Leaderboard for the settings i used
128 5 84.6 link --epochs 5 --bs 64 --lr 3e-3 --mixup 0
I have used only one GPU
As per the below link i setup the env for 128 pixel size , 64 bs 5 epochs.
https://docs.fast.ai/dev/test.html#getting-reproducible-results
Baseline results under reproducible settings
| epoch | train_loss | valid_loss | accuracy | top_k_accuracy | time |
|---|---|---|---|---|---|
| 0 | 1.623444 | 1.875800 | 0.480000 | 0.898000 | 01:58 |
| 1 | 1.412858 | 1.535896 | 0.618000 | 0.934000 | 01:48 |
| 2 | 1.246371 | 1.275325 | 0.726000 | 0.956000 | 01:47 |
| 3 | 1.083522 | 0.980761 | 0.816000 | 0.986000 | 01:47 |
| 4 | 1.011306 | 0.938234 | 0.842000 | 0.990000 | 01:47 |
Results of my technique
| epoch | train_loss | valid_loss | accuracy | top_k_accuracy | time |
|---|---|---|---|---|---|
| 0 | 1.757266 | 1.883168 | 0.488000 | 0.880000 | 02:26 |
| 1 | 1.376854 | 1.267548 | 0.706000 | 0.964000 | 02:15 |
| 2 | 1.192331 | 1.129854 | 0.772000 | 0.966000 | 02:15 |
| 3 | 1.044082 | 0.962126 | 0.826000 | 0.986000 | 02:15 |
| 4 | 0.973471 | 0.918321 | 0.858000 | 0.984000 | 02:15 |
As i used the reproducible settings, the results are completely reproducible without any change.
Is this valid enough to consider or should i be doing any further tests
Kindly provide ur suggestion so i can carry out further testing