This repo contains two implementations of MobileNet V3: one for PyTorch and another for Keras (TF 2.0), including segmentation-specific variants. I have used these for practical applications, they seem to work fine.

MobileNet V3 blocks in this implementation also retain the feature map if there is a downsampling in the block so that the feature map can then be fed into a detection or segmentation head. Keras version doesn't need this because there you can get the relevant op outputs directly, but for PyTorch the way the paper describes head attachment requires that you grab the output from inside the block, so this helps you do it easily.

There are also PyTorch checkpoints for "small" and "large" MobileNet V3 that achieve 67.36% (which I suppose you could interpret as paper's 67.4%) and 74.78% top1 correspondingly. I was not able to achieve "paper" accuracy for the "large" model, even though I'm pretty sure the code is correct.

Also available are the following checkpoints:

• Large 1.0 Detection: top1=73.38%

Keras checkpoints are TBD.

Training setup was as follows:

"large": {
		"1.0": {
				"alpha": 1.0,
				"dropout": 0.2,
				"batch_size": 512,
				"loss": "smooth_ce",
				"loss_kwargs": {"smoothing": 0.1},
				"epochs": 200,
				"optimizer": "sgd",
				"optimizer_kwargs": {
						"momentum": 0.9,
						"weight_decay": 2e-5,
						"nesterov": True
				},
				"scheduler": "cosine",
				"scheduler_kwargs": {
						"num_cycles": 1,
						"peak_lr": 0.8,
						"min_lr": 1e-7,
						"initial_warmup_step_fraction": 0.0,
						"cycle_warmup_step_fraction": 0.1,
				},
				"trainer_kwargs": {"use_ema": True},
		}
},
"small": {
		"1.0": {
				"alpha": 1.0,
				"dropout": 0.2,
				"batch_size": 1024,
				"epochs": 200,
				"loss": "smooth_ce",
				"loss_kwargs": {"smoothing": 0.1},
				"optimizer": "sgd",
				"optimizer_kwargs": {
						"momentum": 0.9,
						"weight_decay": 2e-5,
						"nesterov": True
				},
				"scheduler": "cosine",
				"scheduler_kwargs": {
						"num_cycles": 1,
						"peak_lr": 0.8,
						"min_lr": 1e-7,
						"initial_warmup_step_fraction": 0.0,
						"cycle_warmup_step_fraction": 0.1,
				},
				"trainer_kwargs": {"use_ema": True},
		}
},

Normal augmentation was changed to crop less aggressively with scale setting within [0.2, 1.0] for faster convergence. For the large model the checkpoint is EMA smoothed with decay of 0.9999. Small model did not ultimately benefit from EMA.

PyTorch implementation requires PyTorch 1.3.0+, Keras requires TF 2.0.0+.

Contributions are welcome. Please use Black to format the code.