TensorFlow Dataset API

Learning Objectives

Learn how use tf.data to read data from memory
Learn how to use tf.data in a training loop
Learn how use tf.data to read data from disk
Learn how to write production input pipelines with feature engineering (batching, shuffling, etc.)

In this notebook, we will start by refactoring the linear regression we implemented in the previous lab so that is takes its data from atf.data.Dataset, and we will learn how to implement stochastic gradient descent with it. In this case, the original dataset will be synthetic and read by the tf.data API directly from memory.

In a second part, we will learn how to load a dataset with the tf.data API when the dataset resides on disk.



In [ ]:

    
!sudo chown -R jupyter:jupyter /home/jupyter/training-data-analyst



In [ ]:

    
# Ensure the right version of Tensorflow is installed.
!pip freeze | grep tensorflow==2.1 || pip install tensorflow==2.1



In [ ]:

    
import json
import math
import os
from pprint import pprint

import numpy as np
import tensorflow as tf
print(tf.version.VERSION)

Loading data from memory

Creating the dataset

Let's consider the synthetic dataset of the previous section:



In [ ]:

    
N_POINTS = 10
X = tf.constant(range(N_POINTS), dtype=tf.float32)
Y = 2 * X + 10

We begin with implementing a function that takes as input

our $X$ and $Y$ vectors of synthetic data generated by the linear function $y= 2x + 10$
the number of passes over the dataset we want to train on (epochs)
the size of the batches the dataset (batch_size)

and returns a tf.data.Dataset:

Remark: Note that the last batch may not contain the exact number of elements you specified because the dataset was exhausted.

If you want batches with the exact same number of elements per batch, we will have to discard the last batch by setting:

dataset = dataset.batch(batch_size, drop_remainder=True)

We will do that here.



In [ ]:

    
# TODO 1
def create_dataset(X, Y, epochs, batch_size):
    dataset = tf.data.Dataset.from_tensor_slices((X, Y))
    dataset = dataset.repeat(epochs).batch(batch_size, drop_remainder=True)
    return dataset

Let's test our function by iterating twice over our dataset in batches of 3 datapoints:



In [ ]:

    
BATCH_SIZE = 3
EPOCH = 2

dataset = create_dataset(X, Y, epochs=1, batch_size=3)

for i, (x, y) in enumerate(dataset):
    print("x:", x.numpy(), "y:", y.numpy())
    assert len(x) == BATCH_SIZE
    assert len(y) == BATCH_SIZE
assert  EPOCH

Loss function and gradients

The loss function and the function that computes the gradients are the same as before:



In [ ]:

    
def loss_mse(X, Y, w0, w1):
    Y_hat = w0 * X + w1
    errors = (Y_hat - Y)**2
    return tf.reduce_mean(errors)


def compute_gradients(X, Y, w0, w1):
    with tf.GradientTape() as tape:
        loss = loss_mse(X, Y, w0, w1)
    return tape.gradient(loss, [w0, w1])

Training loop

The main difference now is that now, in the traning loop, we will iterate directly on the tf.data.Dataset generated by our create_dataset function.

We will configure the dataset so that it iterates 250 times over our synthetic dataset in batches of 2.



In [ ]:

    
# TODO 2
EPOCHS = 250
BATCH_SIZE = 2
LEARNING_RATE = .02

MSG = "STEP {step} - loss: {loss}, w0: {w0}, w1: {w1}\n"

w0 = tf.Variable(0.0)
w1 = tf.Variable(0.0)

dataset = create_dataset(X, Y, epochs=EPOCHS, batch_size=BATCH_SIZE)

for step, (X_batch, Y_batch) in enumerate(dataset):

    dw0, dw1 = compute_gradients(X_batch, Y_batch, w0, w1)
    w0.assign_sub(dw0 * LEARNING_RATE)
    w1.assign_sub(dw1 * LEARNING_RATE)

    if step % 100 == 0:
        loss = loss_mse(X_batch, Y_batch, w0, w1)
        print(MSG.format(step=step, loss=loss, w0=w0.numpy(), w1=w1.numpy()))
        
assert loss < 0.0001
assert abs(w0 - 2) < 0.001
assert abs(w1 - 10) < 0.001

Loading data from disk

Locating the CSV files

We will start with the taxifare dataset CSV files that we wrote out in a previous lab.

The taxifare datast files been saved into ../data.

