s_%d' % (self.name, self._label))
            chunked_example.src_data = src_data[i:i+chunk_size_in_bins]
            chunked_examples.append(chunked_example)

        return chunked_examples

Instantiating Example subclass

At this point, you can instantiate your subclass for your data and pass it to one of the factory's methods. It is however simpler to let TFRecorder handle it for you.

Overriding the from_csv_row method

Your Example subclass should override from_csv_row static method to return a new instance of your subclass. The row argument will be passed by a csv reader, and is thus a list of strings.

    @staticmethod
    def from_csv_row(row, **kwargs):
        """
        This is used to read from a csv file a single row that should be enough to recreate this Example.
        Don't forget to cast, as the csv reader returns strings only.

        Args:
            row: list, of strings, as returned by a csv reader for this row of a csv file.
            **kwargs: dict, other parameters passed by the factory. Optional.

        Returns:
            example: an instance of the Example object
        """
        return ToyExample(name=row[0], label=int(row[1]))

Overriding the to_csv_row method

Similarly, your Example subclass should override to_csv_row: it will be used when saving metadata describing what is stored in the train, eval (and optionally test) tfrecord files.

    def to_csv_row(self):
        """
        This is used to write in a csv file a single row that should be enough to recreate this Example afterward.
        """
        return [self.name, self._label]

Preparing the examples.csv file

The information required to instantiate your Example subclass shall be stored in a csv file that will be parsed by the factory. Each row will be passed to the from_csv_row static method of your Example subclass.

Generating the tfrecord files

That's it. The factory will generate the tfrecord files passing it your Example subclass, the path to your csv file and the other parameters that are used by your subclass' load and maybe split methods.

tf_factory.generate_and_save_train_eval_test_tfrecords_files(save_directory_path='/my/path/where/to/save',
                                                             example_class=ToyExample,
                                                             examples_list_filepath='/my/path/to/csv_file',
                                                             src_data_dirpath='/my/path/to/data', 
                                                             normalize=True,
                                                             chunk_size_in_bins=5)

The factory will save in your location in a train, eval and optionally test directories the corresponding tfrecord files.

Generating a tf.data.Dataset

The whole point of using tfrecord files is to stream them into a Dataset object.

To stream the content of tfrecord files stored in a directory, create a Dataset object with the factory. Each @tfrecordableattribute of your Example subclass will be streamed as a tuple where they appear in the same order they have been declared.

tfrecord_filepaths = tf_factory.get_tfrecord_filepaths(dirpath='my/path/where/to/save/train')

dataset = tf_factory.generate_dataset(tfrecord_filepaths)

for l in dataset:
    print(l) # label, data