oukohou / ssr_net_pytorch Goto Github PK

View Code? Open in Web Editor NEW

65.0 2.0 16.0 507 KB

a Pytorch reimplementation of SSRNet.

Python 100.00%

age ssr-net

ssr_net_pytorch's Introduction

author : oukohou
time : 2019-09-26 16:44:48
email : [email protected]

A pytorch reimplementation of SSR-Net.

the official keras version is here: SSR-Net

results on MegaAge_Asian datasets:

-	train	valid	test
version_v1^[1]	train Loss: 22.0870 CA_3: 0.5108, CA_5: 0.7329	val Loss: 44.7439 CA_3: 0.4268, CA_5: 0.6225	test Loss: 35.6759 CA_3: 0.4935, CA_5: 0.6902
original paper	**	**	CA_3: 0.549, CA_5: 0.741
version_v2^[2]	train Loss: 2.9401 CA_3: 0.6326, CA_5: 0.8123	val Loss: 4.7221 CA_3: 0.4438, CA_5: 0.6295	test Loss: 3.9311 CA_3: 0.5151, CA_5: 0.7163

Note:

This SSR-Net model can't fit big learning rate, learning rate should be smaller than 0.002. otherwise the model will very likely always output 0, me myself suspects this is because of the utilizing Tanh as activation function.

And also: Batchsize could severely affect the results. A set of tested params can be :

batch_size = 50
input_size = 64
num_epochs = 90
learning_rate = 0.001 # originally 0.001
weight_decay = 1e-4 # originally 1e-4
augment = False
optimizer_ft = optim.Adam(params_to_update, lr=learning_rate, weight_decay=weight_decay)
criterion = nn.L1Loss()
lr_scheduler = optim.lr_scheduler.StepLR(optimizer_ft, step_size=30, gamma=0.1)

The dataset preprocess is quite easy. For MegaAsian datasets, you can use the ./datasets/read_megaasina_data.py directly; for other datasets, just generate a pandas csv file in format like:
```
filename,age
1.jpg,23
...
```

is OK. But also, remember to change the ./datasets/read_imdb_data.py accordingly.

onnxruntime C++ implementation

thanks to DefTruth 's implementation here: How to convert SSRNet to ONNX and implements with onnxruntime c++.

another small note:

my reading understanding of SSRNet can be found:

on my blog site here:论文阅读-年龄估计_SSRNet
or on zhihu here: 论文阅读-年龄估计_SSRNet.

which was written in Chinese.

ssr_net_pytorch's People

Contributors

Stargazers

Watchers

Forkers

bulingda deftruth jingziyou aascode nodototaofordl starstylesky changkaizhi gogogogojudy gykeve wqz960 lhf981128 jacke121 seulegend fffaded codewithleo1103

ssr_net_pytorch's Issues

咨询疑问

您好，如果我使用输入的不是图片，是3维矩阵（大脑的3维矩阵.mat）来做这样的年龄预测估计，您感觉如果使用您的方法和代码修改后可行吗？

使用处理过后的IMDB数据集运行train_SSR-Net.py报错，初学者不太会，望作者解答，不胜感激

C:\ProgramData\Anaconda3\python.exe D:/PycharmProjects/SSR_Net_Pytorch-master/train_SSR-Net.py
Traceback (most recent call last):
File "D:/PycharmProjects/SSR_Net_Pytorch-master/train_SSR-Net.py", line 214, in
num_epochs_=num_epochs,
File "D:/PycharmProjects/SSR_Net_Pytorch-master/train_SSR-Net.py", line 63, in train_model
for i, (inputs, labels) in enumerate(dataloaders_[phase]):
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 615, in next
batch = self.collate_fn([self.dataset[i] for i in indices])
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 615, in
batch = self.collate_fn([self.dataset[i] for i in indices])
File "D:\PycharmProjects\SSR_Net_Pytorch-master\datasets\read_imdb_data.py", line 35, in getitem
image_, image_path_ = self.read_images(index)
File "D:\PycharmProjects\SSR_Net_Pytorch-master\datasets\read_imdb_data.py", line 52, in read_images
filename = self.images_df.iloc[index_].Filename
File "C:\ProgramData\Anaconda3\lib\site-packages\pandas\core\generic.py", line 4372, in getattr
return object.getattribute(self, name)
AttributeError: 'Series' object has no attribute 'Filename'