Check that it is the case in the cell below, and, if not, regenerate the taxifare dataset by running the provious lab notebook:



In [ ]:

    
!ls -l ../data/taxi*.csv

Use tf.data to read the CSV files

The tf.data API can easily read csv files using the helper function

tf.data.experimental.make_csv_dataset

If you have TFRecords (which is recommended), you may use

tf.data.experimental.make_batched_features_dataset

The first step is to define

the feature names into a list CSV_COLUMNS
their default values into a list DEFAULTS



In [ ]:

    
CSV_COLUMNS = [
    'fare_amount',
    'pickup_datetime',
    'pickup_longitude',
    'pickup_latitude',
    'dropoff_longitude',
    'dropoff_latitude',
    'passenger_count',
    'key'
]
LABEL_COLUMN = 'fare_amount'
DEFAULTS = [[0.0], ['na'], [0.0], [0.0], [0.0], [0.0], [0.0], ['na']]

Let's now wrap the call to make_csv_dataset into its own function that will take only the file pattern (i.e. glob) where the dataset files are to be located:



In [ ]:

    
# TODO 3
def create_dataset(pattern):
    return tf.data.experimental.make_csv_dataset(
        pattern, 1, CSV_COLUMNS, DEFAULTS)


tempds = create_dataset('../data/taxi-train*')
print(tempds)

Note that this is a prefetched dataset, where each element is an OrderedDict whose keys are the feature names and whose values are tensors of shape (1,) (i.e. vectors).

Let's iterate over the two first element of this dataset using dataset.take(2) and let's convert them ordinary Python dictionary with numpy array as values for more readability:



In [ ]:

    
for data in tempds.take(2):
    pprint({k: v.numpy() for k, v in data.items()})
    print("\n")

Transforming the features

What we really need is a dictionary of features + a label. So, we have to do two things to the above dictionary:

Remove the unwanted column "key"
Keep the label separate from the features

Let's first implement a funciton that takes as input a row (represented as an OrderedDict in our tf.data.Dataset as above) and then returns a tuple with two elements:

The first element beeing the same OrderedDict with the label dropped
The second element beeing the label itself (fare_amount)

Note that we will need to also remove the key and pickup_datetime column, which we won't use.



In [ ]:

    
UNWANTED_COLS = ['pickup_datetime', 'key']

# TODO 4a
def features_and_labels(row_data):
    label = row_data.pop(LABEL_COLUMN)
    features = row_data
    
    for unwanted_col in UNWANTED_COLS:
        features.pop(unwanted_col)

    return features, label

Let's iterate over 2 examples from our tempds dataset and apply our feature_and_labels function to each of the examples to make sure it's working:



In [ ]:

    
for row_data in tempds.take(2):
    features, label = features_and_labels(row_data)
    pprint(features)
    print(label, "\n")
    assert UNWANTED_COLS[0] not in features.keys()
    assert UNWANTED_COLS[1] not in features.keys()
    assert label.shape == [1]

Batching

Let's now refactor our create_dataset function so that it takes an additional argument batch_size and batch the data correspondingly. We will also use the features_and_labels function we implemented in order for our dataset to produce tuples of features and labels.



In [ ]:

    
# TODO 4b
def create_dataset(pattern, batch_size):
    dataset = tf.data.experimental.make_csv_dataset(
        pattern, batch_size, CSV_COLUMNS, DEFAULTS)
    return dataset.map(features_and_labels)

Let's test that our batches are of the right size:



In [ ]:

    
BATCH_SIZE = 2

tempds = create_dataset('../data/taxi-train*', batch_size=2)

for X_batch, Y_batch in tempds.take(2):
    pprint({k: v.numpy() for k, v in X_batch.items()})
    print(Y_batch.numpy(), "\n")
    assert len(Y_batch) == BATCH_SIZE

Shuffling

When training a deep learning model in batches over multiple workers, it is helpful if we shuffle the data. That way, different workers will be working on different parts of the input file at the same time, and so averaging gradients across workers will help. Also, during training, we will need to read the data indefinitely.

Let's refactor our create_dataset function so that it shuffles the data, when the dataset is used for training.

We will introduce a additional argument mode to our function to allow the function body to distinguish the case when it needs to shuffle the data (mode == "train") from when it shouldn't (mode == "eval").

Also, before returning we will want to prefetch 1 data point ahead of time (dataset.prefetch(1)) to speedup training:



In [ ]:

    
# TODO 4c
def create_dataset(pattern, batch_size=1, mode="eval"):
    dataset = tf.data.experimental.make_csv_dataset(
        pattern, batch_size, CSV_COLUMNS, DEFAULTS)

    dataset = dataset.map(features_and_labels).cache()

    if mode == "train":
        dataset = dataset.shuffle(1000).repeat()

    # take advantage of multi-threading; 1=AUTOTUNE
    dataset = dataset.prefetch(1)
    return dataset

Let's check that our function work well in both modes:



In [ ]:

    
tempds = create_dataset('../data/taxi-train*', 2, "train")
print(list(tempds.take(1)))



In [ ]:

    
tempds = create_dataset('../data/taxi-valid*', 2, "eval")
print(list(tempds.take(1)))

In the next notebook, we will build the model using this input pipeline.

Copyright 2019 Google Inc. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.