Skip to content

Commit

Permalink
add
Browse files Browse the repository at this point in the history
  • Loading branch information
@WAMAWAMA
WAMAWAMA committed Nov 9, 2022
1 parent eaa1701 commit 8fe8263
Show file tree
Hide file tree
Showing 124 changed files with 439 additions and 9,227 deletions.
30 changes: 21 additions & 9 deletions README.md
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,5 @@
# ωαмα m⚙️dules

(🚧 still under building, but feel free to try current module implementations~)

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

Expand All @@ -11,8 +10,14 @@ Highlights (*Simple-to-use & Function-rich!*)
- Easy and flexible to integrate with any other network
- 🚀 Abundant Pretrained weights: Including 80000+ `2D weights` and 80+ `3D weights`

*Download all pretrained weights from
[[Google Drive]](https://drive.google.com/drive/folders/1Pbgrgf8Y8soc6RGMh3bSUSWVIW2WeEmA?usp=sharing)
or [[Baidu Netdisk `psw: wama` ]](https://pan.baidu.com/s/16sskSM5IuHLbXOC4YF5MQQ?pwd=wama)


## 1. Installation
🔥 [wama_modules](https://github.com/WAMAWAMA/wama_modules) `Basic` `1D` `2D` `3D`
🔥 [wama_modules](https://github.com/WAMAWAMA/wama_modules)
`Basic` `1D/2D/3D`

Install `wama_modules` with command ↓
```
Expand All @@ -21,7 +26,8 @@ pip install git+https://github.com/WAMAWAMA/wama_modules.git

*Or you can directly copy the `wama_modules` folder to use

💧 [segmentation_models_pytorch](https://github.com/qubvel/segmentation_models.pytorch) `Optional` `2D` `100+ pretrained weights`
💧 [segmentation_models_pytorch](https://github.com/qubvel/segmentation_models.pytorch)
`Optional` `2D` `100+ pretrained weights`

<details>
<summary> Introduction and installation command </summary>
Expand All @@ -42,7 +48,8 @@ pip install git+https://github.com/rwightman/pytorch-image-models.git

</details>

💧 [transformers](https://github.com/huggingface/transformers) `Optional` `2D` `80000+ pretrained weights`
💧 [transformers](https://github.com/huggingface/transformers)
`Optional` `2D` `80000+ pretrained weights`

<details>
<summary> Introduction and installation command </summary>
Expand All @@ -57,7 +64,8 @@ pip install transformers

</details>

💧 [timm](https://github.com/rwightman/pytorch-image-models) `Optional` `2D` `400+ pretrained weights`
💧 [timm](https://github.com/rwightman/pytorch-image-models)
`Optional` `2D` `400+ pretrained weights`

<details>
<summary> Introduction and installation command </summary>
Expand Down Expand Up @@ -377,16 +385,20 @@ input = torch.ones([3,3,128,128])

## 5. Guideline 2: Use pretrained weights

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

current pretrained support: (这里给一个表格，来自哪里，多少权重，预训练数据类型，2D还是3D）)
- 2D: smp, timm, radimagenet...
- 3D: medicalnet, 3D resnet, 3D densenet...


Download all pretrained weights from
[[Google Drive]](https://drive.google.com/drive/folders/1Pbgrgf8Y8soc6RGMh3bSUSWVIW2WeEmA?usp=sharing)
or [[Baidu Netdisk `psw: wama` ]](https://pan.baidu.com/s/16sskSM5IuHLbXOC4YF5MQQ?pwd=wama)

### 5.1 smp encoders `2D`

smp (119 pretrained weights)
smp (119 pretrained weights, automatic online download)

```python
import torch
Expand All @@ -399,7 +411,7 @@ _ = [print(i.shape) for i in f_list]
```

### 5.2 timm encoders `2D`
timm (400+ pretrained weights)
timm (400+ pretrained weights, automatic online download)
```python
import timm
m = timm.create_model(
Expand All @@ -411,7 +423,7 @@ _ = [print(i.shape) for i in f_list]
```
### 5.3 Transformers (🤗 Huggingface ) `2D`

transformers, supper powered by Huggingface ( with 80000+ pretrained weights)
transformers, supper powered by Huggingface ( with 80000+ pretrained weights, automatic online download [see modelhub] )

