Please install
nb
withpip install nbnb
, thenb
name was taken already, so we usingnbnb
instead!
Nenural network Blocks (aka: NB, or neural network builder). This library provides massive fancy blocks for you to quick build your powerful networks! Some SOTA tricks and connections such as CSP, ASFF, Attention, BaseConv, Hardswish, Mish all included for quick prototype your model. This is an Arsenal for deeplearning forge.
nb is an idea comes from engineering, we build model with some common blocks, we exploring new ideas with SOTA tricks, but all those thing can be gathered into one single place, and for model quick design and prototyping.
this project is under construct for now, I will update it quickly once I found some new blocks that really works in model. And you are very welcome to send me a pull request and add more features!
-
2022.06.27: Now we add
MobileOne
backbone support! -
2021.03.16: Added some blocks used inside Scaled-YoloV4 (P5,P6,P7). List are:
HarDBlock
;SPPCSP
;VoVCSP
; You can using these blocks to stack your model now.
from nb.torch.blocks.csp_blocks import VoVCSP
-
2021.01.22: Adding Mish activation function. You can call it in your model using the following code:
from nb.torch.base import build_activation_layer act = build_activation_layer(act_cfg=dict(type='Mish'))
-
2021.01.22: Adding Triplet Attention Mechanism. You can plug it in any of your conv net blocks using the following code:
from nb.torch.blocks.attention_blocks import TripletAttention att_mechanism = TripletAttention() rand_tensor = torch.rand(1,3,32,32) output = att_mechanism(rand_tensor)
TripletAttention is a shape preserving tensor which expects a 4-dimensional input (B,C,H,W) and outputs a 4-dimensional output of the same shape (B,C,H,W).
-
2021.01.14: Adding SiLU introduced from pytorch 1.7. And now you can build a activation layer by using:
from nb.torch.base import build_activation_layer act = build_activation_layer(act_cfg=dict(type='SiLU'))
Also PANet module also provided now. BiFPN is on the way. We will also provide more examples on how to using it!
-
2020.09.28: ASFF module added inside nb. We have a ASFF design version of YoloV5 now! Some experiment will add here once we confirm ASFF module enhance the model performance.
-
2020.09.22: New backbone of
Ghostnet
andMobilenetV3
included. Both of them can be used to replace any of your application's backbone. -
2020.09.14: We release a primary version of 0.04, which you can build a simple YoloV5 with nb easily!
pip install nbnb
-
2020.09.12: New backbone SpineNet added:
SpineNet is a backbone model specific for detection, it's a backbone but can do FPN's thing!! More info pls reference google's paper link.
from nb.torch.bakbones.spinenet import SpineNet model = SpineNet()
-
2020.09.11: New added blocks:
resnet.Bottleneck resnet.BasicBlock ConvBase
nb can be installed from PIP, remember the name is nbnb
:
sudo pip3 install nbnb
Here is an example of using NB to build YoloV5!
updates: We have another YoloV5-ASFF version added in example!
import torch
from torch import nn
from nb.torch.blocks.bottleneck_blocks import SimBottleneckCSP
from nb.torch.blocks.trans_blocks import Focus
from nb.torch.blocks.head_blocks import SPP
from nb.torch.blocks.conv_blocks import ConvBase
from nb.torch.utils import device
class YoloV5(nn.Module):
def __init__(self, num_cls=80, ch=3, anchors=None):
super(YoloV5, self).__init__()
assert anchors != None, 'anchor must be provided'
# divid by
cd = 2
wd = 3
self.focus = Focus(ch, 64//cd)
self.conv1 = ConvBase(64//cd, 128//cd, 3, 2)
self.csp1 = SimBottleneckCSP(128//cd, 128//cd, n=3//wd)
self.conv2 = ConvBase(128//cd, 256//cd, 3, 2)
self.csp2 = SimBottleneckCSP(256//cd, 256//cd, n=9//wd)
self.conv3 = ConvBase(256//cd, 512//cd, 3, 2)
self.csp3 = SimBottleneckCSP(512//cd, 512//cd, n=9//wd)
self.conv4 = ConvBase(512//cd, 1024//cd, 3, 2)
self.spp = SPP(1024//cd, 1024//cd)
self.csp4 = SimBottleneckCSP(1024//cd, 1024//cd, n=3//wd, shortcut=False)
# PANet
self.conv5 = ConvBase(1024//cd, 512//cd)
self.up1 = nn.Upsample(scale_factor=2)
self.csp5 = SimBottleneckCSP(1024//cd, 512//cd, n=3//wd, shortcut=False)
self.conv6 = ConvBase(512//cd, 256//cd)
self.up2 = nn.Upsample(scale_factor=2)
self.csp6 = SimBottleneckCSP(512//cd, 256//cd, n=3//wd, shortcut=False)
self.conv7 = ConvBase(256//cd, 256//cd, 3, 2)
self.csp7 = SimBottleneckCSP(512//cd, 512//cd, n=3//wd, shortcut=False)
self.conv8 = ConvBase(512//cd, 512//cd, 3, 2)
self.csp8 = SimBottleneckCSP(512//cd, 1024//cd, n=3//wd, shortcut=False)
def _build_backbone(self, x):
x = self.focus(x)
x = self.conv1(x)
x = self.csp1(x)
x_p3 = self.conv2(x) # P3
x = self.csp2(x_p3)
x_p4 = self.conv3(x) # P4
x = self.csp3(x_p4)
x_p5 = self.conv4(x) # P5
x = self.spp(x_p5)
x = self.csp4(x)
return x_p3, x_p4, x_p5, x
def _build_head(self, p3, p4, p5, feas):
h_p5 = self.conv5(feas) # head P5
x = self.up1(h_p5)
x_concat = torch.cat([x, p4], dim=1)
x = self.csp5(x_concat)
h_p4 = self.conv6(x) # head P4
x = self.up2(h_p4)
x_concat = torch.cat([x, p3], dim=1)
x_small = self.csp6(x_concat)
x = self.conv7(x_small)
x_concat = torch.cat([x, h_p4], dim=1)
x_medium = self.csp7(x_concat)
x = self.conv8(x_medium)
x_concat = torch.cat([x, h_p5], dim=1)
x_large = self.csp8(x)
return x_small, x_medium, x_large
def forward(self, x):
p3, p4, p5, feas = self._build_backbone(x)
xs, xm, xl = self._build_head(p3, p4, p5, feas)
return xs, xm, xl
A simple example to build a layer of conv:
from nb.torch.base.conv_block import ConvBase
a = ConvBase(128, 256, 3, 1, 2, norm_cfg=dict(type="BN"), act_cfg=dict(type="Hardswish"))
Be note that, the reason for us using cfg
to specific norm and activation is for users dynamically switch their configuration of model in yaml format rather than hard code it.
A simple example of using GhostNet:
from nb.torch.backbones.ghostnet import GhostNet
m = GhostNet(num_classes=8)
# if you want FPN output
m = GhostNet(fpn_levels=[4, 5, 6])
A simple example of using MobilenetV3:
from nb.torch.backbones.mobilenetv3_new import MobilenetV3_Small
We list all conv
and block
support in nb here:
conv
:- Conv
- ConvWS: https://arxiv.org/pdf/1903.10520.pdf
- ...
Blocks
:- CSPBlock:
@Lucas Jin all rights reserved.