Accuracy for Testing and Training

[mlviz-dev]

Accuracy for Testing and Training

[mnaumovfb]

Let me answer your questions one-by-one:

1. The command you have provided will perform training. However, during the training you can periodically check the accuracy of your model on the test set. How often you check this accuracy is governed by the --test-freq argument. That's why you periodically see test prints during training. This allows you to generate the plots above.
2. It's not clear to me why exactly is your training accuracy showing up as 0.0%. I can not reproduce this on my machine. Try removing some of the arguments from your command line, such as "--plot-compute-graph $dlrm_extra_option", "--save-model=model.pt" and "--use-gpu" to see if this error persists.
3. The data is split into 7 days, the first six are used for training, while the last is used for testing. You can find more details in the data_utils.py and dlrm_data_*.py files about it.
4. I think that accuracy number being 0.0% is related to your earlier problem (see question 2).
5. This has been answered (see question 1).

[mnaumovfb]

I will close this issue. Please reopen if you are still running into trouble.