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.
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')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
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.
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. IfNonenot compressed.compression_level=9- compression level 0...9.columns=None- list of columns to save, ifNoneall 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. IfNoneit 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- IfNoneschema 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(orbytes),int,float, and for sequences it should be wrapped tolist:[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- ifTrueit castsbytesdata after converting fromtf.string. It tries to cast it toint,floatandstrsequentially. If it's not possible, otherwise keeps as is.
alias pandas_tfrecords.tf2pd