Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Variants for EfficientRep and RepPANNeck #80

Merged
merged 7 commits into from
Sep 30, 2024
Merged

Variants for EfficientRep and RepPANNeck #80

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
9869dee
Added medium and large options for EfficientRep and RepPANNeck with C…
klemen1999 Sep 22, 2024
0786779
updated docs
klemen1999 Sep 22, 2024
3b7cccb
fixed indented epytext field
kozlov721 Sep 23, 2024
c39a85d
added tests for variants
klemen1999 Sep 27, 2024
c792a34
Merge branch 'feat/det_model_variants' of https://github.com/luxonis/…
klemen1999 Sep 27, 2024
30ae9da
added return type hint
klemen1999 Sep 27, 2024
faf6e63
removed extra code
klemen1999 Sep 28, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 19 additions & 14 deletions luxonis_train/nodes/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,13 +74,15 @@ Adapted from [here](https://arxiv.org/pdf/2209.02976.pdf).

**Params**

| Key | Type | Default value | Description |
| ------------- | ----------- | --------------------------- | --------------------------------------------------- |
| channels_list | List\[int\] | \[64, 128, 256, 512, 1024\] | List of number of channels for each block |
| n_repeats | List\[int\] | \[1, 6, 12, 18, 6\] | List of number of repeats of RepVGGBlock |
| in_channels | int | 3 | Number of input channels, should be 3 in most cases |
| depth_mul | int | 0.33 | Depth multiplier |
| width_mul | int | 0.25 | Width multiplier |
| Key | Type | Default value | Description |
| ------------- | ----------------------------------------------------------------- | --------------------------- | --------------------------------------------------------------- |
| variant | Literal\["n", "nano", "s", "small", "m", "medium", "l", "large"\] | "nano" | Variant of the network |
| channels_list | List\[int\] | \[64, 128, 256, 512, 1024\] | List of number of channels for each block |
| n_repeats | List\[int\] | \[1, 6, 12, 18, 6\] | List of number of repeats of RepVGGBlock |
| depth_mul | float | 0.33 | Depth multiplier |
| width_mul | float | 0.25 | Width multiplier |
| block | Literal\["RepBlock", "CSPStackRepBlock"\] | "RepBlock" | Base block used |
| csp_e | float | 0.5 | Factor for intermediate channels when block=="CSPStackRepBlock" |

## RexNetV1_lite

Expand Down Expand Up @@ -143,13 +145,16 @@ Adapted from [here](https://arxiv.org/pdf/2209.02976.pdf).

**Params**

| Key | Type | Default value | Description |
| ------------- | ---------------- | ------------------------------------------------------- | ----------------------------------------- |
| n_heads | Literal\[2,3,4\] | 3 ***Note:** Should be same also on head in most cases* | Number of output heads |
| channels_list | List\[int\] | \[256, 128, 128, 256, 256, 512\] | List of number of channels for each block |
| n_repeats | List\[int\] | \[12, 12, 12, 12\] | List of number of repeats of RepVGGBlock |
| depth_mul | int | 0.33 | Depth multiplier |
| width_mul | int | 0.25 | Width multiplier |
| Key | Type | Default value | Description |
| ------------- | ----------------------------------------------------------------- | ------------------------------------------------------- | --------------------------------------------------------------- |
| variant | Literal\["n", "nano", "s", "small", "m", "medium", "l", "large"\] | "nano" | Variant of the network |
| n_heads | Literal\[2,3,4\] | 3 ***Note:** Should be same also on head in most cases* | Number of output heads |
| channels_list | List\[int\] | \[256, 128, 128, 256, 256, 512\] | List of number of channels for each block |
| n_repeats | List\[int\] | \[12, 12, 12, 12\] | List of number of repeats of RepVGGBlock |
| depth_mul | float | 0.33 | Depth multiplier |
| width_mul | float | 0.25 | Width multiplier |
| block | Literal\["RepBlock", "CSPStackRepBlock"\] | "RepBlock" | Base block used |
| csp_e | float | 0.5 | Factor for intermediate channels when block=="CSPStackRepBlock" |

## ClassificationHead

Expand Down
44 changes: 32 additions & 12 deletions luxonis_train/nodes/backbones/efficientrep/efficientrep.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,11 +1,12 @@
import logging
from typing import Any
from typing import Any, Literal

from torch import Tensor, nn

from luxonis_train.nodes.base_node import BaseNode
from luxonis_train.nodes.blocks import (
BlockRepeater,
CSPStackRepBlock,
RepVGGBlock,
SpatialPyramidPoolingBlock,
)
Expand All @@ -26,23 +27,26 @@ def __init__(
n_repeats: list[int] | None = None,
depth_mul: float | None = None,
width_mul: float | None = None,
block: Literal["RepBlock", "CSPStackRepBlock"] | None = None,
csp_e: float | None = None,
**kwargs: Any,
):
"""Implementation of the EfficientRep backbone.
"""Implementation of the EfficientRep backbone. Supports the
version with RepBlock and CSPStackRepBlock (for larger networks)

Adapted from U{YOLOv6: A Single-Stage Object Detection Framework
for Industrial Applications
<https://arxiv.org/pdf/2209.02976.pdf>}.

