Quantization-Aware Training (QAT)により、エッジデバイス向けのAIモデルを最適化し、リソースが限られた環境でも高い精度と効率を実現します。
Quantization-Aware Training (QAT) is a powerful technique used to optimize deep learning (DL) models, like Ultralytics YOLO models, for deployment on devices with limited computational resources, such as mobile phones or embedded systems. Standard models often use high-precision numbers (like 32-bit floating-point or FP32) for calculations, which demand significant processing power and memory. QAT aims to reduce this demand by preparing the model during the training phase to perform well even when using lower-precision numbers (e.g., 8-bit integers or INT8), thereby bridging the gap between high accuracy and efficient performance on edge devices. This optimization is crucial for enabling complex AI tasks directly on hardware like smartphones or IoT sensors.
Unlike methods that quantize a model after it has been fully trained, QAT integrates the simulation of quantization effects directly into the training process. It introduces operations called 'fake quantization' nodes within the model architecture during training. These nodes mimic the effect of lower precision (e.g., INT8 precision) on model weights and activations during the forward pass, rounding values as they would be in a truly quantized model. However, during the backward pass (where the model learns via backpropagation), gradients are typically calculated and updates applied using standard high-precision floating-point numbers. This allows the model's parameters to adapt and learn to be robust to the precision loss that will occur during actual quantized inference. By "seeing" the effects of quantization during training, the model minimizes the accuracy drop often associated with deploying models in low-precision formats, a key aspect discussed in model optimization strategies. Frameworks like TensorFlow Lite and PyTorch provide tools to implement QAT.
The primary difference lies in when quantization is applied. Model Quantization, often referring to Post-Training Quantization (PTQ), converts a pre-trained, full-precision model to a lower-precision format after training is complete. PTQ is generally simpler to implement as it doesn't require retraining or access to the original training dataset. However, it can sometimes lead to a noticeable decrease in model accuracy, especially for complex models performing tasks like object detection or image segmentation. QAT, by contrast, simulates quantization during training, making the model inherently more robust to precision reduction. This often results in higher accuracy for the final quantized model compared to PTQ, although it requires more computational resources and access to training data. For models like YOLO-NAS, which incorporates quantization-friendly blocks, QAT can yield significant performance benefits with minimal precision loss.
While both techniques involve numerical precision, their goals differ. Mixed Precision training primarily aims to speed up the training process itself and reduce memory usage during training by using a combination of lower-precision (e.g., 16-bit float or FP16) and standard-precision (32-bit float) formats for computations and storage. QAT specifically focuses on optimizing the model for efficient inference using low-precision integer formats (like INT8) after model deployment. While mixed precision helps during training, QAT ensures the final model performs well under the constraints of quantized inference hardware, such as NPUs (Neural Processing Units) or TPUs.
量子化を考慮したトレーニングは、効率が重要なリソース制約のある環境で高度なAIモデルを展開するために不可欠です。
Ultralytics supports exporting models to various formats like ONNX, TensorRT, and TFLite, which are compatible with QAT workflows, enabling efficient deployment across diverse hardware. You can manage and deploy your QAT-optimized models using platforms like Ultralytics HUB. Evaluating model performance using relevant metrics after QAT is essential to ensure accuracy requirements are met.
JA
EN
ZH
KO
RU
DE
FR
ES
PT
IT
AR
TR
VI