```python
import torch
Expand Down
239 changes: 195 additions & 44 deletions tmp.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@
from wama_modules.thirdparty_lib.ResNets3D_kenshohara.resnet import generate_model
from wama_modules.utils import load_weights
m = generate_model(18)
pretrain_path = r"D:\pretrainedweights\ResNets3D_kenshohara\weights\resnet\r3d18_KM_200ep.pth"
pretrain_path = r"D:\pretrainedweights\ResNets3D_kenshohara\kenshohara_ResNets3D_weights\resnet\r3d18_KM_200ep.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights)
f_list = m(torch.ones([2,3,64,64,64]))
Expand All @@ -16,7 +16,7 @@
from wama_modules.thirdparty_lib.ResNets3D_kenshohara.resnet2p1d import generate_model
from wama_modules.utils import load_weights
m = generate_model(18)
pretrain_path = r"D:\pretrainedweights\ResNets3D_kenshohara\weights\resnet2p1d\r2p1d18_K_200ep.pth"
pretrain_path = r"D:\pretrainedweights\ResNets3D_kenshohara\kenshohara_ResNets3D_weights\resnet2p1d\r2p1d18_K_200ep.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights)
f_list = m(torch.ones([2,3,64,64,64]))
Expand Down Expand Up @@ -57,74 +57,225 @@


# todo 3 3D Efficient3D_okankop (39 weights)
if True:
# c3d
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.c3d import get_model
m = get_model() # c3d has no pretrained weights
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]


# mobilenet
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.mobilenet import get_model
from wama_modules.utils import load_weights
m = get_model(width_mult = 1.) # e.g. width_mult = 1 when mobilenet_1.0x
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\mobilenet\jester_mobilenet_1.0x_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

# todo 4 3D MedicalNet_tencent (11 weights)
m = get_model(width_mult = 2.) # e.g. width_mult = 2 when mobilenet_2.0x
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\mobilenet\jester_mobilenet_2.0x_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]


# mobilenetv2
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.mobilenetv2 import get_model
from wama_modules.utils import load_weights
m = get_model(width_mult = 1.) # e.g. width_mult = 1 when mobilenet_1.0x
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\mobilenetv2\jester_mobilenetv2_1.0x_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

# todo 5 3D C3D_jfzhang95 (1 weight)

m = get_model(width_mult = 0.45) # e.g. width_mult = 1 when mobilenet_1.0x
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\mobilenetv2\jester_mobilenetv2_0.45x_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]


# todo 6 3D C3D_yyuanad (1 weight)
# resnet
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.resnet import resnet18, resnet50, resnet101
from wama_modules.utils import load_weights
m = resnet18()
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\resnet\kinetics_resnet_18_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

m = resnet50()
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\resnet\kinetics_resnet_50_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

m = resnet101()
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\resnet\kinetics_resnet_101_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

# todo 7 2D smp (119 weight)
# smp
import torch
from wama_modules.thirdparty_lib.SMP_qubvel.encoders import get_encoder
m = get_encoder('resnet18', in_channels=3, depth=5, weights='ssl')
m = get_encoder('name', in_channels=3, depth=5, weights='ssl')
m = get_encoder('resnet18', in_channels=3, depth=5, weights='ss')
f_list = m(torch.ones([2,3,128,128]))
_ = [print(i.shape) for i in f_list]

# resnext
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.resnext import resnext101
from wama_modules.utils import load_weights
m = resnext101()
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\resnext\jester_resnext_101_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

# todo 8 timm (400+)
import timm
m = timm.create_model(
'adv_inception_v3',
features_only=True,
pretrained=False,)
f_list = m(torch.ones([2,3,128,128]))
_ = [print(i.shape) for i in f_list]

# shufflenet
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.shufflenet import get_model
from wama_modules.utils import load_weights
m = get_model(groups=3, width_mult=1)
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\shufflenet\jester_shufflenet_1.0x_G3_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

m = get_model(groups=3, width_mult=1.5)
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\shufflenet\jester_shufflenet_1.5x_G3_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]


# shufflenetv2
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.shufflenetv2 import get_model
from wama_modules.utils import load_weights
m = get_model(width_mult=1)
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\shufflenetv2\jester_shufflenetv2_1.0x_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]

m = get_model(width_mult=2)
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\shufflenetv2\jester_shufflenetv2_2.0x_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]


