1、基本配置

1.1 导入包和版本查询

import torch
import torch.nn as nn
import torchvision
print(torch.__version__)
print(torch.cuda.is_available())
print(torch.version.cuda)
print(torch.backends.cudnn.version())
print(torch.cuda.get_device_name(0))

1.2 可复现性以及设置随机种子

np.random.seed(0)  
torch.manual_seed(0)  
torch.cuda.manual_seed_all(0)  
  
torch.backends.cudnn.deterministic = True  
torch.backends.cudnn.benchmark = False  

1.3 显卡设置以及清除显存

# 利用 PyTorch 的接口指定单张显卡
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')  

# 如果需要指定多张显卡，比如0，1号显卡。
import os  
os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'  

# 也可以在命令行运行代码时设置显卡：
CUDA_VISIBLE_DEVICES=0,1 python train.py  


# 清除显存
torch.cuda.empty_cache()  

# 也可以使用在命令行重置GPU的指令
nvidia-smi --gpu-reset -i [gpu_id]  

2、张量处理

2.1 查看张量的基本信息

tensor = torch.randn(3,4,5)
print(tensor.type())  # 数据类型
print(tensor.size())  # 张量的shape，是个元组
print(tensor.shape)  # 张量的shape，是个元组
print(tensor.dim())   # 维度的数量

2.2 torch.Tensor 和 np.ndarray 互相转换

# np.ndarray -> torch.Tensor
ndarray = tensor.cpu().numpy()
tensor = torch.from_numpy(ndarray).float()
tensor = torch.from_numpy(ndarray.copy()).float() # If ndarray has negative stride.

# torch.Tensor -> np.ndarray
tensor = torch.randn(3,4,5)
array = tensor.cpu().detach().numpy()

2.3 torch.tensor 与 PIL.Image 转换

# pytorch中的张量默认采用[N, C, H, W]的顺序，并且数据范围在[0,1]，需要进行转置和规范化
# torch.Tensor -> PIL.Image
image = PIL.Image.fromarray(
    torch.clamp(tensor*255, min=0, max=255).byte().permute(1,2,0).cpu().numpy())
image = torchvision.transforms.functional.to_pil_image(tensor)  # Equivalently way

# PIL.Image -> torch.Tensor
path = 'figure.jpg'
tensor = torch.from_numpy(
    np.asarray(PIL.Image.open(path))
).permute(2,0,1).float() / 255

# PIL.Image -> torch.Tensor
tensor = torchvision.transforms.functional.to_tensor(PIL.Image.open(path)) 

2.4 张量的一些操作

# 1、从只包含一个元素的张量中提取值
value = torch.rand(1).item()

# 2、张量形变
tensor = torch.rand(2,3,4)
shape = (6, 4)
tensor = torch.reshape(tensor, shape)

# 3、张量顺序重排
tensor = torch.rand(2,3,4)
tensor = tensor.permute(2,0,1)

# 4、复制张量
# Operation                 |  New/Shared memory | Still in computation graph |
tensor.clone()            # |        New         |          Yes               |
tensor.detach()           # |      Shared        |          No                |
tensor.detach.clone()()   # |        New         |          No                |

# 5、张量拼接
# 注意 torch.cat 和 torch.stack 的区别在于 torch.cat 沿着给定的维度拼接，
# 而 torch.stack 会新增一维。例如当参数是 3 个 10x5 的张量，torch.cat 的结果是 30x5 的张量，
# 而 torch.stack 的结果是3x10x5的张量。

tensor = torch.cat(list_of_tensors, dim=0)
tensor = torch.stack(list_of_tensors, dim=0)

# 6、得到非零元素
torch.nonzero(tensor)               # index of non-zero elements
torch.nonzero(tensor==0)            # index of zero elements
torch.nonzero(tensor).size(0)       # number of non-zero elements
torch.nonzero(tensor == 0).size(0)  # number of zero elements

# 7、判断两个张量相等
torch.allclose(tensor1, tensor2)  # float tensor
torch.equal(tensor1, tensor2)     # int tensor

# 8、张量扩展
# Expand tensor of shape 64*512 to shape 64*512*7*7.
tensor = torch.rand(64,512)
torch.reshape(tensor, (64, 512, 1, 1)).expand(64, 512, 7, 7)

# 9、矩阵乘法
# Matrix multiplcation: (m*n) * (n*p) * -> (m*p).
result = torch.mm(tensor1, tensor2)

# Batch matrix multiplication: (b*m*n) * (b*n*p) -> (b*m*p)
result = torch.bmm(tensor1, tensor2)

# Element-wise multiplication.
result = tensor1 * tensor2

2.5 将整数标签转为 one-hot 编码

# pytorch的标记默认从0开始
tensor = torch.tensor([0, 2, 1, 3])
N = tensor.size(0)
num_classes = 4
one_hot = torch.zeros(N, num_classes).long()
one_hot.scatter_(dim=1, index=torch.unsqueeze(tensor, dim=1), src=torch.ones(N, num_classes).long())

3、timm 模型提取特征流程

3.1 timm 提取特征

def forward_features(self, x):
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.act1(x)
    x = self.maxpool(x)

    if self.grad_checkpointing and not torch.jit.is_scripting():
        x = checkpoint_seq([self.layer1, self.layer2, self.layer3, self.layer4], x, flatten=True)
    else:
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
    return x

def forward_head(self, x, pre_logits: bool = False):
    x = self.global_pool(x)
    if self.drop_rate:
        x = F.dropout(x, p=float(self.drop_rate), training=self.training)
    return x if pre_logits else self.fc(x)

def forward(self, x):
    x = self.forward_features(x)
    x = self.forward_head(x)
    return x

3.2 timm 展示出所有的模型

参考资料

• 部分内容转载自：https://zhuanlan.zhihu.com/p/104019160