SerialTime

SerialTime is a Python serialization tool containing many serialization and deserialization shortcuts with timing. There are many modules that can serialize Python object such as:

However, remembering their API is very difficult, and there are many differences in their API. For example, json.dump only accept file-like object as its argument, so we need to open the file first, and then give the function the file-like object, while joblib.dump only accept file path as its argument. Besides, there is no compression shortcut for Python built-in pickle/cPickle, so we also need to remember how to use something like gzip.

This package aims to solve these problems and provides very simple and unified API shortcuts for some popular serialization methods. In addition, we use BisTiming to calculate the execution time, so you can also easily know how quick the serialization is.

Installation

Install serialtime
```
pip install serialtime
```
If you want to use save_joblib_pkl or load_joblib_pkl:
```
pip install scikit-learn scipy
```
If you want to use try_load_yaml:
```
pip install PyYAML
```

Documentation

Shortcuts

Python built-in pickle/cPickle

serialtime.save_pkl(obj, path, log_description=None, logger=None,
                    logging_level=logging.INFO, verbose_start=True,
                    verbose_end=True, end_in_new_line=True, log_prefix="...")

obj = serialtime.load_pkl(path, log_description=None, logger=None,
                          logging_level=logging.INFO, verbose_start=True,
                          verbose_end=True, end_in_new_line=True, log_prefix="...")

Python built-in pickle/cPickle + gzip

serialtime.save_pklgz(obj, path, log_description=None, logger=None,
                      logging_level=logging.INFO, verbose_start=True,
                      verbose_end=True, end_in_new_line=True, log_prefix="...")

obj = serialtime.load_pklgz(path, log_description=None, logger=None,
                            logging_level=logging.INFO, verbose_start=True,
                            verbose_end=True, end_in_new_line=True, log_prefix="...")

Joblib

serialtime.save_joblib_pkl(obj, path, log_description=None, logger=None,
                           logging_level=logging.INFO, verbose_start=True,
                           verbose_end=True, end_in_new_line=True, log_prefix="...")

obj = serialtime.load_joblib_pkl(path, log_description=None, logger=None,
                                 logging_level=logging.INFO, verbose_start=True,
                                 verbose_end=True, end_in_new_line=True, log_prefix="...")

Interactive trying of loading YAML

Sometimes we want to load the configuration file in the middle of a program. If we run the program very long and the file format is incorrect, the program may directly raise an error and exit, so we don't have any chance to fix the file. serialtime.try_load_yaml can try to load the file, and pause when it encounter any error, and ask you whether to reload the file. We can then fix the file and continue running the program.

serialtime.try_load_yaml(yaml_path)

Dataset wrapper

PartialPreprocessedDataset is used to transparrently reindex the data without moving or copying the original memory.

Sometimes we want to reindex the data, for example:

In [1]: import numpy as np

In [2]: dset = np.asarray([1, 2, 3])

In [3]: dset
Out[3]: array([1, 2, 3])

In [4]: idx = [2, 0]

In [5]: dset2 = dset[idx]

In [6]: dset2
Out[6]: array([3, 1])

However, if the data is very large or it's on disk, this may use too much memory. We may not need all the convenient API in numpy.ndarray or h5py.dataset, but some modules only accept a full numpy.ndarray or h5py.dataset (i.e., keras.image.ImageDataGenerator.flow()). Our solution is to use an object to remember the new index, and translate the index while getting the value. For example:

In [1]: import numpy as np

In [2]: from serialtime import PartialPreprocessedDataset

In [3]: dset = np.asarray([[0, 1], [2, 3], [4, 5]])

In [4]: dset
Out[4]:
array([[0, 1],
       [2, 3],
       [4, 5]])

In [5]: idx = [2, 0]

In [6]: dset2 = PartialPreprocessedDataset(dset, idx, shape=(2,), preprocess_func=lambda x: x*2)

We can also use an optional preprocess_func to preprocess the instance while we are getting it. In this example, we just double the values in the array. The shape we give to PartialPreprocessedDataset is the shape of one instance (the shape of the array that we can get after preprocess_func(dset[x])). Then we can do something like:

In [7]: dset2.shape
Out[7]: (2, 2)

In [8]: len(dset2)
Out[8]: 2

In [9]: dset2[0]
Out[9]: array([ 8, 10])

In [10]: dset2[1]
Out[10]: array([0, 2])

In [11]: dset2[2]
IndexError: list index out of range

Testing

For the current environment: python setup test.
For Python 2.7, 3.4, 3.5, 3.6 and installation test: tox.