Introduction

This article will explore Generative Adversarial Networks (GANs) and their remarkable ability to fashion image generation. GANs have revolutionized the field of generative modeling, offering an innovative approach to creating new content through adversarial learning.

Throughout this guide, we will take you on a captivating journey, starting with the foundational concepts of GANs and gradually delving into the intricacies of fashion image generation. With hands-on projects and step-by-step instructions, we will walk you through building and training your GAN model using TensorFlow and Keras.

Get ready to unlock the potential of GANs and witness the magic of AI in the fashion world. Whether you’re a seasoned AI practitioner or a curious enthusiast, “GANS in Vogue” will equip you with the skills and knowledge to create awe-inspiring fashion designs and push the boundaries of generative art. Let’s dive into the fascinating world of GANs and unleash the creativity within!

This article was published as a part of the Data Science Blogathon.

Understanding Generative Adversarial Networks (GANs)

What are GANs?

Generative Adversarial Networks (GANs) consist of two neural networks: the generator and the discriminator. The generator is responsible for creating new data samples, while the discriminator’s task is to distinguish between real data and fake data generated by the generator. The two networks are trained simultaneously through a competitive process, where the generator improves its ability to create realistic samples while the discriminator becomes better at identifying real from fake.

How do GANs Work?

GANs are based on a game-like scenario where the generator and discriminator play against each other. The generator tries to create data that resembles real data, while the discriminator aims to differentiate between real and fake data. The generator learns to create more realistic samples through this adversarial training process.

Key Components of GANs

To build a GAN, we need several essential components:

• Generator: A neural network that generates new data samples.
• Discriminator: A neural network that classifies data as real or fake.
• Latent Space: A random vector space that the generator uses as input to produce samples.
• Training Loop: The iterative process of training the generator and discriminator in alternating steps.

Loss Functions in GANs

The GAN training process relies on specific loss functions. The generator tries to minimize the generator loss, encouraging it to create more realistic data. At the same time, the discriminator aims to minimize the discriminator loss, becoming better at distinguishing real from fake data.

Project Overview: Fashion Image Generation with GANs

Project Goal

In this project, we aim to build a GAN to generate new fashion images that resemble those from the Fashion MNIST dataset. The generated images should capture the essential features of various fashion items, such as dresses, shirts, pants, and shoes.

Dataset: Fashion MNIST

We will use the Fashion MNIST dataset, a popular benchmark dataset containing grayscale images of fashion items. Each image is 28×28 pixels, and there are ten classes in total.

Setting Up the Project Environment

To get started, we must set up our Python environment and install the necessary libraries, including TensorFlow, Matplotlib, and TensorFlow Datasets.

Building the GAN

Import Dependencies and Data

To get started, we must install and import the necessary libraries and load the Fashion MNIST dataset containing a collection of fashion images. We will use this dataset to train our AI model to generate new fashion images.

# Install required packages (only need to do this once)
!pip install tensorflow tensorflow-gpu matplotlib tensorflow-datasets ipywidgets
!pip list

# Import necessary libraries
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, Dense, Flatten, Reshape, LeakyReLU, Dropout, UpSampling2D
import tensorflow_datasets as tfds
from matplotlib import pyplot as plt

# Configure TensorFlow to use GPU for faster computation
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)

# Load the Fashion MNIST dataset
ds = tfds.load('fashion_mnist', split='train')

Visualize Data and Build a Dataset

Next, we will visualize sample images from the Fashion MNIST dataset and prepare the data pipeline. We will perform data transformations and create batches of images for training the GAN.

# Data Transformation: Scale and Vizualize Images
import numpy as np

# Setup data iterator
dataiterator = ds.as_numpy_iterator()

# Visualize some images from the dataset
fig, ax = plt.subplots(ncols=4, figsize=(20, 20))

# Loop four times and get images
for idx in range(4):
    # Grab an image and its label
    sample = dataiterator.next()
    image = np.squeeze(sample['image'])  # Remove the single-dimensional entries
    label = sample['label']

    # Plot the image using a specific subplot
    ax[idx].imshow(image)
    ax[idx].title.set_text(label)

# Data Preprocessing: Scale and Batch the Images
def scale_images(data):
    # Scale the pixel values of the images between 0 and 1
    image = data['image']
    return image / 255.0

