MobileNetV4¶
MobileNetV4 backbone based on MobileNetV4 -- Universal Models for the Mobile Ecosystem.
Field list¶
| Field |
Description |
|---|---|
name |
(str) Name must be "mobilenetv4" to use MobileNetV4 backbone. |
stage_params[n] case1: Conv2D |
(list) Build Conv2D layer under following format: ['conv', out_channels, kernel_size, stride]. |
stage_params[n] case2: FusedIB |
(list) Build FusedIB block under following format: ['fi', out_channels, hidden_channels, kernel_size, stride]. |
stage_params[n] case3: UniversalInvertedResidualBlock |
(list) Build UniversalInvertedResidualBlock block under following format: ['uir', out_channels, hidden_channels, extra_dw, extra_dw_kernel_size, middle_dw, middle_dw_kernel_size, stride]. |
stage_params[n] case4: MobileMultiQueryAttention2D |
(list) Build MobileMultiQueryAttention2D block under following format: ['mmqa', out_channels, attention_channel, num_attention_heads, query_pooling_stride, key_val_downsample, key_val_downsample_kernel_size, key_val_downsample_stride, stride]. |
Model configuration examples¶
MobileNetV4-conv-small
model:
architecture:
backbone:
name: mobilenetv4
params:
stem_out_channel: 32
stem_kernel_size: 3
stem_stride: 2
final_conv_out_channel: 960
final_conv_kernel_size: 1
final_conv_stride: 1
norm_type: batch_norm
act_type: relu
return_stage_idx: ~
layer_scale: 0.1
stage_params:
# Conv2D: ['conv', out_channels, kernel_size, stride]
# FusedIB: ['fi', out_channels, hidden_channels, kernel_size, stride]
# UniversalInvertedResidualBlock: ['uir', out_channels, hidden_channels, extra_dw, extra_dw_kernel_size, middle_dw, middle_dw_kernel_size, stride]
# MobileMultiQueryAttention2D: ['mmqa', out_channels, attention_channel, num_attention_heads, query_pooling_stride, key_val_downsample, key_val_downsample_kernel_size, key_val_downsample_stride, stride]
-
- ['conv', 32, 3, 2]
- ['conv', 32, 1, 1]
-
- ['conv', 96, 3, 2]
- ['conv', 64, 1, 1]
-
- ['uir', 96, 192, True, 5, True, 5, 2]
- ['uir', 96, 192, False, ~, True, 3, 1]
- ['uir', 96, 192, False, ~, True, 3, 1]
- ['uir', 96, 192, False, ~, True, 3, 1]
- ['uir', 96, 192, False, ~, True, 3, 1]
- ['uir', 96, 384, True, 3, False, ~, 1]
-
- ['uir', 128, 576, True, 3, True, 3, 2]
- ['uir', 128, 512, True, 5, True, 5, 1]
- ['uir', 128, 512, False, ~, True, 5, 1]
- ['uir', 128, 384, False, ~, True, 5, 1]
- ['uir', 128, 512, False, ~, True, 3, 1]
- ['uir', 128, 512, False, ~, True, 3, 1]
MobileNetV4-conv-medium
model:
architecture:
backbone:
name: mobilenetv4
params:
stem_out_channel: 32
stem_kernel_size: 3
stem_stride: 2
final_conv_out_channel: 960
final_conv_kernel_size: 1
final_conv_stride: 1
norm_type: batch_norm
act_type: relu
return_stage_idx: ~
layer_scale: ~
stage_params:
# Conv2D: ['conv', out_channels, kernel_size, stride]
# FusedIB: ['fi', out_channels, hidden_channels, kernel_size, stride]
# UniversalInvertedResidualBlock: ['uir', out_channels, hidden_channels, extra_dw, extra_dw_kernel_size, middle_dw, middle_dw_kernel_size, stride]
# MobileMultiQueryAttention2D: ['mmqa', out_channels, attention_channel, num_attention_heads, query_pooling_stride, key_val_downsample, key_val_downsample_kernel_size, key_val_downsample_stride, stride]
-
- ['fi', 48, 128, 3, 2]
