Training the Network in WSL

Back to Index

Lastly, we will train the network. Since we are in WSL, we don't have a Display Driver. Nvidia itself does not recomend to do so. So, we need to erase/comment some lines in the network code.

Erase/comment any reference to the class GPUAccounting (declared in "runstats.py") made in "utils/deeplearningutilities/torch/trainer.py", and the "IterationTimer()" for the display in the same file.

In a nutshell, erase/comment:

• self._display_iteration_timer = IterationTimer()
• self._gpu_accounting = GPUAccounting()
• The "if" right below "# gpu stats"
• The "if" right below "# print display strings"

It is recomended to install PRCTL for Python so you can load the files in parallel, and get rid of a warning. Therefore:

• sudo apt-get install python3.7-dev libcap-dev python3-setuptools

Then, follow the instructions of the official site:

• git clone http://github.com/seveas/python-prctl
• cd python-prctl
• python setup.py build
• sudo python setup.py install

Look for the "prctl" file in the active miniconda environment, and test the command:

• python3.7 -c "import sys; sys.path.append('/usr/local/lib/python3.7/dist-packages/python_prctl-1.8.1-py3.7-linux-x86_64.egg/'); import prctl; print(prctl); print(sys.version_info)"

Now, write this in the train_network_torch.py:

• sys.path.append('/usr/local/lib/python3.7/dist-packages/python_prctl-1.8.1-py3.7-linux-x86_64.egg/')

Note: Remember to switch the path in blue for the one you found.

Start the training:

• python3 train_network_torch.py default.yalm

Evaluate:

• python3 evaluate_network.py --trainscript train_network_torch.py --cfg default.yaml --weights pretrained_model_weights.pt

The End.

<- Back