@type variant: Literal["n", "nano", "s", "small", "m", "medium", "l", "large"]
@param variant: EfficientRep variant. Defaults to "nano".
The variant determines the depth and width multipliers.
The variant determines the depth and width multipliers, block used and intermediate channel scaling factor.
The depth multiplier determines the number of blocks in each stage and the width multiplier determines the number of channels.
The following variants are available:
- "n" or "nano" (default): depth_multiplier=0.33, width_multiplier=0.25
- "s" or "small": depth_multiplier=0.33, width_multiplier=0.50
- "m" or "medium": depth_multiplier=0.60, width_multiplier=0.75
- "l" or "large": depth_multiplier=1.0, width_multiplier=1.0
- "n" or "nano" (default): depth_multiplier=0.33, width_multiplier=0.25, block=RepBlock, e=None
- "s" or "small": depth_multiplier=0.33, width_multiplier=0.50, block=RepBlock, e=None
- "m" or "medium": depth_multiplier=0.60, width_multiplier=0.75, block=CSPStackRepBlock, e=2/3
- "l" or "large": depth_multiplier=1.0, width_multiplier=1.0, block=CSPStackRepBlock, e=1/2
@type channels_list: list[int] | None
@param channels_list: List of number of channels for each block. If unspecified,
defaults to [64, 128, 256, 512, 1024].
Expand All @@ -53,12 +57,19 @@ def __init__(
@param depth_mul: Depth multiplier. If provided, overrides the variant value.
@type width_mul: float
@param width_mul: Width multiplier. If provided, overrides the variant value.
@type block: Literal["RepBlock", "CSPStackRepBlock"] | None
@param block: Base block used when building the backbone. If provided, overrides the variant value.
@tpe csp_e: float | None
@param csp_e: Factor that controls number of intermediate channels if block="CSPStackRepBlock". If provided,
overrides the variant value.
"""
super().__init__(**kwargs)

var = get_variant(variant)
depth_mul = depth_mul or var.depth_multiplier
width_mul = width_mul or var.width_multiplier
block = block or var.block
csp_e = csp_e or var.csp_e or 0.5

channels_list = channels_list or [64, 128, 256, 512, 1024]
n_repeats = n_repeats or [1, 6, 12, 18, 6]
Expand All @@ -85,11 +96,20 @@ def __init__(
kernel_size=3,
stride=2,
),
BlockRepeater(
block=RepVGGBlock,
in_channels=channels_list[i + 1],
out_channels=channels_list[i + 1],
n_blocks=n_repeats[i + 1],
(
BlockRepeater(
block=RepVGGBlock,
in_channels=channels_list[i + 1],
out_channels=channels_list[i + 1],
n_blocks=n_repeats[i + 1],
)
if block == "RepBlock"
else CSPStackRepBlock(
in_channels=channels_list[i + 1],
out_channels=channels_list[i + 1],
n_blocks=n_repeats[i + 1],
e=csp_e,
)
),
)
self.blocks.append(curr_block)
Expand Down
10 changes: 10 additions & 0 deletions luxonis_train/nodes/backbones/efficientrep/variants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,25 +10,35 @@
class EfficientRepVariant(BaseModel):
depth_multiplier: float
width_multiplier: float
block: Literal["RepBlock", "CSPStackRepBlock"]
csp_e: float | None


def get_variant(variant: VariantLiteral) -> EfficientRepVariant:
variants = {
"n": EfficientRepVariant(
depth_multiplier=0.33,
width_multiplier=0.25,
block="RepBlock",
csp_e=None,
),
"s": EfficientRepVariant(
depth_multiplier=0.33,
width_multiplier=0.50,
block="RepBlock",
csp_e=None,
),
"m": EfficientRepVariant(
depth_multiplier=0.60,
width_multiplier=0.75,
block="CSPStackRepBlock",
csp_e=2 / 3,
),
"l": EfficientRepVariant(
depth_multiplier=1.0,
width_multiplier=1.0,
block="CSPStackRepBlock",
csp_e=1 / 2,
),
}
variants["nano"] = variants["n"]
Expand Down
6 changes: 2 additions & 4 deletions luxonis_train/nodes/blocks/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,15 +4,14 @@
BlockRepeater,
Bottleneck,
ConvModule,
CSPStackRepBlock,
DropPath,
EfficientDecoupledBlock,
FeatureFusionBlock,
KeypointBlock,
LearnableAdd,
LearnableMulAddConv,
LearnableMultiply,
RepDownBlock,
RepUpBlock,
RepVGGBlock,
SpatialPyramidPoolingBlock,
SqueezeExciteBlock,
Expand All @@ -26,18 +25,17 @@
"EfficientDecoupledBlock",
"ConvModule",
"UpBlock",
"RepDownBlock",
"SqueezeExciteBlock",
"RepVGGBlock",
"BlockRepeater",
"CSPStackRepBlock",
"AttentionRefinmentBlock",
"SpatialPyramidPoolingBlock",
"FeatureFusionBlock",
"LearnableAdd",
"LearnableMultiply",
"LearnableMulAddConv",
"KeypointBlock",
"RepUpBlock",
"BasicResNetBlock",
"Bottleneck",
"UpscaleOnline",
Expand Down
Loading