read_imdb_data.py

34 def getitem(self, index):
35 image_, image_path_ = self.read_images(index)
if self.mode in ['train', ]:
label = self.images_df.iloc[index].age
else:
label = image_path_
if self.augment:
image_ = self.augmentor(image_)
image_ = T.Compose([
T.ToPILImage(),
# T.RandomResizedCrop(self.input_size),
# T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])(image_)
return image_.float(), label

51def read_images(self, index_):
52 filename = self.images_df.iloc[index_].Filename
image_path_ = os.path.join(self.base_path, filename)
image = cv2.imread(image_path_)
return image, image_path_

onnxruntime c++ implementation for SSR_Net_Pytorch.

我给这个项目写了一份onnxruntime c++版本的推理，在ort_ssrnet-cn.md .

data augment

inference_images.py中的方法inference_single_image有如下代码：

image_ = cv2.imread(image_path_)
image_ = T.Compose([
T.ToPILImage(),
T.Resize((input_size_, input_size_)),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])(image_)

opencv读取的图像通道顺序为BGR，而[0.485, 0.456, 0.406], [0.229, 0.224, 0.225]是Imagenet的均值和标准差，RGB通道顺序，transforms.ToPILImage() 只是单纯的将格式变为PIL,没有转换通道顺序的功能。此处写法是否有问题？

人脸检测

不好意思，想请教一下模型输入是经过什么人脸检测啊，大概要裁剪到什么样子。

关于代码中self.lambda_index 的意思以及作用

您好，在读您的代码中的时候并没有看出来self.lambda_index 的作用，在原作者论文中也没有，能解释一下吗？

Data preprocessing

Hi there,

The official version needs to preprocess the image dataset, but I can not find this part in your code. Doesn't your code need this process?

Regards,

关于imdb wiki 数据集训练及测试

您好，请问您有用imdb wiki制作的数据集训练和测试嘛，我训练的效果很差，mae 只有8,+

About the used datasets of the pretrained model

Hi, oukohou!

Thanks for the pytorch version of SSR net.

I have ran the inference_image file for the prediction, pretty easy to use, and I have some questions.

1. code error

in the inference_images.py file, line44, the code is

image_ = image_.cuda()

but actually I don't have a gpu, so there will be an error occurred. I think

image_ = image_.to(device)

is better in my opinion.

2. datasets

I have read your read.me file. If I didn't misunderstand you, the only one pretrained model was trained on IMDB and Mega datasets. However, in the train_SSR-net.py file, I saw

from datasets.read_face_age_data import FaceAgeDatasets

Now I'm confused. Did you use face age dataset for training? this one?
Or did you just use IMDB and Mega only?

And if possible, could you tell me which solution(datasets) has the best performance?

Looking forward to hear from you soon.

如何扩大输入尺寸

@oukohou 你好，如果我想扩大模型输入尺寸，大概需要改那些东西？

你这个方法验证集年龄的mse是多少啊

我看你给的指标有对应年龄的3岁5岁误差准确度。对应的MegaAge_Asian上年龄平均误差MAE是多大呢？

How to improve the accuracy of a specific age group？

Hello， thanks for your work！I trained based on megaage and got the accuracy of CA5: 75%, but I found that for a certain age group, such as 50~~60, 60~~70 elderly people, 0~~10, 10~~20 young people, the accuracy of this age group is not ideal. Is it because there are relatively few age data on both sides of the dataset？
I need to invest more datasets in my estimated age？

Did you train the model on the IMDB_WIKI? In the doc, the version1 are trained from scratch on the MegaAge_Asian and version2 are fine-tuned on the MegaAge_Asian.
Did you calculate the MAE of the model on the MegaAge_Asian test dataset? my result is 11.09. Maybe is it a little bigger?