Back to Lab Notes
Lab Notes / Custom MLP and PyTorch on Fashion-MNIST

Learning rate changes the outcome

Across three recorded learning rates, both implementations changed markedly; at 0.0001 the saved test accuracies were 38.47% and 66.04%.

Practical consequence

Inspect training configuration before attributing a gap to architecture or library choice.

01

Question

How do the recorded test accuracies change as the learning rate changes for a custom MLP and a PyTorch implementation?

02

Experiment

Saved Fashion-MNIST runs compare a custom MLP with a PyTorch implementation at three learning rates. The values below are the available test outputs.

03

Recorded results

Recorded test results for these particular runs; no multi-seed protocol was documented.
Learning ratePyTorch test accuracyCustom MLP test accuracy
0.000138.47%66.04%
0.00275.81%81.35%
0.186.87%87.67%

Limits

The runs do not document repeated seeds or establish a general advantage for the custom model over PyTorch. No training curves, timing measurements, or confidence intervals were supplied.

04

Interpretation

In these runs, the learning rate substantially changed the outcome. Configuration is a plausible part of the observed gap.

Practical consequence

Inspect training configuration before attributing a gap to architecture or library choice.

05

Next experiment

Compare matched architectures and training settings across several documented seeds, then report the distribution of test outcomes.

06

Origin & contribution

Educational Fashion-MNIST comparison. Values come from an editorial synthesis of saved notebook outputs supplied for this update; the notebooks are not in this checkout.

The original implementation and precise personal contribution still need source review. The comparison here is a presentation of saved outputs, not a new experiment.

ProjectCustom MLP and PyTorch on Fashion-MNIST

A public project page is not available yet.