Skip to main content

Converter pandas to tfrecords & tfrecords to pandas

Project description

This project was born under impression from spark-tensorflow-connector and implements similar functionality in order to save easy pandas dataframe to tfrecords and to restore tfrecords to pandas dataframe.

It can work as with local files as with AWS S3 files. Please keep in mind, that here tensorflow works with local copies of remote files, which are synced via s3fs with S3. I did this workaround because my tensorflow v2.1.0 didn’t work with S3 directly and raised authentication error Credentials have expired attempting to repull from EC2 Metadata Service, but maybe it’s fixed already.

Quick start

pip install pandas-tfrecords
import pandas as pd
from pandas_tfrecords import pd2tf, tf2pd

df = pd.DataFrame({'A': [1, 2, 3], 'B': ['a', 'b', 'c'], 'C': [[1, 2], [3, 4], [5, 6]]})

# local
pd2tf(df, './tfrecords')
my_df = tf2pd('./tfrecords')

# S3
pd2tf(df, 's3://my-bucket/tfrecords')
my_df = tf2pd('s3://my-bucket/tfrecords')

boolean support (v0.1.6+):

import pandas as pd
from pandas_tfrecords import pd2tf, tf2pd

df = pd.DataFrame({'A': [True, False, True], 'B': ['a', 'b', 'c'], 'C': [[1, False], [3, True], [5, False]]})

pd2tf(df, './tfrecords')

my_df = tf2pd('./tfrecords', schema={'A': bool, 'B': str, 'C': [int, bool]})

# or if needs 0/1 instead of True/False
my_df = tf2pd('./tfrecords', schema={'A': int, 'B': str, 'C': [int, int]})

# or just skip schema, int used by default if bool isn't specified
my_df = tf2pd('./tfrecords')

Converted types

pandas -> tfrecords

bytes, str -> tf.string
int, np.integer, bool, np.bool_ -> tf.int64
float, np.floating -> tf.float32
list, np.ndarray of bytes, str, int, np.integer, bool, np.bool_, float, np.floating -> sequence of tf.string, tf.int64, tf.float32

tfrecords -> pandas

tf.string -> bytes
tf.int64 -> int, bool
tf.float32 -> float
sequence of tf.string, tf.int64, tf.float32 -> list of bytes, int, bool, float

NB! Please pay attention it works only with one-dimentional arrays. It means [1, 2, 3] will be converted to both sides, but [[1,2,3]] won’t be converted to any side. It works that, because spark-tensorflow-connector works similar, and I didn’t learn yet how to implement nested sequences. In order to work with nested sequences you should use reshape.

API

pandas_tfrecords.pandas_to_tfrecords(df, folder, compression_type='GZIP', compression_level=9, columns=None, max_mb=50)

Arguments:

  • df - pandas dataframe. Please keep in mind above info about nested sequences.

  • folder - folder to save tfrecords, local or S3. Please be sure that it doesn’t contain other files or folders, if you want to read from this folder then.

  • compression_type='GZIP' - compression types: 'GZIP', 'ZLIB', None. If None not compressed.

  • compression_level=9 - compression level 0…9.

  • columns=None - list of columns to save, if None all columns will be saved.

  • max_mb=50 - maximum size of uncompressed data to save in single file. If dataframe total size is bigger than this limit, then it will be splitted to several files. If None it isn’t limited and single file will be saved.

alias pandas_tfrecords.pd2tf

pandas_tfrecords.tfrecords_to_pandas(file_paths, schema=None, compression_type='auto', cast=True)

Arguments:

  • file_paths - One or sequence of file paths or folders, local or S3, to read tfrecords from.

  • schema=None - If None schema will be detected automatically. But you can specify which columns you want to read only. It should be a dict, which keys are column names and values are column data types: str (or bytes), int, float, and for sequences it should be wrapped to list: [str] (or [bytes]), [int], [float]. For example:

df = pd.DataFrame({'A': [1, 2, 3], 'B': ['a', 'b', 'c'], 'C': [[1, 2], [3, 4], [5, 6]]})
print(df)
   A  B       C
0  1  a  [1, 2]
1  2  b  [3, 4]
2  3  c  [5, 6]

pd2tf(df, './tfrecords')
tf2pd('./tfrecords', schema={'A': int, 'C': [int]})
   A       C
0  1  [1, 2]
1  2  [3, 4]
2  3  [5, 6]
  • compression_type='auto' - compression type: 'auto', 'GZIP', 'ZLIB', None.

  • cast=True - if True it casts bytes data after converting from tf.string. It tries to cast it to int, float and str sequentially. If it’s not possible, otherwise keeps as is.

alias pandas_tfrecords.tf2pd

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pandas-tfrecords-0.1.6.tar.gz (6.0 kB view hashes)

Uploaded source

Built Distribution

pandas_tfrecords-0.1.6-py3-none-any.whl (7.2 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page