Quantification of Flow-Matching Models
Budget: ₹1,500 – ₹12,500 INR
I am seeking a skilled deep learning researcher/engineer to collaborate on a new project. I have a foundational paper and codebase on the quantization of diffusion models (BiDM) and the official Flow Matching for Generative Models implementation from Facebook Research.
The objective is to apply the binarization/quantization concept from BiDM to Flow Matching models to achieve compact, efficient, and high-quality generative performance. The goal is to produce results suitable for top-tier academic publications.
Key Objectives
• Integrate BiDM-style binarization into the Flow Matching framework.
• Evaluate the impact of quantization on efficiency, generation quality, and training stability.
• Establish a comprehensive benchmark for quantized Flow Matching models.
Key Requirements and Deliverables
1. Implementation
• Develop a quantized Flow Matching model in PyTorch and CUDA.
• Incorporate binarized weights and/or activations with minimal performance loss.
• Implement mechanisms for quantization-aware fine-tuning and efficient inference.
2. Experimental Validation
• Train and evaluate on CIFAR-10, CelebA-HQ, FFHQ, and optionally LSUN-Church or AFHQ.
• Report FID, IS, LPIPS, and KID scores.
• Provide visual and numerical comparisons with full-precision models.
3. Performance Analysis
• Measure model size reduction, inference speedup, and memory usage.
• Analyze the trade-off between compression and generative quality.
4. Comparative Study
• Compare against full-precision Flow Matching, BiDM, and other quantized generative baselines.
5. Reproducibility
• Deliver reusable scripts/notebooks for full pipeline execution.
• Include detailed documentation of hyperparameters, environment setup, and dependencies.
Technical Stack
• Frameworks: PyTorch, CUDA
• Datasets: CIFAR-10, CelebA-HQ, FFHQ, LSUN-Church, AFHQ
• Metrics: FID, IS, LPIPS, KID
• Hardware: GPU (A100/V100 or similar)
Expected Outcomes
• A quantized Flow Matching model demonstrating significant compression and speedup with minimal quality degradation.
• Comprehensive benchmarking and reproducible experimental setup suitable for publication.
The objective is to apply the binarization/quantization concept from BiDM to Flow Matching models to achieve compact, efficient, and high-quality generative performance. The goal is to produce results suitable for top-tier academic publications.
Key Objectives
• Integrate BiDM-style binarization into the Flow Matching framework.
• Evaluate the impact of quantization on efficiency, generation quality, and training stability.
• Establish a comprehensive benchmark for quantized Flow Matching models.
Key Requirements and Deliverables
1. Implementation
• Develop a quantized Flow Matching model in PyTorch and CUDA.
• Incorporate binarized weights and/or activations with minimal performance loss.
• Implement mechanisms for quantization-aware fine-tuning and efficient inference.
2. Experimental Validation
• Train and evaluate on CIFAR-10, CelebA-HQ, FFHQ, and optionally LSUN-Church or AFHQ.
• Report FID, IS, LPIPS, and KID scores.
• Provide visual and numerical comparisons with full-precision models.
3. Performance Analysis
• Measure model size reduction, inference speedup, and memory usage.
• Analyze the trade-off between compression and generative quality.
4. Comparative Study
• Compare against full-precision Flow Matching, BiDM, and other quantized generative baselines.
5. Reproducibility
• Deliver reusable scripts/notebooks for full pipeline execution.
• Include detailed documentation of hyperparameters, environment setup, and dependencies.
Technical Stack
• Frameworks: PyTorch, CUDA
• Datasets: CIFAR-10, CelebA-HQ, FFHQ, LSUN-Church, AFHQ
• Metrics: FID, IS, LPIPS, KID
• Hardware: GPU (A100/V100 or similar)
Expected Outcomes
• A quantized Flow Matching model demonstrating significant compression and speedup with minimal quality degradation.
• Comprehensive benchmarking and reproducible experimental setup suitable for publication.