Introduction to PyTorch Ignite
Neural networks training and evaluation with PyTorch is done faster with the help of a few high-level libraries and one among them is PyTorch Ignite. There are deep learning technical skills involved in the library where it is mostly research-oriented and it can do the processes faster than other libraries. Modules, optimizers, and DataLoaders are present in Ignite so that all the needed classes are present in one library to do the work faster.
What is PyTorch Ignite?
There is an abstraction way to control the PyTorch Ignite and this method is called an engine. The training or evaluation function is run without many values inside the same where events are represented by the Event system in the function. This helps to customize the events in the run. Ignite is basically a training loop abstraction and there are metrics available to calculate and evaluate the models. Handlers are present where pipelines are available along with artifacts and parameters in the system.
Why use PyTorch Ignite?
There are various high-level features in Ignite. The engine and event system is very simple in that there are no codes for loops while doing iterations. There are events and handlers present in Ignite where there is much flexibility. All functions are included in handlers so that inheritance from any interface is not needed. Abstract methods are not overridden here which will reduce the complexity of the code and increase the lines in the code.
Out-of-the-box metrics are used in Ignite models where various regression tasks such as accuracy, precision, confusion matrix, recall, and you are used. Arithmetic operations or any torch methods can be used to handle the metrics where users can compose the metrics by themselves.
There are several built-in handlers available in Ignite that can be used to create any training pipeline so that artifacts can be saved from the pipeline. Handlers also help to log the parameters easily and create metrics based on requirements.
Using PyTorch Ignite Projects
Let us look into the metrics module project of Ignite. All metrics should be in a distributed configuration where the current implementation of a few metrics should be updated first followed by the implementation of tests. Objects are detected with the help of new metrics. New metrics are needed for NLP and GANs as well. Label wise metrics are enabled for all the label-related problems. New API is added to the label wise option in the dataset. Sklearn metrics are supported here along with micro or macro options. Current metrics should be updated only if the present distributed configuration is not supported in the system. NLP tasks include ROUGE and BLEU. Chosen API is implemented along with several tests in the system.
Any one or two metrics should be implemented as output. New tests should be implemented to know whether the metrics get along with all the libraries. Configurable metrics and label wise API metrics are needed for all the setup in the metrics module. Required skills are Python, neural networks in PyTorch, AI related open source projects, and implementing new technicalities in the project. Out of the box, integration is done for several metrics so that all the technical setups are done in the project.
Basic PyTorch Setup
PyTorch can be installed in Windows with the following prerequisites. It must be either Windows 7 or a greater version of Windows 7. Windows Server 2008 is preferred. Python 3 is supported in PyTorch that can be installed via Anaconda or Chocolatey or from the Python website directly. From the website, it is a direct download of the Python application and in Chocolatey, we should run the following command.
choco install python
Either Anaconda or pip is needed to install PyTorch libraries in the system. It is good to install Anaconda as it has both Python and PyTorch. A 64-bit graphical installer is needed for Anaconda where it is just simple as clicking on the PyTorch link and running the installation link. If Python was downloaded from the website or installed via Chocolatey, pip is installed directly in the system.
If CUDA is present in the system, install Anaconda via CUDA or else install via the options present in the website directly. We can test whether Python and PyTorch are installed using the following code.
Print(“Hello World”)
Import torch
A = torch.rand(3,6)
Print(A)
PyTorch Ignite Examples
from torchtext import data
from torchtext import datasets
from torchtext.vocab import Vectors
import torchtext
import torch
import torch.nn as nn
import torch.nn.functional as Fun
SEEDS = 134
torch.manual_seed(SEEDS)
torch.cuda.manual_seed(SEEDS)
import pathlib
import numpy
import sklearn.metrics
import pprint
import csv
import sys
csv.field_size_limit(sys.maxsize)
pip install pytorch-ignite
from ignite.engine import Engine, Events
from ignite.metrics import Accuracy, Loss, RunningAverage, Precision, Recall
from ignite.handlers import ModelCheckpoint, EarlyStopping
from ignite.contrib.handlers import ProgressBar
input_path = '../input/dataloaded-dataset-2021-clf'
vectors_path = '../input/glove890b1300dtxt/glove.890B.1300d.txt'
cache_path = '../input/glove890b1300dtxt'
%matplotlib inline
import pandas
dataframe = pd.read_csv(f'{input_path}/train.csv')
dataframe.head()
dataframe.label.value_counts()
system = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
training_loader = data.Iterator(training_dataset, batch_size=32, system='cuda', shuffle=True, sort=False)
value_loader = data.Iterator(value_dataset, batch_size=32, system='cuda', shuffle=False, sort=False)
testing_loader = data.Iterator(testing_dataset, batch_size=32, system='cuda', shuffle=False, sort=False)
batch_load = next(iter(training_loader))
print(batch_load)
next(iter(testing_loader))
class Model(nn.Module):
def __init__(self, vocabulary_size, embed_dim, kernel_size, number_filters,
number_classes, d_problem, mode, hidden_dimen, lstm_units,
embed_vectors=None):
super(Model, self).__init__()
self.vocabulary_size = vocabulary_size
self.embed_dim = embed_dim
self.kernel_size = kernel_size
self.number_filters = number_filters
self.number_classes = number_classes
self.d_problem = d_problem
self.mode = mode
self.embed = nn.Embedding(vocabulary_size, embed_dim, padding_index=1)
self.load_embeddings(embed_vectors)
self.convol = nn.ModuleList([nn.Conv1d(input_channels=embed_dim,
output_channels=number_filters,
kernel_size=kel, stride=1) for kel in kernel_size])
self.conv2 = nn.ModuleList([nn.Conv1d(input_channels=embed_dim,
output_channels=number_filters,
kernel_size=kel, stride=1) for kel in kernel_sizes])
self.conv_body = nn.ModuleList([nn.Conv1d(input_channels=embed_dim,
output_channels=number_filters,
kernel_size=kel, stride=1) for kel in kernel_sizes])
self.lstm1 = nn.LSTM(embed_dim, lstm_units, bidirectional=True, batch_first=True)
self.lstm2 = nn.LSTM(lstm_units * 2, lstm_units, bidirectional=False, batch_first=True)
self.lstm_body = nn.LSTM(embed_dim, lstm_units, bidirectional=True, batch_first=True)
self.dropout = nn.Dropout(d_problem)
self.fc = nn.Linear(len(kernel_size) * number_filters, hidden_dimen)
self.fc_body = nn.Linear(len(kernel_size) * numbere_filters, hidden_dinm)
Conclusion
Modular code is more in Ignite but technical code is less when compared to PyTorch. There is more control to Ignite codes than any other PyTorch libraries. Maximum tools are provided to reduce the coupling in the system and to improve cohesion. If the configurations have more parameters, those are avoided in Ignite and it implements new use cases in the system.
Recommended Articles
This is a guide to PyTorch Ignite. Here we discuss the Introduction, What is PyTorch Ignite?, Why use PyTorch Ignite?, Examples. You may also have a look at the following articles to learn more –
