This hugging face repo contains the model checkpoints and loggings.
🔭 Project of Interest:
My primary focus lies in Diffusion Models. I've been deeply engaged with the advancements in this exciting area of generative AI.
I possess a strong theoretical and mathematical understanding, and I have implemented key diffusion model architectures such as DDPM, DDIM, and Latent Diffusion Models.
As an optional third task, I explored Foundation Models. This was a new and enriching experience, providing valuable insights into large-scale pre-training and the adaptation of learned representations for diverse downstream tasks like classification and super-resolution.
📊 1. Common Multi-class Classification:
I fine-tuned a Resnet-18 model (pretrained on ImageNet) to classify gravitational lensing images into the following categories:
- No Substructure
- ️Subhalo Substructure
- Vortex Substructure
Model performance was rigorously evaluated using key metrics: Accuracy, Kappa, ROC-AUC curve, and the MSE loss curve.
🌌 2. Specific Test IV. Diffusion Models:
I successfully trained a DDIM model to generate realistic strong gravitational lensing images.
To accommodate larger batch sizes, the model training was distributed across 2 GPUs.
The foundational architecture code for the DDIM/DDPM implementation draws inspiration from this excellent open-source repository.
🧠 3. Specific Test VI. Foundation Model:
I undertook the pre-training of a ViT as an encoder, coupled with a lightweight decoder, utilizing the Masked AutoEncoder (MAE) technique.
Task VI-A (Classification): I employed the pre-trained ViT with an MLP head as the encoder for classification tasks.
Task VI-B (Super-Resolution): I integrated the pre-trained ViT with a more powerful decoder to perform super-resolution.
This insightful blog post proved invaluable for understanding and implementing the Masked AutoEncoder.
✨ It was a truly engaging project, and I'm delighted to have had such an exciting and insightful learning experience. Thank you for the opportunity! ✨
