A very simple, three layered neural network, as excercise while reading Grokking Deep Learning by Andrew W. Trask. Goal is to evaluate it on MNIST, and to add features as I progress in the book.
| model_id | alpha | epochs | hidden_layer_size | dropout | error | accuracy |
|---|---|---|---|---|---|---|
| 0cd32548-f205-11eb-a7dc-7c04d0d54c7a | 0.005 | 20 | 264 | no | 0.18 | 0.945 |
| 524cdfd4-f227-11eb-ae71-7c04d0d54c7a | 0.05 | 350 | 264 | no | 0.09 | 0.974 |
| 14d06eb2-f300-11eb-86a7-7c04d0d54c7a | 0.005 | 20 | 264 | yes | 0.095 | 0.972 |
I'm a bit suspicious of the reported accuracy, since it is considerable higher than in the book (where 0.81 test accuracy is reported, with a model with hidden_size=100, dropout, alpha 0.005 and 300 iterations). I'll have to make some more runs to find out why.
Run python benchmark.py [model_id] to re-run a benchmark of a certain model, or python benchmark.py for a run with the current configurations in benchmark.py.