«««< HEAD
layout: post title: “torch” subtitle: “ "工具"” date: 2020-10-28 18:00:00 mathjax: true author: “zwt” header-img: “img/post-bg-2015.jpg” catalog: false tags: - 工具 —
预备知识
设置随机种子
1
2
3
4
5
6
7
8
9
10
11
import torch
import numpy as np
import random
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
# 设置随机数种子
setup_seed(20)
tensor
创建未初始化的Tensor:5*3
1
x = torch.empty(5,3)
创建随机初始化Tensor
1
x = torch.rand(,3)
创建全为0的Tensor
1
x = torch.zeros(5,3,dtype=torch.long)
返回的tensor默认具有相同的dtype和device
1
x.new_ones(5,3,dtype=torch.float64)
指定新的数据类型
1
torch.randn_like(x, dtype=torch.float)
获取形状
1
2
torch.size()
torch.shape
加法，可以指定输出
1
2
result = torch.empty(5,3)
torch.add(y,y,out=result)
模型保存：
1
2
3
4
5
6
7
8
9
10
11
只保存参数：
# 保存
torch.save(model.state_dict(), '\parameter.pkl')
# 加载
model = TheModelClass(...)
model.load_state_dict(torch.load('\parameter.pkl'))
保存完整模型：
# 保存
torch.save(model, '\model.pkl')
# 加载
model = torch.load('\model.pkl')
sentence2teansformer
转onnx
1
2
3
pip install transformers[onnx]
python -m transformers.onnx --model=./models/my-128dim-model onnx/
python -m transformers.onnx --model=./output/training_multi-task-learning2 --atol=2e-5 onnx/v1/
tensorboard
1
2
3
4
5
6
7
8
9
10
11
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter('./runs')
writer.add_scalar('LOSS/Train_loss', float(avg_train_loss),(epoch + 1))
writer.add_scalar('LOSS/Valid_loss', float(avg_valid_loss), (epoch + 1))
writer.add_scalar('ACC/Train_lacc', float(avg_train_acc), (epoch + 1))
writer.add_scalar('ACC/Valid_lacc', float(avg_valid_acc), (epoch + 1))
cd runs/
tensorboard --logdir ./(writer地址)
或：
tensorboard --logdir=./(writer地址)
参考
=======
layout: post title: “torch” subtitle: “ "工具"” date: 2020-10-28 18:00:00 mathjax: true author: “zwt” header-img: “img/post-bg-2015.jpg” catalog: false tags: - 工具 —
- TOC
预备知识
设置随机种子
1
2
3
4
5
6
7
8
9
10
11
import torch
import numpy as np
import random
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
# 设置随机数种子
setup_seed(20)
tensor
创建未初始化的Tensor:5*3
1
x = torch.empty(5,3)
创建随机初始化Tensor
1
x = torch.rand(,3)
创建全为0的Tensor
1
x = torch.zeros(5,3,dtype=torch.long)
返回的tensor默认具有相同的dtype和device
1
x.new_ones(5,3,dtype=torch.float64)
指定新的数据类型
1
torch.randn_like(x, dtype=torch.float)
获取形状
1
2
torch.size()
torch.shape
加法，可以指定输出
1
2
result = torch.empty(5,3)
torch.add(y,y,out=result)
模型保存：
1
2
3
4
5
6
7
8
9
10
11
只保存参数：
# 保存
torch.save(model.state_dict(), '\parameter.pkl')
# 加载
model = TheModelClass(...)
model.load_state_dict(torch.load('\parameter.pkl'))
保存完整模型：
# 保存
torch.save(model, '\model.pkl')
# 加载
model = torch.load('\model.pkl')
sentence2teansformer
转onnx
1
2
3
pip install transformers[onnx]
python -m transformers.onnx --model=./models/my-128dim-model onnx/
python -m transformers.onnx --model=./output/training_multi-task-learning2 --atol=2e-5 onnx/v1/
tensorboard
1
2
3
4
5
6
7
8
9
10
11
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter('./runs')
writer.add_scalar('LOSS/Train_loss', float(avg_train_loss),(epoch + 1))
writer.add_scalar('LOSS/Valid_loss', float(avg_valid_loss), (epoch + 1))
writer.add_scalar('ACC/Train_lacc', float(avg_train_acc), (epoch + 1))
writer.add_scalar('ACC/Valid_lacc', float(avg_valid_acc), (epoch + 1))
cd runs/
tensorboard --logdir ./(writer地址)
或：
tensorboard --logdir=./(writer地址)
参考
d5005ecd9eaf3ce32260dc1b7d831c7f2a1a85f3
