user@devops:~$ cat README.md
DCGAN Generative
# Description
DCGAN (Deep Convolutional Generative Adversarial Network) implemented from scratch in PyTorch to generate synthetic grayscale images of clothing items using the Fashion-MNIST dataset. The Generator (215K parameters) uses Dense + Upsample + Conv2d layers to transform random noise vectors (latent_dim=100) into 28×28 grayscale images. The Discriminator (10K parameters) classifies images as real or generated using strided convolutions with LeakyReLU. Alternating adversarial training with label smoothing (0.9/0.1), Adam optimizer (lr=2e-4), 25 epochs on 4000 Fashion-MNIST samples. Results: D_loss=0.85, G_loss=1.55, D_accuracy=95%. Includes temporal evolution visualization with samples every 5 epochs, D vs G loss curves, and an 8×8 final grid of generated images.
# Key features
$ Generator with 3 convolutional layers (Upsample + Conv2d) for gradual 8×8 → 14×14 → 28×28 upscaling
$ Discriminator with 2 strided Conv2d layers, BatchNorm, LeakyReLU, and Dropout for stability
$ Label smoothing (0.9 real / 0.1 fake) to prevent the Discriminator from becoming overconfident
$ Normal(0, 0.02) weight initialization as recommended by the DCGAN paper
$ Visual evolution: samples saved every 5 epochs to observe progressive improvement
$ Alternating G/D adversarial training with Adam optimizer (β₁=0.5)
$ Final 8×8 grid showing 64 synthetically generated images
$ Fashion-MNIST dataset with 10 clothing categories (t-shirts, shoes, bags, etc.)
# Gallery
# Technologies used
