MixNet¶
MixNet backbone based on MixConv: Mixed Depthwise Convolutional Kernels.
Similar with MobileNetV3, we provide a configuration that allows users to define each MixDepthBlock in MixNet individually. You can specify in a list format the number and form of each MixDepthBlock for every stage. Using this, we pre-construct and provide MixNet family models.
Compatibility matrix¶
| Supporting necks | Supporting heads | torch.fx | NetsPresso |
|---|---|---|---|
|
FPN YOLOPAFPN |
FC ALLMLPDecoder AnchorDecoupledHead AnchorFreeDecoupledHead |
Supported | Supported |
Field list¶
| Field |
Description |
|---|---|
name |
(str) Name must be "mixnet" to use MixNet backbone. |
params.stem_channels |
(int) Output dimension of the first convolution layer. |
params.wid_mul |
(float) Ratio for adjusting the input/output dimensions for the entire model. |
params.dep_mul |
(float) Ratio for adjusting the num_block value for the entire model. |
params.dropout_rate |
(float) Dropout ratio applied to all MixDepthBlock. |
stage_params[n].expansion_ratio |
(list[int]) Determines the output dimension of the expansion phase in each MixDepthBlock. Expands the input dimensions by multiplying the expansion_ratio. |
stage_params[n].out_channels |
(list[int]) Output dimensions of each MixDepthBlock. |
stage_params[n].num_blocks |
(list[int]) Repetition count for each MixDepthBlock. |
stage_params[n].kernel_sizes |
(list[list[int]]) Various kernel sizes used within each MixDepthBlock. |
stage_params[n].num_exp_groups |
(list[int]) The number of convolution groups in the expansion phase of MixDepthBlock. |
stage_params[n].num_poi_groups |
(list[int]) The number of convolution groups in the final point-wise convolution of MixDepthBlock. |
stage_params[n].stride |
(list[int]) Stride values for each MixDepthBlock. |
stage_params[n].act_type |
(list[str]) Activation function for each MixDepthBlock. Supporting activation functions are described in [here] |
stage_params[n].se_reduction_ratio |
(list[int]) Reduction factor for calculating the output dimension of the squeeze-and-excitation block. If None, the squeeze-and-excitation block is not applied. |
Model configuration examples¶
MixNet-s
model:
architecture:
backbone:
name: mixnet
params:
stem_channels: 16
wid_mul: 1.0
dep_mul: 1.0
dropout_rate: 0.
stage_params:
-
expansion_ratio: [1, 6, 3]
out_channels: [16, 24, 24]
num_blocks: [1, 1, 1]
kernel_sizes: [[3], [3], [3]]
num_exp_groups: [1, 2, 2]
num_poi_groups: [1, 2, 2]
stride: [1, 2, 1]
act_type: ["relu", "relu", "relu"]
se_reduction_ratio: [~, ~, ~]
-
expansion_ratio: [6, 6]
out_channels: [40, 40]
num_blocks: [1, 3]
kernel_sizes: [[3, 5, 7], [3, 5]]
num_exp_groups: [1, 2]
num_poi_groups: [1, 2]
stride: [2, 1]
act_type: ["swish", "swish"]
se_reduction_ratio: [2, 2]
-
expansion_ratio: [6, 6, 6, 3]
out_channels: [80, 80, 120, 120]
num_blocks: [1, 2, 1, 2]
kernel_sizes: [[3, 5, 7], [3, 5], [3, 5, 7], [3, 5, 7, 9]]
num_exp_groups: [1, 1, 2, 2]
num_poi_groups: [2, 2, 2, 2]
stride: [2, 1, 1, 1]
act_type: ["swish", "swish", "swish", "swish"]
se_reduction_ratio: [4, 4, 2, 2]
-
expansion_ratio: [6, 6]
out_channels: [200, 200]
num_blocks: [1, 2]
kernel_sizes: [[3, 5, 7, 9, 11], [3, 5, 7, 9]]
num_exp_groups: [1, 1]
num_poi_groups: [1, 2]
stride: [2, 1]
act_type: ["swish", "swish"]
se_reduction_ratio: [2, 2]
MixNet-m
model:
architecture:
backbone:
name: mixnet
params:
stem_channels: 24
wid_mul: 1.0
dep_mul: 1.0
dropout_rate: 0.
