Name	Name	Last commit message	Last commit date
Latest commit History 34 Commits
.idea	.idea
wama_modules	wama_modules
Document_allmodules.md	Document_allmodules.md
README.md	README.md
requirements.txt	requirements.txt
setup.py	setup.py
tmp.py	tmp.py

ωαмα m⚙️dules

*A PyTorch module library for building 1D/2D/3D networks flexibly ~*

(Simple-to-use & Function-rich!)

Highlights

Simple code that show whole forward processes succinctly
Output rich features and attention map for fast reuse
Support 1D / 2D / 3D networks (CNNs, GNNs, Transformers...)
Easy and flexible to integrate with any other network
🚀 Abundant Pretrained weights: Including 80000+ 2D weights and 80+ 3D weights

*Download pretrained weights from [Google Drive] or [Baidu Netdisk psw: wama ]

*All modules are detailed in [Document] (🚧 still under building)

1. Installation

🔥 wama_modules Basic 1D/2D/3D

Install wama_modules with command ↓

pip install git+https://github.com/WAMAWAMA/wama_modules.git

Other ways to install (or use) wama_modules

Way1: Download code and run python setup.py install
Way2: Directly copy the folder wama_modules into your project path

💧 segmentation_models_pytorch Optional 2D 100+ pretrained weights

Introduction and installation command

segmentation_models_pytorch (called smp) is a 2D CNN lib including many backbones and decoders, which is highly recommended to install for cooperating with this library.

*Our codes have already contained smp, but you can still install the latest version with the code below.

Install with pip ↓

pip install segmentation-models-pytorch

Install the latest version ↓

pip install git+https://github.com/rwightman/pytorch-image-models.git

💧 transformers Optional 2D 80000+ pretrained weights

Introduction and installation command

transformer (powered by Huggingface) is a lib including super abundant CNN and Transformer structures, which is highly recommended to install for cooperating with this library.

Install transformer with pip ↓

pip install transformers

💧 timm Optional 2D 400+ pretrained weights

Introduction and installation command

timm is a lib includes abundant CNN and Transformer structures, which is highly recommended to install for cooperating with this library.

Install with pip ↓

pip install timm

Install the latest version ↓

pip install git+https://github.com/rwightman/pytorch-image-models.git

2. Update list

2022/11/11: The birthday of this code, version v0.0.1
...

3. Main modules and network architectures

An overview of this repo (let's call wama_modules as wm)

File	Description	Main class or function
`wm.utils`	Some operations on tensors and pre-training weights	`resizeTensor()` `tensor2array()` `load_weights()`
`wm.thirdparty_lib`	2D/3D network structures (CNN/GNN/Transformer) from other repositories, and all are with pre-trained weights 🚀	`MedicalNet` `C3D` `3D_DenseNet` `3D_shufflenet` `transformers.ConvNextModel` `transformers.SwinModel` `Radimagenet`
`wm.Attention`	Some attention-based plugins	`SCSEModule` `NonLocal`
`wm.BaseModule`	Basic modules(layers). For example, BottleNeck block (ResBlock) in ResNet, and DenseBlock in DenseNet, etc.	`MakeNorm()` `MakeConv()` `MakeActive()` `VGGBlock` `ResBlock` `DenseBlock`
`wm.Encoder`	Some encoders such like ResNet or DenseNet, but with more flexibility for building the network modularly, and 1D/2D/3D are all supported	`VGGEncoder` `ResNetEncoder` `DenseNetEncoder`
`wm.Decoder`	Some encoders with more flexibility for building the network modularly, and 1D/2D/3D are all supported	`UNet_decoder`
`wm.Neck`	Modules for making the multi-scale features (from encoder) interact with each other to generate stronger features	`FPN`
`wm.Transformer`	Some self-attention or cross-attention modules, which can be used to build ViT, DETR or TransUnet	`TransformerEncoderLayer` `TransformerDecoderLayer`

How to build your networks modularly and freely? 👉 See 'Guideline 1: Build networks modularly' below ~
How to use pretrained model with wm.thirdparty_lib? 👉 See 'Guideline 2: Use pretrained weights' below ~

4. Guideline 1: Build networks modularly

How to build a network modularly? Here's a paradigm:

'Design architecture according to tasks, pick modules according to architecture'

So, network architectures for different tasks can be viewed modularly such as:

vgg = vgg_encoder + cls_head
Unet = encoder + decoder + seg_ead
resnet = resnet_encoder + cls_head
densenet = densenet_encoder + cls_head
a multi-task net for classification and segmentation = encoder + decoder + cls_head + seg_head

For example, build a 3D resnet50

import wama_modules as ws
import torch

encoder = ws.resnet(input_channel = 3, per_stage_channel = [8,16,32,64], dim=3)
decoder = ws.unet(encoder = encoder, output_channel = 3, dim=3)

input = torch.ones([3,3,128,128])

Here are more demos shown below ↓ (Click to view codes, or visit the demo folder)

*Todo-demo list (under preparation and coming soon...) ↓

Demo1: Build a 2D vgg16

Demo2: Build a 3D resnet50

Demo3: Build a 3D densenet121

Demo4: Build a Unet

Demo5: Build a Unet with a resnet50 encoder

Demo6: Build a Unet with a resnet50 encoder and a FPN

Demo7: Build a multi-task model for segmentation and classification

Demo8: Build a C-tran model for multi-label classification

Demo9: Build a Q2L model for multi-label classification

Demo10: Build a ML-Decoder model for multi-label classification

Demo11: Build a ML-GCN model for multi-label classification

Demo12: Build a UCTransNet model for segmentation

Demo13: Build a model for multiple inputs (1D signal and 2D image)

Demo14: Build a 2D Unet with pretrained Resnet50 encoder (1D signal and 2D image)

Demo15: Build a 3D DETR model for object detection

Demo16: Build a 3D VGG with SE-attention module for multi-instanse classification

5. Guideline 2: Use pretrained weights

(*All pretrained weights are from third-party codes or repos)

Currently available pre-training models are shown below ↓

	Module name	Number of pretrained weights	Pretrained data	Dimension
1	`.ResNets3D_kenshohara`	21	video	3D
2	`.VC3D_kenshohara`	13	video	3D
3	`.Efficient3D_okankop`	39	video	3D
4	`.MedicalNet_tencent`	11	medical image	3D
5	`.C3D_jfzhang95`	1	video	3D
6	`.C3D_yyuanad`	1	video	3D
7	`.SMP_qubvel`	119	image	2D
8	`timm`	400+	image	2D
9	`transformers`	80000+	video/image	2D/3D
10	`radimagenet`	1	medical image	2D

*Download all pretrained weights from [Google Drive] or [Baidu Netdisk psw: wama ]