In step 6 I said dropout + weight decay "do not raise validation accuracy in a way you would notice". Do not believe it — check it.

Write a new file seeds.py that, for two settings — baseline (dropout=0, weight_decay=0) and dropout=0.3, weight_decay=1e-4 — and three seeds (0, 1, 2), trains an MLP (256, 256): 10 000 training images, 5 000 validation, 30 epochs, Adam lr=0.001, no early stopping. For each setting print the training accuracies, the validation accuracies and the validation mean.

Details that matter:

• torch.manual_seed(seed) before each training, and build the model after it;
• measure the accuracies with evaluate after fit;
• fit prints a line per epoch — silence it (verbose=False).

My machine (it takes ~6 min — 6 trainings of 30 epochs, and that is normal):

baseline                 train 95.8% 97.3% 97.0%  val 85.8% 87.3% 86.9%  mean val 86.7%
dropout 0.3 + wd 1e-4    train 93.1% 92.7% 94.0%  val 86.7% 86.2% 87.7%  mean val 86.9%

Answer three questions in writing, one sentence each:

1. Is training accuracy systematically lower with regularisation? (Is that visible in all three seeds?)
2. Is the difference of the validation means larger than the difference between seeds within the same row?
3. What number would you tell a friend if they asked "how accurate is your model"?

The third question is the most important. My answer would be: "about 86–87 % — and I would not try to be more precise, because runs differ by up to 1.5 points".

Your numbers may differ by a few tenths — what matters is that you answer 1 and 2 from what you see, not from my table.