Otimizando o uso da GPU no Google Colab

I was doing a computer vision project for my master's degree and, like all transformer training, it was taking too long. I considered paying for Google Colab Pro to use the A100 GPU, but a colleague told me they didn't see much difference compared to the free T4 GPU. I found that with some model adjustments, it's possible to optimize GPU usage. I'll explain some optimizations that helped me accelerate fine-tuning significantly. My project involved detecting violence in videos, which requires processing enormous amounts of numerical data. Using mixed precision training, the system decides to perform most operations using "half-precision" (FP16), which uses half the memory and is faster on GPUs. The model's "knowledge" (weights) is still stored with "full precision" (FP32) to avoid losing important details. Adjusting the batch size allows the GPU to process more data in parallel, making better use of its resources. Additionally, a larger batch size provides a more accurate and less noisy estimate of the gradient, leading to a more stable optimization path.