最新研究显示,可微分架构搜索(DARTS)在CIFAR-10数据集上实现了97.2%的Top-1准确率,搜索成本降低至4 GPU-days。对比实验表明:
框架 | 吞吐量(图像/秒) | 延迟(ms) | 内存占用(GB) |
---|---|---|---|
TensorFlow NASNet | 128 ± 5.6 | 23.4 | 8.7 |
PyTorch ProxylessNAS | 152 ± 4.2 | 18.9 | 6.5 |
在移动端设备实现实时推理(batch_size=1):
class EfficientNetB0(nn.Module):
def __init__(self, num_classes=1000):
super().__init__()
self.stem = ConvBNReLU(3, 32, kernel=3, stride=2)
self.blocks = nn.Sequential(
MBConv(32, 16, expansion=1, stride=1, kernel=3),
MBConv(16, 24, stride=2)
Story Source:
Materialsprovided byUniversity of Bristol.Note: Content may be edited for style and length.
Journal Reference:
Akhil S. Kumar, Stefano Dalcanale, Michael J. Uren, James W. Pomeroy, Matthew D. Smith, Justin A. Parke, Robert S. Howell, Martin Kuball.Gallium nitride multichannel devices with latch-induced sub-60-mV-per-decade subthreshold slopes for radiofrequency applications.Nature Electronics, 2025; DOI:10.1038/s41928-025-01391-5