# Reload the dataset
ds = tfds.load('fashion_mnist', split='train')

# Apply the scale_images preprocessing step to the dataset
ds = ds.map(scale_images)

# Cache the dataset for faster processing during training
ds = ds.cache()

# Shuffle the dataset to add randomness to the training process
ds = ds.shuffle(60000)

# Batch the dataset into smaller groups (128 images per batch)
ds = ds.batch(128)

# Prefetch the dataset to improve performance during training
ds = ds.prefetch(64)

# Check the shape of a batch of images
ds.as_numpy_iterator().next().shape

In this step, we first visualize four random fashion images from the dataset using the matplotlib library. This helps us understand what the images look like and what we want our AI model to learn.

After visualizing the images, we proceed with data preprocessing. We scale the pixel values of the images between 0 and 1, which helps the AI model learn better. Imagine scaling the brightness of images to be suitable for learning.

Next, we batch the images into groups of 128 (a batch) to train our AI model. Think of batches as dividing a big task into smaller, manageable chunks.

We also shuffle the dataset to add some randomness so the AI model doesn’t learn the images in a fixed order.

Finally, we prefetch the data to prepare it for the AI model’s learning process, making it run faster and more efficiently.

At the end of this step, we have visualized some fashion images, and our dataset is prepared and organized for training the AI model. We are now ready to move on to the next step, where we will build the neural network to generate new fashion images.

Build the Generator

The generator is crucial to the GAN, creating new fashion images. We will design the generator using TensorFlow’s Sequential API, incorporating layers like Dense, LeakyReLU, Reshape, and Conv2DTranspose.

# Import the Sequential API for building models
from tensorflow.keras.models import Sequential

# Import the layers required for the neural network
from tensorflow.keras.layers import (
    Conv2D, Dense, Flatten, Reshape, LeakyReLU, Dropout, UpSampling2D
)

def build_generator():
    model = Sequential()

    # First layer takes random noise and reshapes it to 7x7x128
    # This is the beginning of the generated image
    model.add(Dense(7 * 7 * 128, input_dim=128))
    model.add(LeakyReLU(0.2))
    model.add(Reshape((7, 7, 128)))

    # Upsampling block 1
    model.add(UpSampling2D())
    model.add(Conv2D(128, 5, padding='same'))
    model.add(LeakyReLU(0.2))

    # Upsampling block 2
    model.add(UpSampling2D())
    model.add(Conv2D(128, 5, padding='same'))
    model.add(LeakyReLU(0.2))

    # Convolutional block 1
    model.add(Conv2D(128, 4, padding='same'))
    model.add(LeakyReLU(0.2))

    # Convolutional block 2
    model.add(Conv2D(128, 4, padding='same'))
    model.add(LeakyReLU(0.2))

    # Convolutional layer to get to one channel
    model.add(Conv2D(1, 4, padding='same', activation='sigmoid'))

    return model

# Build the generator model
generator = build_generator()
# Display the model summary
generator.summary()

The generator is a deep neural network responsible for generating fake fashion images. It takes random noise as input, and its output is a 28×28 grayscale image that looks like a fashion item. The goal is to learn how to generate images that resemble real fashion items.

Several Layers of the Model

The model consists of several layers:

1. Dense Layer: The first layer takes random noise of size 128 and reshapes it into a 7x7x128 tensor. This creates the initial structure of the generated image.
2. Upsampling Blocks: These blocks gradually increase the image’s resolution using the UpSampling2D layer, followed by a convolutional layer and a LeakyReLU activation. The Upsampling2D layer doubles the resolution of the image along both dimensions.
3. Convolutional Blocks: These blocks further refine the generated image. They consist of convolutional layers with LeakyReLU activations.
4. Convolutional Layer: The final convolutional layer reduces the channels to one, effectively creating the output image with a sigmoid activation to scale the pixel values between 0 and 1.

At the end of this step, we will have a generator model capable of producing fake fashion images. The model is now ready for training in the next steps of the process.

Build the Discriminatory

Starting with the foundational concepts of GANs and gradually delving into the intricacies of fashion image generation. With hands-on projects and step-by-step instructions, we will walk you through building and training your GAN model using TensorFlow and Keras.