stage_params:
-
expansion_ratio: [1, 6, 3]
out_channels: [24, 32, 32]
num_blocks: [1, 1, 1]
kernel_sizes: [[3], [3, 5, 7], [3]]
num_exp_groups: [1, 2, 2]
num_poi_groups: [1, 2, 2]
stride: [1, 2, 1]
act_type: ["relu", "relu", "relu"]
se_reduction_ratio: [~, ~, ~]
-
expansion_ratio: [6, 6]
out_channels: [40, 40]
num_blocks: [1, 3]
kernel_sizes: [[3, 5, 7, 9], [3, 5]]
num_exp_groups: [1, 2]
num_poi_groups: [1, 2]
stride: [2, 1]
act_type: ["swish", "swish"]
se_reduction_ratio: [2, 2]
-
expansion_ratio: [6, 6, 6, 3]
out_channels: [80, 80, 120, 120]
num_blocks: [1, 3, 1, 3]
kernel_sizes: [[3, 5, 7], [3, 5, 7, 9], [3], [3, 5, 7, 9]]
num_exp_groups: [1, 2, 1, 2]
num_poi_groups: [1, 2, 1, 2]
stride: [2, 1, 1, 1]
act_type: ["swish", "swish", "swish", "swish"]
se_reduction_ratio: [4, 4, 2, 2]
-
expansion_ratio: [6, 6]
out_channels: [200, 200]
num_blocks: [1, 3]
kernel_sizes: [[3, 5, 7, 9], [3, 5, 7, 9]]
num_exp_groups: [1, 1]
num_poi_groups: [1, 2]
stride: [2, 1]
act_type: ["swish", "swish"]
se_reduction_ratio: [2, 2]
MixNet-l
model:
architecture:
backbone:
name: mixnet
params:
stem_channels: 24
wid_mul: 1.3
dep_mul: 1.0
dropout_rate: 0.
stage_params:
-
expansion_ratio: [1, 6, 3]
out_channels: [24, 32, 32]
num_blocks: [1, 1, 1]
kernel_sizes: [[3], [3, 5, 7], [3]]
num_exp_groups: [1, 2, 2]
num_poi_groups: [1, 2, 2]
stride: [1, 2, 1]
act_type: ["relu", "relu", "relu"]
se_reduction_ratio: [~, ~, ~]
-
expansion_ratio: [6, 6]
out_channels: [40, 40]
num_blocks: [1, 3]
kernel_sizes: [[3, 5, 7, 9], [3, 5]]
num_exp_groups: [1, 2]
num_poi_groups: [1, 2]
stride: [2, 1]
act_type: ["swish", "swish"]
se_reduction_ratio: [2, 2]
-
expansion_ratio: [6, 6, 6, 3]
out_channels: [80, 80, 120, 120]
num_blocks: [1, 3, 1, 3]
kernel_sizes: [[3, 5, 7], [3, 5, 7, 9], [3], [3, 5, 7, 9]]
num_exp_groups: [1, 2, 1, 2]
num_poi_groups: [1, 2, 1, 2]
stride: [2, 1, 1, 1]
act_type: ["swish", "swish", "swish", "swish"]
se_reduction_ratio: [4, 4, 2, 2]
-
expansion_ratio: [6, 6]
out_channels: [200, 200]
num_blocks: [1, 3]
kernel_sizes: [[3, 5, 7, 9], [3, 5, 7, 9]]
num_exp_groups: [1, 1]
num_poi_groups: [1, 2]
stride: [2, 1]
act_type: ["swish", "swish"]
se_reduction_ratio: [2, 2]