Record the energy consumption of your cpu and gpu. Check our documentation for usage.
This repo is largely inspired from this experiment Tracker
Running Average Power Limit (RAPL) and its linux interface : powercap
RAPL is introduced in the Intel processors starting with the Sandy bridge architecture in 2011.
Your linux os supports RAPL if the following folder is not empty:
/sys/class/powercap/intel-rapl/
Empty folder? If your cpu is very recent, it is worth to check the most recent linux kernels.
Install pytorch, then,
pip install -r requirements.txt
python setup.py install
You need to authorize the reading of the rapl related files:
sudo chmod -R 755 /sys/class/powercap/intel-rapl/
See examples/example_exp_deep_learning.py.
Essentially, you instantiate an experiment and place the code you want to measure between a start and stop signal.
from deep_learning_power_measure.power_measure import experiment, parsers
driver = parsers.JsonParser("output_folder")
exp = experiment.Experiment(driver)
p, q = exp.measure_yourself(period=2)
###################
# place here the code that you want to profile
################
q.put(experiment.STOP_MESSAGE)
This will save the recordings as json file in the output_folder. You can display them with:
from deep_learning_power_measure.power_measure import experiment, parsers
driver = parsers.JsonParser(output_folder)
exp_result = experiment.ExpResults(driver)
exp_result.print()
We use a wrapper for torchinfo to extract statistics about your model. To obtain them, add additional parameters:
net = ... the model you are using for your experiment
input_size = ... (batch_size, *data_point_shape)
exp = experiment.Experiment(driver, model=net, input_size=input_size)