The discriminator plays a critical role in distinguishing between real and fake images. We will design the discriminator using TensorFlow’s Sequential API, incorporating Conv2D, LeakyReLU, Dropout, and Dense layers.

def build_discriminator():
    model = Sequential()

    # First Convolutional Block
    model.add(Conv2D(32, 5, input_shape=(28, 28, 1)))
    model.add(LeakyReLU(0.2))
    model.add(Dropout(0.4))

    # Second Convolutional Block
    model.add(Conv2D(64, 5))
    model.add(LeakyReLU(0.2))
    model.add(Dropout(0.4))

    # Third Convolutional Block
    model.add(Conv2D(128, 5))
    model.add(LeakyReLU(0.2))
    model.add(Dropout(0.4))

    # Fourth Convolutional Block
    model.add(Conv2D(256, 5))
    model.add(LeakyReLU(0.2))
    model.add(Dropout(0.4))

    # Flatten the output and pass it through a dense layer
    model.add(Flatten())
    model.add(Dropout(0.4))
    model.add(Dense(1, activation='sigmoid'))

    return model

# Build the discriminator model
discriminator = build_discriminator()
# Display the model summary
discriminator.summary()

The discriminator is also a deep neural network for classifying whether an input image is real or fake. It inputs a 28×28 grayscale image and outputs a binary value (1 for real, 0 for fake).

The model consists of several layers:

1. Convolutional Blocks: These blocks process the input image with convolutional layers, followed by LeakyReLU activations and dropout layers. The dropout layers help prevent overfitting by randomly dropping some neurons during training.
2. Flatten and Dense Layers: The output from the last convolutional block is flattened to a 1D vector and passed through a dense layer with sigmoid activation. The sigmoid activation squashes the output between 0 and 1, representing the probability of the image being real.

At the end of this step, we will have a discriminator model capable of classifying whether an input image is real or fake. The model is now ready to be integrated into the GAN architecture and trained in the next steps.

Construct the Training Loop

Set up Losses and Optimizer

Before building the training loop, we need to define the loss functions and optimizers that will be used to train both the generator and discriminator.

# Import the Adam optimizer and Binary Cross Entropy loss function
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.losses import BinaryCrossentropy

# Define the optimizers for the generator and discriminator
g_opt = Adam(learning_rate=0.0001)  # Generator optimizer
d_opt = Adam(learning_rate=0.00001)  # Discriminator optimizer

# Define the loss functions for the generator and discriminator
g_loss = BinaryCrossentropy()  # Generator loss function
d_loss = BinaryCrossentropy()  # Discriminator loss function

• We are using the Adam optimizer for both the generator and discriminator. Adam is an efficient optimization algorithm that adapts the learning rate during training.
• For the loss functions, we are using Binary Cross Entropy. This loss function is commonly used for binary classification problems, suitable for our discriminator’s binary classification task (real vs. fake).

Build Subclassed Model

Next, we will build a subclassed model that combines the generator and discriminator models into a single GAN model. This subclassed model will train the GAN during the training loop.

from tensorflow.keras.models import Model

class FashionGAN(Model):
    def __init__(self, generator, discriminator, *args, **kwargs):
        # Pass through args and kwargs to the base class
        super().__init__(*args, **kwargs)

        # Create attributes for generator and discriminator models
        self.generator = generator
        self.discriminator = discriminator

    def compile(self, g_opt, d_opt, g_loss, d_loss, *args, **kwargs):
        # Compile with the base class
        super().compile(*args, **kwargs)

        # Create attributes for optimizers and loss functions
        self.g_opt = g_opt
        self.d_opt = d_opt
        self.g_loss = g_loss
        self.d_loss = d_loss

    def train_step(self, batch):
        # Get the data for real images
        real_images = batch
        # Generate fake images using the generator with random noise as input
        fake_images = self.generator(tf.random.normal((128, 128, 1)), training=False)

        # Train the discriminator
        with tf.GradientTape() as d_tape:
            # Pass real and fake images through the discriminator model
            yhat_real = self.discriminator(real_images, training=True)
            yhat_fake = self.discriminator(fake_images, training=True)
            yhat_realfake = tf.concat([yhat_real, yhat_fake], axis=0)

            # Create labels for real and fake images
            y_realfake = tf.concat([tf.zeros_like(yhat_real), tf.ones_like(yhat_fake)], axis=0)

