Wildly different training results on Jupyter Notebook Locally (Mac) vs Kaggle

fastai
1

Beginner here - I am running example code from the chapter 1 intro, and I am a bit confused why running the code below on my Mac (with Jupyter or VSCode kernel) provides an accuracy of 0.00000 and overall different train results than running the exact same code on kaggle. Here’s the code and screenshots of differences:

from fastai.tabular.all import *
path = untar_data(URLs.ADULT_SAMPLE)

dls = TabularDataLoaders.from_csv(path/'adult.csv', path=path, y_names="salary",
    cat_names = ['workclass', 'education', 'marital-status', 'occupation',
                 'relationship', 'race'],
    cont_names = ['age', 'fnlwgt', 'education-num'],
    procs = [Categorify, FillMissing, Normalize])

learn = tabular_learner(dls, metrics=accuracy)

learn.fit_one_cycle(3)

Local Jupyter

image
image1554×752 77.6 KB

Kaggle

image
image791×555 43.9 KB

As you can see the accuracy of the training is 0.00 for all epochs locally.

What am I missing?

2

Did you ever figure it out?

3

Nope, still looking for an answer!

4

I figured the solution to be setting the device to cpu on the dls before training. This did the job.
dls.device = torch.device('cpu')