Create notebook that reproduces *one* point in Fig 2a of the Zeus paper

[ffund]

As a starting point for #4 let's write a notebook that reproduces one point in that figure.

[Deepak-Work]

The DeepSpeech experiment became too complicated. The librispeech data needs to be preprocessed from .flac file to .csv, which in itself becomes a very complicated task.

[Deepak-Work]

Switched to Imagenet experiment using Resnet-50. The notebook uploaded currently has not been run in its entirety. Also, missing an explanation of steps

[ffund]

Comments -

• take notes about how long each part takes, to put in the instructions
• write about how/why of all choices. e.g. refer to the specific section of Zeus repo where you found

  normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

• perhaps we are only supposed to be limited by min accuracy
• change to validation accuracy (min accuracy)
• change order in train loop to match example in this notebook

[ffund]

• Use pynvml to set power limits
• power limit values will be 100-250W, batch size will be powers of 2 from 8 to 1024
• following their comment, not specifying initial LR for Adadelta
• in your loop:
  • iterate over n_max_iter (clarify how long it will take to generate each point)
  • in each iteration, randomly sample from the ranges of power limit/batch size (exclude values that have already been selected)
  • write a json file with experiment details for each point (time and energy measurements, experiment parameters - power limit and batch size, also training and validation accuracy at the end, number of epochs)

[Deepak-Work]

The imagenet data is too huge for the node. When extracting the downloaded .tar files, the system gives a "Disk Full Error"

Compresses .tar files have the size: 137GB (Train) and 7GB (Test). These when uncompressed give error.

Continuing the experiment with a subset of data present here

[ffund]

The imagenet data is too huge for the node. When extracting the downloaded .tar files, the system gives a "Disk Full Error"

Compresses .tar files have the size: 137GB (Train) and 7GB (Test). These when uncompressed give error.

Continuing the experiment with a subset of data present here

Refer to #6