            # Add some noise to the true outputs to make training more robust
            noise_real = 0.15 * tf.random.uniform(tf.shape(yhat_real))
            noise_fake = -0.15 * tf.random.uniform(tf.shape(yhat_fake))
            y_realfake += tf.concat([noise_real, noise_fake], axis=0)

            # Calculate the total discriminator loss
            total_d_loss = self.d_loss(y_realfake, yhat_realfake)

        # Apply backpropagation and update discriminator weights
        dgrad = d_tape.gradient(total_d_loss, self.discriminator.trainable_variables)
        self.d_opt.apply_gradients(zip(dgrad, self.discriminator.trainable_variables))

        # Train the generator
        with tf.GradientTape() as g_tape:
            # Generate new images using the generator with random noise as input
            gen_images = self.generator(tf.random.normal((128, 128, 1)), training=True)

            # Create the predicted labels (should be close to 1 as they are fake images)
            predicted_labels = self.discriminator(gen_images, training=False)

            # Calculate the total generator loss (tricking the discriminator to classify the fake images as real)
            total_g_loss = self.g_loss(tf.zeros_like(predicted_labels), predicted_labels)

        # Apply backpropagation and update generator weights
        ggrad = g_tape.gradient(total_g_loss, self.generator.trainable_variables)
        self.g_opt.apply_gradients(zip(ggrad, self.generator.trainable_variables))

        return {"d_loss": total_d_loss, "g_loss": total_g_loss}

# Create an instance of the FashionGAN model
fashgan = FashionGAN(generator, discriminator)

# Compile the model with the optimizers and loss functions
fashgan.compile(g_opt, d_opt, g_loss, d_loss)

• We create a subclassed FashionGAN model that extends the tf.keras.models.Model class. This subclassed model will handle the training process for the GAN.
• In the train_step method, we define the training loop for the GAN:
  • We first obtain authentic images from the batch and generate fake images using the generator model with random noise as input.
  • Then, we train the discriminator:
    • We use a gradient tape to calculate the discriminator’s loss concerning real and fake images. The goal is to make the discriminator classify authentic images as 1 and fake images as 0.
    • We add some noise to the true outputs to make the training more robust and less prone to overfitting.
    • The total discriminator loss is calculated as the binary cross entropy between the predicted and target labels.
    • We apply backpropagation to update the discriminator’s weights based on the calculated loss.
  • Next, we train the generator:
    • We generate new fake images using the generator with random noise as input.
    • We calculate the total generator loss as the binary cross entropy between the predicted labels (generated images) and the target labels (0, representing fake images).
    • The generator aims to “fool” the discriminator by generating images that the discriminator classifies as real (with a label close to 1).
    • We apply backpropagation to update the generator’s weights based on the calculated loss.
  • Finally, we return the total losses for the discriminator and generator during this training step.

The FashionGAN model is now ready to be trained using the training dataset in the next step.

Build Callback

Callbacks in TensorFlow are functions that can be executed during training at specific points, such as the end of an epoch. We will create a custom callback called ModelMonitor to generate and save images at the end of each epoch to monitor the progress of the GAN.

import os
from tensorflow.keras.preprocessing.image import array_to_img
from tensorflow.keras.callbacks import Callback

class ModelMonitor(Callback):
    def __init__(self, num_img=3, latent_dim=128):
        self.num_img = num_img
        self.latent_dim = latent_dim

    def on_epoch_end(self, epoch, logs=None):
        # Generate random latent vectors as input to the generator
        random_latent_vectors = tf.random.uniform((self.num_img, self.latent_dim, 1))
        # Generate fake images using the generator
        generated_images = self.model.generator(random_latent_vectors)
        generated_images *= 255
        generated_images.numpy()
        for i in range(self.num_img):
            # Save the generated images to disk
            img = array_to_img(generated_images[i])
            img.save(os.path.join('images', f'generated_img_{epoch}_{i}.png'))

• The ModelMonitor callback takes two arguments: num_img, which specifies the number of images to generate and save at the end of each epoch, and latent_dim, which is the dimension of the random noise vector used as input to the generator.
• During the on_epoch_end method, the callback generates num_img random latent vectors and passes them as input to the generator. The generator then generates fake images based on these random vectors.
• The generated images are scaled to the 0-255 range and saved as PNG files in the “images” directory. The filenames include the epoch number to keep track of the progress over time.

