Where is the MobileNetV4?

[csw92805]

I only find MobileNetV1,V2,V3 and V5 in the timm models. So where is the MobileNetV4? And what model is MobileNetV5? I noticed that Google seems to have only released up to V4.

[brianhou0208]

MobileNetV4 variants are implemented within the mobilenetv3.py file.

pytorch-image-models/timm/models/mobilenetv3.py

Lines 783 to 1036 in ae4d1bb

	def _gen_mobilenet_v4(
	variant: str,
	channel_multiplier: float = 1.0,
	group_size: Optional[int] = None,
	pretrained: bool = False,
	**kwargs,
	) -> MobileNetV3:
	"""Creates a MobileNet-V4 model.
	Paper: https://arxiv.org/abs/2404.10518
	Args:
	variant: Model variant name.
	channel_multiplier: Multiplier to number of channels per layer.
	group_size: Group size for grouped convolutions.
	pretrained: Load pretrained weights.
	**kwargs: Additional model arguments.
	Returns:
	MobileNetV3 model instance.
	"""
	num_features = 1280
	if 'hybrid' in variant:
	layer_scale_init_value = 1e-5
	if 'medium' in variant:
	stem_size = 32
	act_layer = resolve_act_layer(kwargs, 'relu')
	arch_def = [
	# stage 0, 112x112 in
	[
	'er_r1_k3_s2_e4_c48' # FusedIB (EdgeResidual)
	],
	# stage 1, 56x56 in
	[
	'uir_r1_a3_k5_s2_e4_c80', # ExtraDW
	'uir_r1_a3_k3_s1_e2_c80', # ExtraDW
	],
	# stage 2, 28x28 in
	[
	'uir_r1_a3_k5_s2_e6_c160', # ExtraDW
	'uir_r1_a0_k0_s1_e2_c160', # FFN
	'uir_r1_a3_k3_s1_e4_c160', # ExtraDW
	'uir_r1_a3_k5_s1_e4_c160', # ExtraDW
	'mqa_r1_k3_h4_s1_v2_d64_c160', # MQA w/ KV downsample
	'uir_r1_a3_k3_s1_e4_c160', # ExtraDW
	'mqa_r1_k3_h4_s1_v2_d64_c160', # MQA w/ KV downsample
	'uir_r1_a3_k0_s1_e4_c160', # ConvNeXt
	'mqa_r1_k3_h4_s1_v2_d64_c160', # MQA w/ KV downsample
	'uir_r1_a3_k3_s1_e4_c160', # ExtraDW
	'mqa_r1_k3_h4_s1_v2_d64_c160', # MQA w/ KV downsample
	'uir_r1_a3_k0_s1_e4_c160', # ConvNeXt
	],
	# stage 3, 14x14in
	[
	'uir_r1_a5_k5_s2_e6_c256', # ExtraDW
	'uir_r1_a5_k5_s1_e4_c256', # ExtraDW
	'uir_r2_a3_k5_s1_e4_c256', # ExtraDW
	'uir_r1_a0_k0_s1_e2_c256', # FFN
	'uir_r1_a3_k5_s1_e2_c256', # ExtraDW
	'uir_r1_a0_k0_s1_e2_c256', # FFN
	'uir_r1_a0_k0_s1_e4_c256', # FFN
	'mqa_r1_k3_h4_s1_d64_c256', # MQA
	'uir_r1_a3_k0_s1_e4_c256', # ConvNeXt
	'mqa_r1_k3_h4_s1_d64_c256', # MQA
	'uir_r1_a5_k5_s1_e4_c256', # ExtraDW
	'mqa_r1_k3_h4_s1_d64_c256', # MQA
	'uir_r1_a5_k0_s1_e4_c256', # ConvNeXt
	'mqa_r1_k3_h4_s1_d64_c256', # MQA
	'uir_r1_a5_k0_s1_e4_c256', # ConvNeXt
	],
	# stage 4, 7x7 in
	[
	'cn_r1_k1_s1_c960' # Conv
	],
	]
	elif 'large' in variant:
	stem_size = 24
	act_layer = resolve_act_layer(kwargs, 'gelu')
	arch_def = [
	# stage 0, 112x112 in
	[
