Quantification après entraînement
La quantification post-formation (PTQ) est une technique utilisé en apprentissage automatique, particularly in apprentissage profond models, to optimize the performance of trained models for deployment. This process involves converting the weights and activations of a réseau neuronal from high precision (typically 32-bit floating point) to lower precision formats (such as 8-bit integers). The primary goals of PTQ are to reduce the memory footprint of the model and to accelerate inference times, which is particularly beneficial for running models on edge devices and mobile platforms.
PTQ is typically applied after the model has been fully trained and validated. This means that the model has already learned to perform its task effectively. During PTQ, quantization algorithms analyze the distribution of weights and activations, allowing them to determine how best to map these values to a lower precision format while minimizing the loss of accuracy.
Il existe différentes méthodes de quantification post-formation, notamment :
- Quantification uniforme : This method equally distributes the range of floating-point values into fixed intervals for the integer representation.
- Quantification dynamique: Here, weights are quantized dynamically during inference, which allows for some flexibility and can help maintain accuracy.
- Quantification statique : This approach involves a calibration step where representative input data is used to determine the optimal scale and zero-point for quantization.
Bien que PTQ soit efficace en réduction de la taille du modèle and improving inference speed, it can sometimes lead to a decrease in accuracy. Therefore, it is essential to evaluate the model’s performance post-quantization to ensure that it still meets the required standards for its intended application.