🐛 Describe the bug When combining the new torch.nn.utils.parametrizations.weight_norm() parametrization, torch.compile() fails on it: import torch # Create model. module = torch.nn.Conv3d( in_channels=4, out_channels=4, kernel_size=3, bi...
torch.nn.utils.parametrizations 模块主要用于参数化神经网络层的权重,但它并不包含 weight_norm 功能。weight_norm 是一种正则化技术,用于重新参数化神经网络的权重,使其具有更好的泛化能力。然而,这个功能在 PyTorch 中是通过 torch.nn.utils.weight_norm 提供的,而不是 torch.nn.utils.parametrizations。 查找we...
Tensors and Dynamic neural networks in Python with strong GPU acceleration - Fix nn.utils.parametrizations.weight_norm for BFloat16 (#114785) · pytorch/pytorch@d4128b1