	'er_r1_k3_s2_e4_c48', # FusedIB (EdgeResidual)
	],
	# stage 1, 56x56 in
	[
	'uir_r1_a3_k5_s2_e4_c96', # ExtraDW
	'uir_r1_a3_k3_s1_e4_c96', # ExtraDW
	],
	# stage 2, 28x28 in
	[
	'uir_r1_a3_k5_s2_e4_c192', # ExtraDW
	'uir_r3_a3_k3_s1_e4_c192', # ExtraDW
	'uir_r1_a3_k5_s1_e4_c192', # ExtraDW
	'uir_r2_a5_k3_s1_e4_c192', # ExtraDW
	'mqa_r1_k3_h8_s1_v2_d48_c192', # MQA w/ KV downsample
	'uir_r1_a5_k3_s1_e4_c192', # ExtraDW
	'mqa_r1_k3_h8_s1_v2_d48_c192', # MQA w/ KV downsample
	'uir_r1_a5_k3_s1_e4_c192', # ExtraDW
	'mqa_r1_k3_h8_s1_v2_d48_c192', # MQA w/ KV downsample
	'uir_r1_a5_k3_s1_e4_c192', # ExtraDW
	'mqa_r1_k3_h8_s1_v2_d48_c192', # MQA w/ KV downsample
	'uir_r1_a3_k0_s1_e4_c192', # ConvNeXt
	],
	# stage 3, 14x14in
	[
	'uir_r4_a5_k5_s2_e4_c512', # ExtraDW
	'uir_r1_a5_k0_s1_e4_c512', # ConvNeXt
	'uir_r1_a5_k3_s1_e4_c512', # ExtraDW
	'uir_r2_a5_k0_s1_e4_c512', # ConvNeXt
	'uir_r1_a5_k3_s1_e4_c512', # ExtraDW
	'uir_r1_a5_k5_s1_e4_c512', # ExtraDW
	'mqa_r1_k3_h8_s1_d64_c512', # MQA
	'uir_r1_a5_k0_s1_e4_c512', # ConvNeXt
	'mqa_r1_k3_h8_s1_d64_c512', # MQA
	'uir_r1_a5_k0_s1_e4_c512', # ConvNeXt
	'mqa_r1_k3_h8_s1_d64_c512', # MQA
	'uir_r1_a5_k0_s1_e4_c512', # ConvNeXt
	'mqa_r1_k3_h8_s1_d64_c512', # MQA
	'uir_r1_a5_k0_s1_e4_c512', # ConvNeXt
	],
	# stage 4, 7x7 in
	[
	'cn_r1_k1_s1_c960', # Conv
	],
	]
	else:
	assert False, f'Unknown variant {variant}.'
	else:
	layer_scale_init_value = None
	if 'small' in variant:
	stem_size = 32
	act_layer = resolve_act_layer(kwargs, 'relu')
	arch_def = [
	# stage 0, 112x112 in
	[
	'cn_r1_k3_s2_e1_c32', # Conv
	'cn_r1_k1_s1_e1_c32', # Conv
	],
	# stage 1, 56x56 in
	[
	'cn_r1_k3_s2_e1_c96', # Conv
	'cn_r1_k1_s1_e1_c64', # Conv
	],
	# stage 2, 28x28 in
	[
	'uir_r1_a5_k5_s2_e3_c96', # ExtraDW
	'uir_r4_a0_k3_s1_e2_c96', # IR
	'uir_r1_a3_k0_s1_e4_c96', # ConvNeXt
	],
	# stage 3, 14x14 in
	[
	'uir_r1_a3_k3_s2_e6_c128', # ExtraDW
	'uir_r1_a5_k5_s1_e4_c128', # ExtraDW
	'uir_r1_a0_k5_s1_e4_c128', # IR
	'uir_r1_a0_k5_s1_e3_c128', # IR
	'uir_r2_a0_k3_s1_e4_c128', # IR
	],
	# stage 4, 7x7 in
	[
	'cn_r1_k1_s1_c960', # Conv
	],
	]
	elif 'medium' in variant:
	stem_size = 32
	act_layer = resolve_act_layer(kwargs, 'relu')
	arch_def = [
	# stage 0, 112x112 in
	[
	'er_r1_k3_s2_e4_c48', # FusedIB (EdgeResidual)
	],
	# stage 1, 56x56 in
	[
	'uir_r1_a3_k5_s2_e4_c80', # ExtraDW
	'uir_r1_a3_k3_s1_e2_c80', # ExtraDW
	],
	# stage 2, 28x28 in
	[
	'uir_r1_a3_k5_s2_e6_c160', # ExtraDW
	'uir_r2_a3_k3_s1_e4_c160', # ExtraDW
	'uir_r1_a3_k5_s1_e4_c160', # ExtraDW
	'uir_r1_a3_k3_s1_e4_c160', # ExtraDW
	'uir_r1_a3_k0_s1_e4_c160', # ConvNeXt