-
- ['uir', 80, 192, True, 3, True, 5, 2]
- ['uir', 80, 160, True, 3, True, 3, 1]
-
- ['uir', 160, 480, True, 3, True, 5, 2]
- ['uir', 160, 640, True, 3, True, 3, 1]
- ['uir', 160, 640, True, 3, True, 3, 1]
- ['uir', 160, 640, True, 3, True, 5, 1]
- ['uir', 160, 640, True, 3, True, 3, 1]
- ['uir', 160, 640, True, 3, False, ~, 1]
- ['uir', 160, 320, False, ~, False, ~, 1]
- ['uir', 160, 640, True, 3, False, ~, 1]
-
- ['uir', 256, 960, True, 5, True, 5, 2]
- ['uir', 256, 1024, True, 5, True, 5, 1]
- ['uir', 256, 1024, True, 3, True, 5, 1]
- ['uir', 256, 1024, True, 3, True, 5, 1]
- ['uir', 256, 1024, False, ~, False, ~, 1]
- ['uir', 256, 1024, True, 3, False, ~, 1]
- ['uir', 256, 512, True, 3, True, 5, 1]
- ['uir', 256, 1024, True, 5, True, 5, 1]
- ['uir', 256, 1024, False, ~, False, ~, 1]
- ['uir', 256, 1024, False, ~, False, ~, 1]
- ['uir', 256, 512, True, 5, False, ~, 1]
MobileNetV4-conv-large
model:
architecture:
backbone:
name: mobilenetv4
params:
stem_out_channel: 24
stem_kernel_size: 3
stem_stride: 2
final_conv_out_channel: 960
final_conv_kernel_size: 1
final_conv_stride: 1
norm_type: batch_norm
act_type: relu
return_stage_idx: ~
layer_scale: ~
stage_params:
# Conv2D: ['conv', out_channels, kernel_size, stride]
# FusedIB: ['fi', out_channels, hidden_channels, kernel_size, stride]
# UniversalInvertedResidualBlock: ['uir', out_channels, hidden_channels, extra_dw, extra_dw_kernel_size, middle_dw, middle_dw_kernel_size, stride]
# MobileMultiQueryAttention2D: ['mmqa', out_channels, attention_channel, num_attention_heads, query_pooling_stride, key_val_downsample, key_val_downsample_kernel_size, key_val_downsample_stride, stride]
-
- ['fi', 48, 96, 3, 2]
-
- ['uir', 96, 192, True, 3, True, 5, 2]
- ['uir', 96, 384, True, 3, True, 3, 1]
-
- ['uir', 192, 384, True, 3, True, 5, 2]
- ['uir', 192, 768, True, 3, True, 3, 1]
- ['uir', 192, 768, True, 3, True, 3, 1]
- ['uir', 192, 768, True, 3, True, 3, 1]
- ['uir', 192, 768, True, 3, True, 5, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['uir', 192, 768, True, 3, False, ~, 1]
-
- ['uir', 512, 768, True, 5, True, 5, 2]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, True, 3, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, True, 3, 1]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
MobileNetV4-hybrid-medium
model:
architecture:
backbone:
name: mobilenetv4
params:
stem_out_channel: 32
stem_kernel_size: 3
stem_stride: 2
final_conv_out_channel: 960
final_conv_kernel_size: 1
final_conv_stride: 1
norm_type: batch_norm
act_type: relu
return_stage_idx: ~
layer_scale: 0.1
stage_params:
# Conv2D: ['conv', out_channels, kernel_size, stride]
# FusedIB: ['fi', out_channels, hidden_channels, kernel_size, stride]
# UniversalInvertedResidualBlock: ['uir', out_channels, hidden_channels, extra_dw, extra_dw_kernel_size, middle_dw, middle_dw_kernel_size, stride]
# MobileMultiQueryAttention2D: ['mmqa', out_channels, attention_channel, num_attention_heads, query_pooling_stride, key_val_downsample, key_val_downsample_kernel_size, key_val_downsample_stride, stride]
-
- ['fi', 48, 128, 3, 2]
-
- ['uir', 80, 192, True, 3, True, 5, 2]
- ['uir', 80, 160, True, 3, True, 3, 1]
-
- ['uir', 160, 480, True, 3, True, 5, 2]
