weixin_43918682的博客

gitlab

可能会遇到有一个多gpu的训练后保存的模型，但是后续要在单gpu的机子上使用，在torxh.load时报错。这是因为nn.DataParallel会在模型参数结构前面加一个module.比如你是这么save的state = {'epoch': epoch, 'state_dict': self.model.state_dict(), 'optimizer': self.optimizer.state_dict(), '

PointTransformerSeg((backbone): Backbone((fc1): Sequential((0): Linear(in_features=19, out_features=32, bias=True)(1): ReLU()(2): Linear(in_features=32, out_features=32, bias=True))(transformer1): TransformerBlock((fc1): Linear(in_features=32, out_

pytorch中的 Diceloss 记录https://github.com/BloodAxe/pytorch-toolbelt/blob/0.4.1/pytorch_toolbelt/losses/dice.py输入：y_pred: NxCxHxW y_true: NxHxWfrom typing import Listimport torchimport torch.nn.functional as Ffrom pytorch_toolbelt.utils.torch

get_device2 numpy的item() 获取对应的值np.random.seed(123)x = np.random.randint(9, size=(3, 3))xarray([[2, 2, 6], [1, 3, 6], [1, 0, 1]])x.item(3)1x.item(7)0x.item((0, 1))2x.item((2, 2))1tensor中的item（），获取值并保留2位小数>>> x = tor

模型打印理解my_model：ptBEVnet((PPmodel): Sequential((0): BatchNorm1d(9, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(1): Linear(in_features=9, out_features=64, bias=True)(2): BatchNorm1d(64, eps=1e-05, momentum=0.1, affine=True, track_ru

Swin transformer学习记录新一代backbone，！！核心为Swin Transformer block：MSA+ SW_MSA实现了类似CNN的pooling效果

FCN 论文代码解读二–模型构建FCN参数类别数class FCN(nn.Module): def __init__(self, num_classes): super().__init__() self.stage1 = pretrained_net.features[:7] # 一第一个block conv relu bn pooling 一共6层 self.stage2 = pretrained_net.features[7:14]

FCN论文以及源码拆分详解（一）FCN 论文：Fully Convolutional Networks for Semantic Segmentation参考github代码。摘要：开山之作-----state of-the-art segmentationOur key insight is to build “fully convolutional” networks that take input of arbitrary size and produce correspondingly-s