Hello all! I have been playing with swift a little bit and put together a basic imagenette demo you can look at here:
You will need to be using a nightly/tensorflow master if you want to try and run it yourself. If anybody has suggestions for refactoring/improvements, let me know!
Trains pretty consistently to 80+% using vanilla SGD:
[Epoch 67] Accuracy: 406/500 (0.812) Loss: 0.639303
[Epoch 68] Accuracy: 410/500 (0.82) Loss: 0.65327513
[Epoch 69] Accuracy: 403/500 (0.806) Loss: 0.65153205
[Epoch 70] Accuracy: 403/500 (0.806) Loss: 0.6486639
[Epoch 71] Accuracy: 420/500 (0.84) Loss: 0.60659444
[Epoch 72] Accuracy: 423/500 (0.846) Loss: 0.5867116
[Epoch 73] Accuracy: 416/500 (0.832) Loss: 0.5800408
[Epoch 74] Accuracy: 424/500 (0.848) Loss: 0.5674498
[Epoch 75] Accuracy: 422/500 (0.844) Loss: 0.55907977
[Epoch 76] Accuracy: 421/500 (0.842) Loss: 0.5679178
[Epoch 77] Accuracy: 415/500 (0.83) Loss: 0.5917381
[Epoch 78] Accuracy: 420/500 (0.84) Loss: 0.5697638
[Epoch 79] Accuracy: 421/500 (0.842) Loss: 0.5626675
[Epoch 80] Accuracy: 417/500 (0.834) Loss: 0.5851209