This folder presents the installation that we made on a Lenovo Legion Y920 with i7 and a GPU NVIDIA GeForce 1070 in my former company. Having laptops was for a better agility for our work in Applied Deep learning. This guide would give the possibility to anyone to build their own installation so that they can practice freely. Although this guide is straight forward and ends up in a perfectly working setup, you will need some linux skills in order to fix the issues that will be coming after this installation (if you update the cuda drivers or other things that might eventually happen).
This installation was made on a specific model of computer and you might have issues transfering it to another, but it gives a lot of solution to a wide range of problems we encountered. I believe it is a rather complete guide.
By following this guide you'll install this environement:
On a Lenovo Legion Y920 - Core i7-7820HK - 16 Go - SSD 256 Go + HDD 1 To - Screen 17.3" LED Full HD - NVIDIA GeForce GTX 1070 8 Go you'll have:
- A dual boot Ubuntu 18.04 LTS/ Windows 10
- Cuda 10.0
- Cudnn 7.6
- Python 3.7
- Tensorflow-gpu 2.0.0
- Pytorch 1.4 for cuda 10.0
Along with other libraries that are quite handy for manipulating data and visualizations
You'll need:
- A bootable key with Ubuntu 18.04
- A paperclip (or any pointy thing you'll find)
- Patience (seriously)
- The freshly bought PC
To install ubuntu we will have to disable some of windows functionalities. Once you are in windows:
- Open a command line
- Type
diskmgmt.mscand enter - The window of the partitions of the PC opens
- Delete the partition on the Hard Drive (1 To)
Then you need to disable the "fast boot":
- Go to the parameters
- Select "Power Management"
- Clic on "Change parameters currently unavailable"
- Untick "fast boot (recommanded)"
- Clic on "Save changes"
On this computer Windows is installed on the SSD, we will therefore install Ubuntu on the Hard Drive.
Enter the BIOS of the computer (on the computer you have a button on the right side which can be pressed with a pointy tool)
On the configuration tab, change "SATA Controller mode" to "AHSI" save and quit the BIOS, the computer attemps to start on windows, turn it off
Plug the Ubuntu 18.04 bootable key
Start using the button on the right side to enter BIOS menu, start the computer on the USB drive
Ubuntu opens and clic on install Ubuntu, select your langage. You can decide to just wipe out Windows and install ubuntu instead on the SSD and skip to the next section but here we will choose "Something Else" with "normal installation".
In the space we just freed on the hard drive we will create 3 partitions:
Select the 1 To space and clic on "+"
- Primary partition
- Beginning of this space (200 000 Mo)
- File ext4
- Mounting point "/"
On the remaining free space, clic on "+":
- Primary partition
- Beginning of this space (1024 Mo)
- Swap file
On the remaining free space, clic on "+":
- Primary partition
- Beginning of this space (all the remaining space)
- File ext4
- Mounting point "/home"
Device where the booting program will be installed: "NVMe device", the SSD
Clic on install now
Ubuntu will ask you some other information, for the password, select something easy to type because the installations will ask for it an awful lot
Reboot the computer, once the screen is black (but not off) remove the bootable key and press enter to turn off.
Remark: if you want to start Windows you'll have to return to the BIOS and change back the SATA Controler Mode back to intel RST
On this PC we noticed that the installation of Ubuntu had some unexpected results, the wifi chip, the bluetooth chip are not working at first, we suggest you use an ethernet cable but we will solve this problem right away
- Problem with the chips
- Check the state of the wifi chip:
sudo lshw -C networkrfkill listIf there are 2 wifi chips it means that this is the issuesudo tee /etc/modprobe.d/ideapad.conf <<< "blacklist ideapad_laptop"- Reboot
-
Update Ubuntu 18.04 (a window should suggest you to update once you are connected to the web)
-
Install some dependancies:
sudo apt get updatesudo apt get upgradesudo apt-get --assume-yes install tmux build-essential gcc g++ make binutils software-properties-common git
- Check compatibility of the graphic card:
lspci | grep -i nvidia - Some informations about the GPU should be displayed otherwise go here to solve the issue
- Install some dependancies, python3.7, pip3.7:
sudo apt-get install build-essential cmake unzip zip dkms freeglut3 freeglut3-dev libxi-dev libxmu-devsudo add-apt-repository ppa:deadsnakes/ppasudo apt-get updatesudo apt-get install python3.7 python3.7-dev pylintsudo python3.7 get-pip.pysudo pip3.7 install testresources numpy matplotlib
- Download CUDA 10.0 at this link (Linux-x86_64 -> Ubuntu-18.04 -> deb (local))
- Install CUDA 10.0: The installation of cuda needs to be done after installing the right drivers for the computer otherwise the computer will freeze and there won't be any easy way to get back from this. We found the solution here, to sum it up:
sudo add-apt-repository ppa:graphics-drivers/ppasudo apt-get updatesudo apt-get install nvidia-driver-410- Reboot the PC
Then in a terminal:
nvidia-smiSome informations about the drivers should be displayed
To finish the installation of CUDA open a terminal in the folder where you downloaded CUDA 10 at step 4
sudo dpkg -i cuda-repo-ubuntu1804-10-0-local-10.0.130-410.48_1.0-1_amd64.debsudo apt-key add /var/cuda-repo-10-0-local-10.0.130-410.48/7fa2af80.pubsudo apt-get updatesudo apt-get install cuda
Then you need to add some PATH variables:
sudo gedit ~/.bashrc- In your .bashrc add these lines:
export PATH=/usr/local/cuda-10.0/bin:/usr/local/cuda-10.0/NsightCompute-1.0${PATH:+:${PATH}}export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64/${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
save, exit and reboot
- Register at https://developer.nvidia.com/cudnn
- Download:
- cuDNN v7.6.1 Runtime Library for Ubuntu18.04 (Deb)
- cuDNN v7.6.1 Developer Library for Ubuntu18.04 (Deb)
- cuDNN v7.6.1 Code Samples and User Guide for Ubuntu18.04 (Deb)
- In a terminal where the files were downloaded:
sudo dpkg -i libcudnn7_7.*.debsudo dpkg -i libcudnn7-dev_7.*.debsudo dpkg -i libcudnn7-doc_7.*.deb
- Check the installation in a terminal:
cp -r /usr/src/cudnn_samples_v7/ $HOMEcd $HOME/cudnn_samples_v7/mnistCUDNNmake clean && make./mnistCUDNN
If the message "Test passed" is displayed then everything is ok
In a terminal:
sudo pip3.7 install tensorflow-gpu==2.0.0
You can check the installation by running this script:
import tensorflow as tf
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(input_shape=(28, 28)),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(10, activation='softmax')
])
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.fit(x_train, y_train, epochs=5)
model.evaluate(x_test, y_test)sudo pip3.7 install https://download.pytorch.org/whl/cu100/torch-1.4.0%2Bcu100-cp37-cp37m-linux_x86_64.whl torchvision
You can check the installation:
import torch
print(torch.__version__)
print(torch.cuda.is_available())sudo pip3.7 install scikit-learn networkX jupyter pandas scipy h5py seaborn
To be able to monitor the load of the GPU and CPU we found glances to be quite handy, to install it:
sudo pip3.7 install nvidia-ml-py3sudo pip3.7 install glances
You can now start glances by simply typing glances in the terminal
This guide gives a good overview of all that is necessary to build a deep learning setting from a freshly bought PC, sadly it is not general it had been tested on 2 different PCs that came from the Lenovo Legion collection.