Creating a CLIP training workflow based on the ViT-B Architecture (Framework: Open CLIP). Dataset used: laion-coco-nllb.

openclip latency performance experiments

Generate flower images from text using a DDPM-based U-Net model. Evaluate results with CLIP Score and FID, analyze embeddings using FiftyOne, and track experiments with Weights & Biases. Tech: PyTorch, OpenCLIP, FiftyOne, WandB