In this project , I implemenetd a complete end-to-end pipeline for :
- analyzing Algeria’s climate
- forecasting future trends
- building an interactive intelligent agent for querying climate information.
- It leverages historical climate data, machine learning, reinforcement learning, and interactive visualization.
In this Notebook I implemenetd a Generative Adversarial Network (GAN) to generate realistic cartoon faces.
Using the cartoon-faces-googles-cartoon-set dataset , the model consists of : - a Generator that transforms random latent vectors into 64×64 RGB images
- a Discriminator that classifies images as real or fake.
- Through adversarial training, the Generator gradually learns to produce images that can fool the Discriminator, while the Discriminator continually improves its ability to distinguish real from generated images.
- The training process demonstrates the characteristic GAN dynamics, with alternating improvements in both Generator and Discriminator performance.
Keywords: Generative Adversarial Network, image synthesis, cartoon faces, adversarial training, PyTorch.
This notebook presents my implementation of a U-Net autoencoder for removing watermarks from images.
- Using the CLWD dataset, the model learns a mapping from watermarked images to their clean counterparts.
- The encoder extracts hierarchical features and compresses the input, the bottleneck captures high-level patterns differentiating watermark and content, and the decoder reconstructs the clean image.
- Skip connections allow fine details from the encoder to aid reconstruction.
- The network is trained with a combined loss of pixel-wise L1 and SSIM, balancing numerical accuracy with perceptual similarity.
- Results demonstrate effective watermark removal while preserving image details, with steady convergence on both training and validation sets.
Keywords: U-Net, autoencoder, watermark removal, image restoration, structural similarity, TensorFlow/Keras.
In this Notebook , I explored the use of Conditional Variational Autoencoders (CVAE) for facial expression generation.
- Using the CelebA dataset, a CVAE was trained to generate smiling faces from non-smiling inputs by conditioning both the encoder and decoder on a binary smile label.
- The model learns a structured latent representation of facial features, enabling controlled manipulation of expressions while preserving individual identity.
- Evaluation demonstrates successful generation of realistic smiling faces .
Keywords: Variational Autoencoder, conditional generation, facial expression, CelebA dataset, latent representation.
In this notebook I implemented a Transformer neural network for forecasting hourly household electricity consumption.
- The model leverages self-attention mechanisms and positional encoding to capture temporal dependencies in sequential data.
- Using past consumption patterns, the model predicts future values, demonstrating high accuracy in short-term forecasting.
- Results indicate that Transformers can effectively model complex temporal structures in time series data, providing a scalable approach for energy demand prediction and temporal pattern learning.
Keywords: Transformer, time series forecasting, electricity consumption, self-attention, temporal modeling.