Train the GAN

Now that we have set up the GAN model and the custom callback, we can start the training process using the fit method. We will train the GAN for sufficient epochs to allow the generator and discriminator to converge and learn from each other.

# Train the GAN model
hist = fashgan.fit(ds, epochs=20, callbacks=[ModelMonitor()])

• We use the fit method of the FashionGAN model to train the GAN.
• We set the number of epochs to 20 (you may need more epochs for better results).
• We pass the ModelMonitor callback to save generated images at the end of each epoch.
• The training process will iterate over the dataset, and for each batch, it will update the weights of the generator and discriminator models using the training loop defined earlier.

The training process can take some time, depending on your hardware and the number of epochs. After training, we can review the performance of the GAN by plotting the discriminator and generator losses. This will help us understand how well the models have been trained and whether there is any sign of convergence or mode collapse. Let’s move on to the next step, to review the performance of the GAN.

Review Performance and Test the Generator

Review performance

After training the GAN, we can review its performance by plotting the discriminator and generator losses over the training epochs. This will help us understand how well the GAN has learned and whether there are any issues, such as mode collapse or unstable training.

import matplotlib.pyplot as plt

# Plot the discriminator and generator losses
plt.suptitle('Loss')
plt.plot(hist.history['d_loss'], label='d_loss')
plt.plot(hist.history['g_loss'], label='g_loss')
plt.legend()
plt.show()

• We use matplotlib to plot the discriminator and generator losses over the training epochs.
• The x-axis represents the epoch number, and the y-axis represents the corresponding losses.
• The discriminator loss (d_loss) and generator loss (g_loss) should ideally decrease over epochs as the GAN learns.

Test out the Generator

After training the GAN and reviewing its performance, we can test the generator by generating and visualizing new fashion images. First, we will load the weights of the trained generator and use it to generate new images.

# Load the weights of the trained generator
generator.load_weights('generator.h5')

# Generate new fashion images
imgs = generator.predict(tf.random.normal((16, 128, 1)))

# Plot the generated images
fig, ax = plt.subplots(ncols=4, nrows=4, figsize=(10, 10))
for r in range(4):
    for c in range(4):
        ax[r][c].imshow(imgs[(r + 1) * (c + 1) - 1])

• We load the weights of the trained generator from the saved file using generator.load_weights(‘generator.h5’).
• We generate new fashion images by passing random latent vectors to the generator. The generator interprets these random vectors and generates corresponding images.
• We use matplotlib to display the generated images in a 4×4 grid.

Save the Model

Finally, if you are satisfied with the performance of your GAN, you can save the generator and discriminator models for future use.

# Save the generator and discriminator models
generator.save('generator.h5')
discriminator.save('discriminator.h5')

• We save the generator and discriminator models to disk using the save method.
• The models will be saved in the current working directory with filenames “generator.h5” and “discriminator.h5,” respectively.
• Saving the models allows you to use them later to generate more fashion images or to continue the training process.

And that concludes the process of building and training a GAN for generating fashion images using TensorFlow and Keras! GANs are powerful models for generating realistic data and can be applied to other tasks.

Remember that the quality of the generated images depends on the architecture of the GAN, the number of training epochs, the dataset size, and other hyperparameters. Feel free to experiment and fine-tune the GAN to achieve better results. Happy generating!

Additional Improvements and Future Directions

Congratulations on completing the GAN for generating fashion images! Now, let’s explore some additional improvements and future directions you can consider to enhance the GAN’s performance and generate even more realistic and diverse fashion images.

Hyperparameter Tuning

Tuning hyperparameters can significantly impact the GAN’s performance. Experiment with different learning rates, batch sizes, number of training epochs, and architecture configurations for the generator and discriminator. Hyperparameter tuning is essential to GAN training, as it can lead to better convergence and more stable results.

Use Progressive Growing

The progressive, growing technique starts training the GAN with low-resolution images and gradually increases the image resolution during training. This approach helps stabilize training and produces higher-quality images. Implementing progressive growth can be more complex but often leads to improved results.

