This repo contains two implementations of MobileNet V3: one for PyTorch and another for Keras (TF 2.0).
There's also a PyTroch checkpoint for "small" MobileNet V3 that achieves 66.718% top1 and 86.975% top5 accuracies on ImageNet. Training was done with SGD, lr 0.05, wd 1e-5, 150 epochs for initial training and then 2 50 epoch runs with lower peak LR. Cosine schedule with warmup.
Because TF is much slower than PyTorch, training for the Keras model is still underway, but the model is straight-across port, so it should, hypothetically, achieve the same accuracy results.
Contributions are welcome. Please use Black to format the code.