- ['uir', 160, 320, False, ~, False, ~, 1]
- ['uir', 160, 640, True, 3, True, 3, 1]
- ['uir', 160, 640, True, 3, True, 5, 1]
- ['mmqa', 160, 256, 4, ~, True, 3, 2, 1]
- ['uir', 160, 640, True, 3, True, 3, 1]
- ['mmqa', 160, 256, 4, ~, True, 3, 2, 1]
- ['uir', 160, 640, True, 3, False, ~, 1]
- ['mmqa', 160, 256, 4, ~, True, 3, 2, 1]
- ['uir', 160, 640, True, 3, True, 3, 1]
- ['mmqa', 160, 256, 4, ~, True, 3, 2, 1]
- ['uir', 160, 640, True, 3, False, ~, 1]
-
- ['uir', 256, 960, True, 5, True, 5, 2]
- ['uir', 256, 1024, True, 5, True, 5, 1]
- ['uir', 256, 1024, True, 3, True, 5, 1]
- ['uir', 256, 1024, True, 3, True, 5, 1]
- ['uir', 256, 512, False, ~, False, ~, 1]
- ['uir', 256, 512, True, 3, True, 5, 1]
- ['uir', 256, 512, False, ~, False, ~, 1]
- ['uir', 256, 1024, False, ~, False, ~, 1]
- ['mmqa', 256, 256, 4, ~, False, ~, ~, 1]
- ['uir', 256, 1024, True, 3, False, ~, 1]
- ['mmqa', 256, 256, 4, ~, False, ~, ~, 1]
- ['uir', 256, 1024, True, 5, True, 5, 1]
- ['mmqa', 256, 256, 4, ~, False, ~, ~, 1]
- ['uir', 256, 1024, True, 5, False, ~, 1]
- ['mmqa', 256, 256, 4, ~, False, ~, ~, 1]
- ['uir', 256, 1024, True, 5, False, ~, 1]
MobileNetV4-hybrid-large
model:
architecture:
backbone:
name: mobilenetv4
params:
stem_out_channel: 24
stem_kernel_size: 3
stem_stride: 2
final_conv_out_channel: 960
final_conv_kernel_size: 1
final_conv_stride: 1
norm_type: batch_norm
act_type: gelu
return_stage_idx: ~
layer_scale: 0.1
stage_params:
# Conv2D: ['conv', out_channels, kernel_size, stride]
# FusedIB: ['fi', out_channels, hidden_channels, kernel_size, stride]
# UniversalInvertedResidualBlock: ['uir', out_channels, hidden_channels, extra_dw, extra_dw_kernel_size, middle_dw, middle_dw_kernel_size, stride]
# MobileMultiQueryAttention2D: ['mmqa', out_channels, attention_channel, num_attention_heads, query_pooling_stride, key_val_downsample, key_val_downsample_kernel_size, key_val_downsample_stride, stride]
-
- ['fi', 48, 96, 3, 2]
-
- ['uir', 96, 192, True, 3, True, 5, 2]
- ['uir', 96, 384, True, 3, True, 3, 1]
-
- ['uir', 192, 384, True, 3, True, 5, 2]
- ['uir', 192, 768, True, 3, True, 3, 1]
- ['uir', 192, 768, True, 3, True, 3, 1]
- ['uir', 192, 768, True, 3, True, 3, 1]
- ['uir', 192, 768, True, 3, True, 5, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['mmqa', 192, 384, 8, ~, True, 3, 2, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['mmqa', 192, 384, 8, ~, True, 3, 2, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['mmqa', 192, 384, 8, ~, True, 3, 2, 1]
- ['uir', 192, 768, True, 5, True, 3, 1]
- ['mmqa', 192, 384, 8, ~, True, 3, 2, 1]
- ['uir', 192, 768, True, 3, False, ~, 1]
-
- ['uir', 512, 768, True, 5, True, 5, 2]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, True, 3, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['uir', 512, 2048, True, 5, True, 3, 1]
- ['uir', 512, 2048, True, 5, True, 5, 1]
- ['mmqa', 512, 512, 8, ~, False, ~, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['mmqa', 512, 512, 8, ~, False, ~, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['mmqa', 512, 512, 8, ~, False, ~, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]
- ['mmqa', 512, 512, 8, ~, False, ~, ~, 1]
- ['uir', 512, 2048, True, 5, False, ~, 1]