Implement Wasserstein GAN (WGAN)

Consider using the Wasserstein GAN (WGAN) with a gradient penalty instead of the standard GAN loss. WGAN can provide more stable training and better gradients during the optimization process. This can lead to improved convergence and fewer mode collapses.

Data Augmentation

Apply data augmentation techniques to the training dataset. This can include random rotations, flips, translations, and other transformations. Data augmentation helps the GAN generalize better and can prevent overfitting the training set.

Include Label Information

If your dataset contains label information (e.g., clothing categories), you can try conditioning the GAN on the label information during training. This means providing the generator and discriminator with additional information about the clothing type, which can help the GAN generate more category-specific fashion images.

Use a Pretrained Discriminator

Using a pretrained discriminator can help accelerate training and stabilize the GAN. You can train the discriminator on a classification task using the fashion MNIST dataset independently and then use this pretrained discriminator as a starting point for the GAN training.

Collect a Larger and More Diverse Dataset

GANs often perform better with larger and more diverse datasets. Consider collecting or using a larger dataset that contains a wider variety of fashion styles, colors, and patterns. A more diverse dataset can lead to more diverse and realistic generated images.

Explore Different Architectures

Experiment with different generator and discriminator architectures. There are many variations of GANs, such as DCGAN (Deep Convolutional GAN), CGAN (Conditional GAN), and StyleGAN. Each architecture has its strengths and weaknesses, and trying different models can provide valuable insights into what works best for your specific task.

Use Transfer Learning

If you can access pre-trained GAN models, you can use them as a starting point for your fashion GAN. Fine-tuning a pre-trained GAN can save time and computational resources while achieving good results.

Monitor Mode Collapse

Mode collapse occurs when the generator collapses to produce only a few types of images. Monitor your generated samples for signs of mode collapse and adjust the training process accordingly if you notice this behavior.

Building and training GANs is an iterative process, and achieving impressive results often requires experimentation and fine-tuning. Keep exploring, learning, and adapting your GAN to generate even better fashion images!

That concludes our journey in creating a fashion image GAN using TensorFlow and Keras. Feel free to explore other GAN applications, such as generating art, faces, or 3D objects. GANs have revolutionized the field of generative modeling and continue to be an exciting area of research and development in the AI community. Good luck with your future GAN projects!

Conclusion

In conclusion, Generative Adversarial Networks (GANs) represent a cutting-edge technology in artificial intelligence that has revolutionized the creation of synthetic data samples. Throughout this guide, we have gained a deep understanding of GANs and successfully built a remarkable project: a GAN for generating fashion images.

Key Points

1. GANs: GANs consist of two neural networks, the generator, and the discriminator, which use adversarial training to create realistic data samples.
2. Project Goal: We aimed to develop a GAN that generates fashion images resembling those in the Fashion MNIST dataset.
3. Dataset: The Fashion MNIST dataset, with grayscale images of fashion items, served as the basis for our fashion image generator.
4. Building the GAN: We constructed the generator and discriminator using TensorFlow’s Sequential API, incorporating layers like Dense, Conv2D, and LeakyReLU.
5. GAN Training Loop: We employed a carefully designed training loop to optimize the generator and discriminator iteratively.
6. Improvements: We explored several techniques to enhance the GAN’s performance, including hyperparameter tuning, progressive growing, Wasserstein GAN, data augmentation, and conditional GAN.
7. Evaluation: We discussed evaluation metrics such as Inception Score and FID to assess the quality of the generated fashion images objectively.
8. Fine-tuning and Transfer Learning: By fine-tuning the generator and utilizing pretrained models, we aimed to achieve more diverse and realistic fashion image generation.
9. Future Directions: There are countless opportunities for further improvements and research in GANs, including hyperparameter optimization, progressive growing, Wasserstein GAN, and more.

In summary, this comprehensive guide provided a solid foundation for understanding GANs, the intricacies of their training, and how they can be applied to fashion image generation. We demonstrated the potential for creating sophisticated and realistic artificial data by exploring various techniques and advancements. As GANs evolve, they are poised to transform various industries, including art, design, healthcare, and more. Embracing the innovative power of GANs and exploring their limitless possibilities is a thrilling endeavor that will undoubtedly shape the future of artificial intelligence.

Frequently Asked Questions

The media shown in this article is not owned by Analytics Vidhya and is used at the Author’s discretion.