Optimize Image CNN Performance
Budget: $30 – $250 USD
I’m working on a personal research project that relies on convolutional neural networks to analyse image data, and I’ve reached the point where the baseline architecture no longer meets my goals. I need the model to be both more accurate and noticeably faster at inference without sacrificing one for the other.
You’ll start with my existing checkpoints and training scripts (Python, TensorFlow/Keras—PyTorch equivalents are fine if you would rather port). The task is to redesign or fine-tune the network, apply the right data-augmentation or regularisation tricks, and then compress, prune, quantise, or otherwise accelerate it so that I see measurable gains in both metrics. I’m happy to experiment with advanced techniques such as knowledge distillation or Neural Architecture Search as long as the final solution remains reproducible on a single-GPU workstation.
Deliverables
• Updated training and inference code with clear comments
• Trained model weights reflecting the improved accuracy and speed
• A concise report comparing baseline vs. enhanced results (dataset, hardware specs, accuracy %, average inference time per image)
I’ll provide the dataset and the current benchmark numbers as soon as we begin. Let me know which frameworks or optimisation libraries you prefer and how you plan to validate each improvement.
You’ll start with my existing checkpoints and training scripts (Python, TensorFlow/Keras—PyTorch equivalents are fine if you would rather port). The task is to redesign or fine-tune the network, apply the right data-augmentation or regularisation tricks, and then compress, prune, quantise, or otherwise accelerate it so that I see measurable gains in both metrics. I’m happy to experiment with advanced techniques such as knowledge distillation or Neural Architecture Search as long as the final solution remains reproducible on a single-GPU workstation.
Deliverables
• Updated training and inference code with clear comments
• Trained model weights reflecting the improved accuracy and speed
• A concise report comparing baseline vs. enhanced results (dataset, hardware specs, accuracy %, average inference time per image)
I’ll provide the dataset and the current benchmark numbers as soon as we begin. Let me know which frameworks or optimisation libraries you prefer and how you plan to validate each improvement.