# squeezenet
import torch
from wama_modules.thirdparty_lib.Efficient3D_okankop.models.squeezenet import get_model
from wama_modules.utils import load_weights
m = get_model()
pretrain_path = r"D:\pretrainedweights\Efficient3D_okankop\Efficient3D_okankop_weights\squeezenet\jester_squeezenet_RGB_16_best.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 3, 64, 64, 64]))
_ = [print(i.shape) for i in f_list]


# todo 4 3D MedicalNet_tencent (11 weights)
if True:
import torch
from wama_modules.utils import load_weights
from wama_modules.thirdparty_lib.MedicalNet_Tencent.model import generate_model
m = generate_model(18)
pretrain_path = r"D:\pretrainedweights\MedicalNet_Tencent\MedicalNet_weights\resnet_18_23dataset.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')['state_dict']
m = load_weights(m, pretrain_weights, drop_modelDOT=True)
f_list = m(torch.ones([2, 1, 64, 64, 64])) # input channel is 1 (not 3 for video)
_ = [print(i.shape) for i in f_list]


# todo 5 3D C3D_jfzhang95 (1 weight)

# todo 9 transformers (80000+ weights)
import torch
from transformers import ConvNextModel
from wama_modules.utils import load_weights
# Initializing a model (with random weights) from the convnext-tiny-224 style configuration
m = ConvNextModel.from_pretrained('facebook/convnext-base-224-22k')
f = m(torch.ones([2,3,224,224]), output_hidden_states=True)
f_list = f.hidden_states
_ = [print(i.shape) for i in f_list]

weights = m.state_dict()
m1 = ConvNextModel(m.config)
m = load_weights(m, weights)

# todo 6 3D C3D_yyuanad (1 weight)



# todo 7 2D smp (119 weight)
if True:
# smp
import torch
from wama_modules.thirdparty_lib.SMP_qubvel.encoders import get_encoder
m = get_encoder('resnet18', in_channels=3, depth=5, weights='ssl')
f_list = m(torch.ones([2,3,128,128]))
_ = [print(i.shape) for i in f_list]


# todo 8 timm (400+)
if True:
import timm
m = timm.create_model(
'adv_inception_v3',
features_only=True,
pretrained=True,)
f_list = m(torch.ones([2,3,128,128]))
_ = [print(i.shape) for i in f_list]

import torch
from transformers import SwinModel
from wama_modules.utils import load_weights

m = SwinModel.from_pretrained('microsoft/swin-base-patch4-window12-384')
f = m(torch.ones([2,3,384,384]), output_hidden_states=True)
f_list = f.reshaped_hidden_states # For transformer, should use reshaped_hidden_states
_ = [print(i.shape) for i in f_list]
# todo 9 transformers (80000+ weights)
if True:
import torch
from transformers import ConvNextModel
from wama_modules.utils import load_weights
# Initializing a model (with random weights) from the convnext-tiny-224 style configuration
m = ConvNextModel.from_pretrained('facebook/convnext-base-224-22k')
f = m(torch.ones([2,3,224,224]), output_hidden_states=True)
f_list = f.hidden_states
_ = [print(i.shape) for i in f_list]
# reload weights
weights = m.state_dict()
m1 = ConvNextModel(m.config)
m = load_weights(m, weights)

weights = m.state_dict()
m1 = SwinModel(m.config)
m = load_weights(m, weights)

import torch
from transformers import SwinModel
from wama_modules.utils import load_weights
m = SwinModel.from_pretrained('microsoft/swin-base-patch4-window12-384')
f = m(torch.ones([2,3,384,384]), output_hidden_states=True)
f_list = f.reshaped_hidden_states # For transformer, should use reshaped_hidden_states
_ = [print(i.shape) for i in f_list]
# reload weights
weights = m.state_dict()
m1 = SwinModel(m.config)
m = load_weights(m, weights)



# todo 10 radimagenet (4 weights)
if True:
import torch
from wama_modules.utils import load_weights
from wama_modules.thirdparty_lib.SMP_qubvel.encoders import get_encoder
m = get_encoder('resnet50', in_channels=3, depth=5, weights=None)
pretrain_path = r"D:\pretrainedweights\radimagnet\RadImageNet_models-20221104T172755Z-001\RadImageNet_models\RadImageNet-ResNet50_notop_torch.pth"
pretrain_weights = torch.load(pretrain_path, map_location='cpu')
m = load_weights(m, pretrain_weights)
f_list = m(torch.ones([2,3,128,128]))
_ = [print(i.shape) for i in f_list]



Expand Down
Loading

0 comments on commit 8fe8263

Please sign in to comment.