算能SG2380混合精度进阶
算能SG2380混合精度进阶🧮 INT4量化实战python
复制
下载
from sophgo import quantmodel = quant.CalibratedModel('resnet18.onnx')model.quantize( precision='int4', calibration_data=dataset, dynamic_ranges=True # 启用动态范围)model.export('resnet18_int4.sophgo')
📊 精度保持率
模型INT32精度INT4精度损失率
ResNet1869.8%68.1%2.4%
MobileNetV375.2%74.5%0.9%
页:
[1]