-
for some reason, my watt-integration is considerably smaller than what pinpoint accumulates automatically
-
for some very strange reason, keras watts look completely different at 80%, apparently because is the only training data amount with a perfect batch size (48000/128=375), all others dont divide evenly
-
measure sleep/idle consumption
-
check pytorch model and keras model
- maybe check for the random seed stuff
-
test different pinpoint execution styles (and maybe validate values against nvidia-smi)
-
check energy calculation (sampling interval, correct integration, subtract idle power)
-
write script to run all sub-experiments directly consecutively
-
conda install -n
-
conda remove -n
-
conda env list
-
conda env remove -n
-
conda create -n python=3.9
-
for tensorflow: https://www.tensorflow.org/install/pip working, but replace 11.2 with 11.7
-
for pytorch:
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
- nvidia-smi -i 0 -q -d POWER
- make log files that are opened inside python scripts 666 (r+w for everyone)
- two consecutive GPU watt values are always exactly the same, probably can not be sampled that often
-
conda activate tf; python3 tensorflow_mnist.py
-
conda activate pytorch; python3 pytorch_mnist.py
-
currently paths not working for accuracy logs yet