	'uir_r1_a0_k0_s1_e2_c160', # ExtraDW
	'uir_r1_a3_k0_s1_e4_c160', # ConvNeXt
	],
	# stage 3, 14x14in
	[
	'uir_r1_a5_k5_s2_e6_c256', # ExtraDW
	'uir_r1_a5_k5_s1_e4_c256', # ExtraDW
	'uir_r2_a3_k5_s1_e4_c256', # ExtraDW
	'uir_r1_a0_k0_s1_e4_c256', # FFN
	'uir_r1_a3_k0_s1_e4_c256', # ConvNeXt
	'uir_r1_a3_k5_s1_e2_c256', # ExtraDW
	'uir_r1_a5_k5_s1_e4_c256', # ExtraDW
	'uir_r2_a0_k0_s1_e4_c256', # FFN
	'uir_r1_a5_k0_s1_e2_c256', # ConvNeXt
	],
	# stage 4, 7x7 in
	[
	'cn_r1_k1_s1_c960', # Conv
	],
	]
	elif 'large' in variant:
	stem_size = 24
	act_layer = resolve_act_layer(kwargs, 'relu')
	arch_def = [
	# stage 0, 112x112 in
	[
	'er_r1_k3_s2_e4_c48', # FusedIB (EdgeResidual)
	],
	# stage 1, 56x56 in
	[
	'uir_r1_a3_k5_s2_e4_c96', # ExtraDW
	'uir_r1_a3_k3_s1_e4_c96', # ExtraDW
	],
	# stage 2, 28x28 in
	[
	'uir_r1_a3_k5_s2_e4_c192', # ExtraDW
	'uir_r3_a3_k3_s1_e4_c192', # ExtraDW
	'uir_r1_a3_k5_s1_e4_c192', # ExtraDW
	'uir_r5_a5_k3_s1_e4_c192', # ExtraDW
	'uir_r1_a3_k0_s1_e4_c192', # ConvNeXt
	],
	# stage 3, 14x14in
	[
	'uir_r4_a5_k5_s2_e4_c512', # ExtraDW
	'uir_r1_a5_k0_s1_e4_c512', # ConvNeXt
	'uir_r1_a5_k3_s1_e4_c512', # ExtraDW
	'uir_r2_a5_k0_s1_e4_c512', # ConvNeXt
	'uir_r1_a5_k3_s1_e4_c512', # ExtraDW
	'uir_r1_a5_k5_s1_e4_c512', # ExtraDW
	'uir_r3_a5_k0_s1_e4_c512', # ConvNeXt
	],
	# stage 4, 7x7 in
	[
	'cn_r1_k1_s1_c960', # Conv
	],
	]
	else:
	assert False, f'Unknown variant {variant}.'
	model_kwargs = dict(
	block_args=decode_arch_def(arch_def, group_size=group_size),
	head_bias=False,
	head_norm=True,
	num_features=num_features,
	stem_size=stem_size,
	fix_stem=channel_multiplier < 1.0,
	round_chs_fn=partial(round_channels, multiplier=channel_multiplier),
	norm_layer=partial(nn.BatchNorm2d, **resolve_bn_args(kwargs)),
	act_layer=act_layer,
	layer_scale_init_value=layer_scale_init_value,
	**kwargs,

MobileNetV5 is Vision Encoder from Gemma3n:
https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/#mobilenet-v5:-new-state-of-the-art-vision-encoder

[rwightman]

Brian's answer is correct, mnv4 has the same higher level structure as mnv3 so it's implemented there with two new blocks added.

mnv5 was for Gemma 3n encoder, it could be trained stand alone too but for most 'mobile' applications not clear that's better than mnv4 hybrid. For scaling up maybe? mnv5 uses the same blocks as mnv4 but it has a different head setup w/ multi-scale fusion. So implemented as its own file for clear separation for the gemma integration.

mnv1 and mnv2 have a structure that's simila to efficientnet, which is why they are implemented there.