Serialization is a process of storing an object as a stream of bytes or characters in order to transmit it over a network or store it on the disk to recreate it along with its state whenever required. The reverse process is called deserialization.
In Python, the Pickle module provides us the means to serialize and deserialize the python objects. Pickle is a powerful library that can serialize many complex and custom objects that other library fails to do. Just like pickle, there is a cPickle module that shares the same methods as pickle, but it is written in C. The cPickle module is written as a C function instead of a class format.
Note:
cPickle was a Python library that provided a faster implementation of the pickle library, which is used for serializing and de-serializing Python objects.
cPickle has been deprecated in favor of the pickle library, which is now implemented in C and offers similar performance benefits.
Difference between Pickle and cPickle:
- Pickle uses python class-based implementation while cPickle is written as C functions. As a result, cPickle is many times faster than pickle.
- Pickle is available in both python 2.x and python 3.x while cPickle is available in python 2.x by default. To use cPickle in python 3.x, we can import _pickle.
- cPickle does not support subclass from pickle. cPickle is better if subclassing is not important otherwise Pickle is the best option.
Since both pickle and cPickle share the same interface, we can use both of them in the same way. Below is an example code as a reference:
Python3
try: # In python 2.x it is available as default import cPickle as pickleexcept ImportError: # In python 3.x cPickle is not available import pickleimport random# A custom class to demonstrate pickling class ModelTrainer: def __init__(self) -> None: self.weights = [0,0,0] def train(self): for i in range(len(self.weights)): self.weights[i] = random.random() def get_weights(self): return self.weights# Create an object model = ModelTrainer()# Populate the datamodel.train()print('Weights before pickling', model.get_weights())# Open a file to write bytesp_file = open('model.pkl', 'wb')# Pickle the objectpickle.dump(model, p_file)p_file.close()# Deserialization of the filefile = open('model.pkl','rb')new_model = pickle.load(file)print('Weights after pickling', new_model.get_weights()) |
Output:
Weights before pickling [0.6089721131909885, 0.7891019431265203, 0.5653418337976294]
Weights after pickling [0.6089721131909885, 0.7891019431265203, 0.5653418337976294]
In the above code, we have created a custom class ModelTrainer that initializes a list of 0’s. The train() method populates the list with some random values and get_weight() method returns the generated values. Next, we have created the model object and printed the generated weights. We have created a new file in ‘wb’ (Write bytes) mode. The dump() method dumped the object as bytes stream into the file. Verification is done by loading the file in a new object and printing the weights.
Pickle module is very powerful for python objects. But it can only preserve the data, not the class structure. Hence, any custom class object won’t load if we don’t provide the class definition. Below is an example when depickling fails:
Python3
try: # In python 2.x it is available as default import cPickle as pickleexcept ImportError: # In python 3.x cPickle is not available import pickle# Deserialization of the filefile = open('model.pkl','rb')new_model = pickle.load(file)print('Weights of model', new_model.get_weights()) |
Output:
Traceback (most recent call last):
File “des.py”, line 12, in <module>
new_model = pickle.load(file)
AttributeError: Can’t get attribute ‘ModelTrainer’ on <module ‘__main__’ from ‘des.py’>
The above error was generated because our current script doesn’t know about the class of this object. Thus, we can say that pickle is only preserving the data inside the object but it cannot save the methods and class structure.
To rectify the above error, we must provide the class definition to the script. Below is an example of how to correctly load custom objects:
Python3
try: # In python 2.x it is available as default import cPickle as pickleexcept ImportError: # In python 3.x cPickle is not available import pickleimport random# If the file is available,# we can use import statement to import the class# A custom class to demonstrate picklingclass ModelTrainer: def __init__(self) -> None: self.weights = [0, 0, 0] def train(self): for i in range(len(self.weights)): self.weights[i] = random.random() def get_weights(self): return self.weights# Deserialization of the filefile = open('model.pkl', 'rb')new_model = pickle.load(file)print('Weights of model', new_model.get_weights()) |
Output:
Weights of model [0.6089721131909885, 0.7891019431265203, 0.5653418337976294]
We have provided a reference for ModelTrainer class. The script now recognizes the class, and it can call the constructor again to build the object. Instead of typing the whole class code, we can simply import it from the previous file.
Serialization as string
We can also serialize an object as a string. Pickle and cPickle modules provide dumps() and loads() methods. The dumps() method takes the object as the parameter and returns the encoded string. The load() method does the reverse. It takes the encoded string and returns the original object. Below is the code to serialize a custom object as a string.
Python3
try: # In python 2.x it is available as default import cPickle as pickleexcept ImportError: # In python 3.x cPickle is not available import pickleimport random# A custom class to demonstrate pickling class ModelTrainer: def __init__(self) -> None: self.weights = [0,0,0] def train(self): for i in range(len(self.weights)): self.weights[i] = random.random() def get_weights(self): return self.weights# Create an object model = ModelTrainer()# Populate the datamodel.train()print('Weights before pickling', model.get_weights())# Pickle the objectbyte_string = pickle.dumps(model)print("The bytes of object are:",byte_string)# Deserialization of the object using same byte stringnew_model = pickle.loads(byte_string)print('Weights after depickling', new_model.get_weights()) |
Output:
Weights before pickling [0.923474126606742, 0.34909608824193983, 0.3761122243447367]
The bytes of object are: b’\x80\x03c__main__\nModelTrainer\nq\x00)\x81q\x01}q\x02X\x07\x00\x00\x00weightsq\x03]q\x04(G?\xed\x8d\x19\x9c\x8fL\xc3G?\xd6W\x97\x1e\x8aHHG?\xd8\x129\x01\xcb\xee\xf2esb.’
Weights after depickling [0.923474126606742, 0.34909608824193983, 0.3761122243447367]
