Today I was trying to train the complete CIFAR-10 dataset on Xavier. The CIFAR-10 dataset is 60,000 32x32 RGB images. After installing Tensorflow I decided to give it a go. My 2.1 GB of memory usage soared to over the 15.4 GB limit during training. Then my Xavier froze. This is no good. The short term solution was to create a 8GB Swap file.
-
-$ df -h
This will show your file system and how much space you have left. -
-$ sudo fallocate -l 8.0G /swapfile
Create a 8 GB of swap -
-$ sudo chmod 600 /swapfile
Change file permissions -
-$ sudo mkswap /swapfile
-
-$ sudo swapon /swapfile
-
-$ free -m
This will show you the swap file is on. You can also pull up the System Monitor
However this is only temporary. If you reboot, swap file is gone. -
-$ sudo nano /etc/fstab
-
Within this file add the line “/swapfile none swap 0 0”. Do not include the quotes.
Exit and save file -
Now you can reboot and your Swap will be activated.
-
Then you can successfully train the CIFAR-10 dataset.
Ideally you would want to create a swap partition on an attached SSD or M.2 SSD and not use any of the internal memory.
