摘要：
火炬的y不需要转换成一个热的。它会自动帮助
Torch笔记
import torch
import numpy as np
import torch.nn as nn
a_np = np.random.rand(10,100)
numpy知识回顾
a_np.dtype # 数据类型
a_np.ndim #维度个数
a_np.shape # 形状 整数元祖
a_np.dtype=np.int32 # 修改数据类型
获取tensor基本信息
tensor_a = torch.from_numpy(a_np)
tensor_a.type() # 获取tensor类型
tensor_a.size() # 获取维度特征 并返回数组
tensor_a.size(0) #获取第一个维度信息
tensor_a.dim() # 获取维度个数
tensor_a.device # 返回设备类型
tensor数据类型转换
tensor_b = tensor_a.float()
tensor_b = tensor_a.to(torch.float)
设备类型转换
tensor_b = tensor_b.cuda()
tensor_b = tensor_b.cpu()
tensor_b = tensor_a.to(torch.device(0))
tensor 转换到 numpy数据
tensor_b = tensor_b.numpy()
# 注意*gpu上的tensor不能直接转为numpy
ndarray = tensor_b.cpu().numpy()
# numpy 转 torch.Tensor
tensor = torch.from_numpy(ndarray)
tensor的维度变换
- squeeze_
tensor_a.t_().size() # 对tensor进行转置
tensor_a.t_().size()
tensor_a.unsqueeze_(2) #
unsqu_tensor = tensor_a.squeeze_() # 将tensor中维度元素数为1 的全部压缩掉
- reshape()
tensor = torch.reshape(tensor_a, [50,20])
# 也可以用tensor.reshape(50,20)
tensor.size()
- tanspose()
tensor.transpose(1,2) #交换tensor的1，2维度
从只包含一个元素的张量中提取值
tensor_a = tensor_a.cuda()
tensor_a[1][1].item()
tensor_a.device
拼接张量
- 例如当参数是3个10×5的张量，torch.cat的结果是30×5的张量，而torch.stack的结果是3×10×5的张量。
temp1 = torch.from_numpy(np.random.rand(5,4))
temp2 = torch.from_numpy(np.random.rand(5,4))
temp1.size()
temp2.size()
temp3 = torch.stack([temp1,temp2],dim=0)
temp4 = torch.cat([temp1,temp2],dim=0)
temp3.size()
temp4.size()
矩阵运算
- 矩阵乘法
# Matrix multiplication: (m*n) * (n*p) -> (m*p).
result = torch.mm(tensor1, tensor2)
NOTE:This function does not broadcast. For broadcasting matrix products, see torch.matmul().
result = torch.matmul(tensor1, tensor2) # 支持broadcast
- tensor中对应元素相乘
tensor1*tensor2 # 支持broadcast
- 矩阵求最大组索引
res,index = torch.max(tensor,dim=-1)
维度理解 ：填入哪个维度则哪个维度消失 比如sum，softmax
- 求平均值 ==avg_pooling
avg = torch.mean(tensor,dim=-1)
打印模型信息
- 参数量
- 模型结构 使用torch summary
class myNet(nn.Module):
def __init__(self,*other_para):
super(myNet,self).__init__()
self.embedding_layer = nn.Embedding(10,3)
net = myNet()
num_parameters = sum(torch.numel(parameter) for parameter in net.parameters())
from torchsummary import summary
summary(net,input_size=(2,2))
模型初始化
# Common practise for initialization.
for layer in model.modules():
if isinstance(layer, torch.nn.Conv2d):
torch.nn.init.kaiming_normal_(layer.weight, mode='fan_out',
nonlinearity='relu')
if layer.bias is not None:
torch.nn.init.constant_(layer.bias, val=0.0)
elif isinstance(layer, torch.nn.BatchNorm2d):
torch.nn.init.constant_(layer.weight, val=1.0)
torch.nn.init.constant_(layer.bias, val=0.0)
elif isinstance(layer, torch.nn.Linear):
torch.nn.init.xavier_normal_(layer.weight)
if layer.bias is not None:
torch.nn.init.constant_(layer.bias, val=0.0)
# Initialization with given tensor.
layer.weight = torch.nn.Parameter(tensor)
计算Softmax输出的准确率
score = model(images)
prediction = torch.argmax(score, dim=1)
num_correct = torch.sum(prediction == labels).item()
accuruacy = num_correct / labels.size(0)
保存模型
- torch.save(my_model.state_dict(), "params.pkl")
加载模型
- 先初始化model网络结构
- model.load_state_dict(torch.load("params.pkl"))
- torch.nn.CrossEntropyLoss的输入不需要经过Softmax。torch.nn.CrossEntropyLoss等价于torch.nn.functional.log_softmax + torch.nn.NLLLoss。
- torch的y不用转成onehot 它会自动帮你转 这点比较怪异，相比之下keras在多分类时 y必须是onehot的才能计算损失
- model.train() ：启用 BatchNormalization 和 Dropout
- model.eval() ：不启用 BatchNormalization 和 Dropout
- 用del及时删除不用的中间变量，节约GPU存储。
- 使用inplace操作可节约GPU存储，如