Huawei cloud server, 4-core 8G memory, no graphics, performance calculation improvise, to catch up with double 11 was less than 1000, it can also be cost-effective, plan to configure a set of training densenet environment.
First comes python version 2.7, due next year, it will no longer maintained, so installed a conda.
wget https://repo.continuum.io/archive/Anaconda3-5.3.0-Linux-x86_64.sh
I found too slow, looking for a source of Tsinghua University, a lot faster.
wget https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/Anaconda3-4.3.0-Linux-x86_64.sh
(Here 4.3.0 can be replaced 5.3.0,4.3.0 the python version 3.6 seems to be, back when you install pytorch and torchvision may also need to upgrade python version)
chmod 777 anaconda3.4.3.0-Linux-x86_64.sh
./anaconda3.4.3.0-Linux-x86_64.sh
Has been yes, after installation, you will be prompted whether to join the system variables, click yes, and then execute the command: source ~ / .bachrc
Enter python --version to check whether the upgrade from version 2.7 to 3.6
Enter conda list, check whether the installed conda
Then install pytorch and torchvision
conda config --add channels https://repo.continuum.io/pkgs/free/
conda config --add channels https://repo.continuum.io/pkgs/main/
conda config --set show_channel_urls yes
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
conda install pytorch torchvision
Can also conda install pytorch torchvision cudatoolkit = 10.0.130, because there is no cuda graphics card, so do not take it does not matter.
Also this: Conda install pytorch -c pytorch
conda install torchvision -c pytorch
After installed, use the test program
import torch
train_on_gpu = torch.cuda.is_available()
if not train_on_gpu:
print('Training on CPU ...')
else:
print('Training on GPU ...')
The data sets uploaded to the server, start training.
1: lr_scheduler reported a fault, look to upgrade pytorch and torchvision
conda install pytorch==0.4.0
conda upgrade torchvision
2: Report a mistake, pytorch: ValueError: optimizing a parameter that does not require gradients
But this wrong, there has never been the same in the windows environment and centos7 + python 2.7 both environments.
There are two solutions, one is the param.requires_grad = False ➡param.requires_grad = True, but the memory may not carry.
Another option is to optimizer = optim.Adadelta (model.parameters ()) ➡ optimizer = optim.Adadelta (filter (lambda p: p.requires_grad, model.parameters ()))
The second method is recommended, because the first method the derivative accounting for slower memory.
3：报了个错：json.decoder.JSONDecodeError: Expecting property name enclosed in double quo
Json file seems to last more than a "," No, exclude this log to get the wrong watch.