laminarflow 0.0.7

Streamline your TensorFlow workflow.

LaminarFlow

Streamline your TensorFlow workflow.

Installation

pip install laminarflow

Usage

TFRecord Datasets

LaminarFlow has two classes for writing to and reading from TFRecord datasets, DatasetWriter and DatasetReader.

When creating your datasets with DatasetWriter, you can pass in raw Python or Numpy data, and it will automatically get converted into TensorFlow Examples or SequenceExamples and be written to a TFRecord file.

Then when reading from the TFRecord file, DatasetReader takes care of creating the input pipeline that will parse your stored Examples or SequenceExamples, prepare them as needed (batching, padding, shuffling, etc.), then pass them to your TensorFlow Estimator, implementing the recommended best practices as outlined in TensorFlow's Input Pipline Performance Guide.

To demonstrate, we'll create some datasets.

import laminarflow as lf

train_writer = lf.DatasetWriter('data/train.tfrecord')
test_writer = lf.DatasetWriter('data/test.tfrecord')

train_writer.write({
  'input': [3.1, 4.1, 5.9],
  'label': 2
})

train_writer.write({
  'input': [2.7, 1.8, 2.8],
  'label': 1
})

test_writer.write({
  'input': [0.1, 1.2, 3.5],
  'label': 8
})

train_writer.close()
test_writer.close()

We create a DatasetWriter, then call the write method on it for each TensorFlow Example or SequenceExample we want to add to the dataset. When we call the write method, we pass in a dictionary where the keys are the feature names and the values are the feature values. The values can be Numpy arrays or any values that can be converted into Numpy arrays, such as Python ints, floats, or lists of ints or floats. The shape of the values can be multidimensional, but must be the same between Examples. Creating SequenceExamples, which support variable length data, is discussed below.

When we are done writing data with a Writer, we call the close() method on it.

The data will be written to a TFRecord file and the shapes and data types of your features will be stored in a separate metadata JSON file, which will have the same filename as the TFRecord file, except the extension will be changed to '.json'.

data/
├── test.json
├── test.tfrecord
├── train.json
└── train.tfrecord

We can then train a model on our datasets.

train_dataset = lf.DatasetReader('data/train.tfrecord')
test_dataset = lf.DatasetReader('data/test.tfrecord')

estimator = tf.estimator.Estimator(
  model_fn=model_fn,
  model_dir=MODEL_DIR,
  params=PARAMS)

train_spec = tf.estimator.TrainSpec(
    input_fn=train_dataset.input_fn,
    max_steps=1000)

eval_spec = tf.estimator.EvalSpec(
    input_fn=test_dataset.input_fn)

tf.estimator.train_and_evaluate(
    estimator=estimator,
    train_spec=train_spec,
    eval_spec=eval_spec)

Calling lf.DatasetReader('data/train.tfrecord') creates a dataset using the TFRecord file and its corresponding metadata JSON file. The path to the metadata JSON file data/train.json is inferred from the TFRecord path.

The created dataset has an input_fn method that you can pass in as the input function to a TensorFlow Estimator. The input_fn method automatically creates the input pipeline for your dataset.

For a more complete example of creating datasets, training a model, and making predictions with that model, check out: xor.py

Using a with Statement

A DatasetWriter can also be created using a with statement, in which case the close() method does not need to be called.

with lf.DatasetWriter('data/train.tfrecord') as train_writer:
  train_writer.write({
    'input': [1.4, 1.4, 2.1],
    'label': 3
  })

SequenceExamples

The default behavior of the write method is to write a TensorFlow Example. To write a SequenceExample, instead of passing in features to the first parameter of the write method, pass in features using the context_features and sequence_features parameters.

train_writer.write(
  context_features={
    'category': 7
  },
  sequence_features={
    'inputs': [[1.4, 0.0], [1.4, 0.0], [1.4, 0.0]],
    'labels': [3, 5, 3]
  })

train_writer.write(
  context_features={
    'category': 5
  },
  sequence_features={
    'inputs': [[1.4, 0.0], [1.4, 0.0]],
    'labels': [3, 5]
  })

Passing in context_features is optional, but if used, their values must have the same shape between SequenceExamples, similar to Example features.

The shape of the sequence_features values must have a rank of at least 1. The length of the first dimension must be the same for all sequence_features within a SequenceExample, but can vary between SequenceExamples. And the lengths of the rest of the dimensions can vary between features, but must be the same between SequenceExamples.

When a batch of SequenceExamples is created, any sequences that are shorter than the longest sequence will be padded with zeros.

The length of each sequence will be extracted from the data as one of the steps in the input pipeline when reading from the dataset. The lengths of the sequences will be made available as one of the feature values passed into the model_fn, features['lengths']. It will be a batch size length list of ints, that are the lengths of each of the sequences in the batch before that sequence was possibly padded with zeros.