This project implements a GAN in PyTorch to synthesize images from noise. Its Generator upsamples latent vectors via transposed convolutions, BatchNorm, ReLU, and Tanh, while its Discriminator downsamples images with convolutions, LeakyReLU, and Sigmoid. Trained with BCELoss, label smoothing, and Adam, it supports weight initialization, graceful interrupts, Base64 visualization, pre-trained models, real-time loss tracking, and checkpointing. It’s device-agnostic (GPU/CPU) and leverages PyTorch’s ImageFolder for flexible training and monitoring.
This project presents a Variational Autoencoder (VAE) in PyTorch for image reconstruction and generation, featuring a convolutional Encoder-Decoder pair that learns a 256-dimensional latent space using the reparameterization trick. Trained with MSE loss and KL-divergence via Adam optimization, it includes face-mixing by blending latent representations, outputting Base64-encoded images. With mixed precision and checkpointing, it’s optimized for CUDA and showcases latent space exploration using the CelebA-HQ dataset.
This project is a lightweight, educational framework for building and training deep neural networks with NumPy only. By avoiding heavier libraries, it highlights the core math behind forward and backward passes. Each component (layers, activations, losses, optimizers) is written from scratch, making it easy to follow how data flows through a network and how parameters update during training. This is ideal for students or enthusiasts who want hands-on insight into deep learning’s fundamentals without relying on opaque black-box implementations.
This project introduces a system for real-time hand tracking and gesture recognition, enabling users to interact with a digital canvas by sketching or drawing directly with their pointing finger. Designed to detect and interpret hand movements dynamically, it offers an interface for control through natural gestures. Currently, the implementation operates on a CPU, which results in noticeably slower performance when executed within a browser environment, impacting the responsiveness of the experience
This Telegram-based solution streamlines buying, activating, and renewing VPN subscriptions under a SaaS model. Hosted on a dedicated server with WireGuard, it ensures top-notch speed, reliability, and flexible network control. Users interact through a dedicated Telegram bot to subscribe, check their remaining subscription time, and easily request support or updates. The system integrates payment handling, subscription validation, and automated notifications, delivering a smooth, user-friendly experience with minimal oversight.