shamangary / ssr-net

[IJCAI18] SSR-Net: A Compact Soft Stagewise Regression Network for Age Estimation

Academic Kickstart

Academic makes it easy to create a beautiful website for free using Markdown, Jupyter, or RStudio. Customize anything on your site with widgets, themes, and language packs. Check out the latest demo of what you'll get in less than 10 minutes, or view the showcase.

Academic Kickstart provides a minimal template to kickstart your new website.

• Get Started
• View the documentation
• Ask a question
• Request a feature or report a bug
• Updating? View the Update Guide and Release Notes
• Support development of Academic:
  • Donate a coffee
  • Become a backer on Patreon
  • Decorate your laptop or journal with an Academic sticker
  • Wear the T-shirt

Install

You can choose from one of the following four methods to install:

• one-click install using your web browser (recommended)
• install on your computer using Git with the Command Prompt/Terminal app
• install on your computer by downloading the ZIP files
• install on your computer with RStudio

Then personalize your new site.

Ecosystem

• Academic Admin: An admin tool to import publications from BibTeX or import assets for an offline site
• Academic Scripts: Scripts to help migrate content to new versions of Academic

License

Copyright 2017-present George Cushen.

Released under the MIT license.

ssr-net's Issues

keras version

你好，我能不能咨询一下，你在进行训练和测试的时候，采用的keras和Python的版本是多少呢？谢谢！

project environment

hello ,would u plz give detailed information about the project environment ,such as keras and tensorflow version ,cuda and cudnn?
when i run "python3 SSRNET_train.py --input /home/xd/workspace/SSR-Net/data/imdb.npz --db imdb --netType1 4 --netType2 4", the folloing errors occur:
Epoch 1/90
Exception in thread Thread-1:
Traceback (most recent call last):
File "/usr/lib/python3.5/threading.py", line 914, in _bootstrap_inner
self.run()
File "/usr/lib/python3.5/threading.py", line 862, in run
self._target(*self._args, **self._kwargs)
File "/home/xx/.local/lib/python3.5/site-packages/keras/engine/training.py", line 606, in data_generator_task
generator_output = next(self._generator)
File "/home/xx/workspace/SSR-Net/training_and_testing/TYY_generators.py", line 49, in data_generator_reg
yield augment_data(np.array(p)),np.array(q)
File "/home/xx/workspace/SSR-Net/training_and_testing/TYY_generators.py", line 27, in augment_data
images[i] = tf.contrib.keras.preprocessing.image.random_rotation(images[i], 20, row_axis=0, col_axis=1, channel_axis=2)
AttributeError: module 'tensorflow.contrib' has no attribute 'keras'
File "/home/xx/.local/lib/python3.5/site-packages/keras/engine/training.py", line 1852, in fit_generator
str(generator_output))
ValueError: output of generator should be a tuple (x, y, sample_weight) or (x, y). Found: None
Your reply would be appreciated greatly.

Training error

I have some problems to start training. See 2 errors below.

Error 1: it is looking for a pre-trained weights but not available from repo
cmd:> python SSRNET_train.py --input ../data/imdb_db.npz --db imdb--netType1 1 --netType2 1

OSError: Unable to open file (Unable to open file: name = 'imdb_models/ssrnet_3_3_3_64_0.25_0.25/ssrnet_3_3_3_64_0.25_0.25.h5', errno = 2, error message = 'no such file or directory', flags = 0, o_flags = 0)

Error 2: I solve error1 by replace the imdb_models with from another pre-train folder, I hv picked "ssrnet_3_3_3_64_1.0_1.0", it throws me another error below:

UnicodeDecodeError: 'rawunicodeescape' codec can't decode bytes in position 838-839: truncated \uXXXX

Details:

Total params: 40,915
Trainable params: 40,531
Non-trainable params: 384

---

DEBUG:root:Saving model...
Traceback (most recent call last):
File "SSRNET_train.py", line 136, in
main()
File "SSRNET_train.py", line 95, in main
f.write(model.to_json())
File "C:\Users\default.LAPTOP-2CI68M4P\Anaconda3\envs\xrvision2\lib\site-packages\keras\engine\topology.py", line 2618, in to_json
model_config = self._updated_config()
File "C:\Users\default.LAPTOP-2CI68M4P\Anaconda3\envs\xrvision2\lib\site-packages\keras\engine\topology.py", line 2585, in _updated_config
config = self.get_config()
File "C:\Users\default.LAPTOP-2CI68M4P\Anaconda3\envs\xrvision2\lib\site-packages\keras\engine\topology.py", line 2322, in get_config
layer_config = layer.get_config()
File "C:\Users\default.LAPTOP-2CI68M4P\Anaconda3\envs\xrvision2\lib\site-packages\keras\layers\core.py", line 656, in get_config
function = func_dump(self.function)
File "C:\Users\default.LAPTOP-2CI68M4P\Anaconda3\envs\xrvision2\lib\site-packages\keras\utils\generic_utils.py", line 175, in func_dump
code = marshal.dumps(func.code).decode('raw_unicode_escape')
UnicodeDecodeError: 'rawunicodeescape' codec can't decode bytes in position 838-839: truncated \uXXXX

Is the method applicable to CACD dataset?

Hi there @shamangary ,

I am wondering whether your method is applicable to CACD dataset: https://bcsiriuschen.github.io/CARC/ ?

In addition, why did you not crop the face images when pre-processing? I am thinking that removing more background noise is possible to get better performance.

Regards,

Why stage1 (layer4) has no pooling layer?

Excuse me. I read your code and I didn't see any pooling layer(AvgPooling or MaxPooling) in Stage1(code: layer4). But it does actually exist in Stage3(code: layer1) and Stage2(code: layer2).

Why is this difference? Thank you.

How did you crop and aligh MORPH II?

Hi, Thanks for sharing this awesome source code. It helps me a lot. Currently, I am trying to reproduce your result on morph2 dataset. Your script use morph2_db_align.npz file to train the network. Because the file's name has a '_align' at the end, so I suppose you made an preprocessing step for this dataset. Can you tell me which preprocessing you used?

请问您在预处理数据的时候对图像进行人脸识别并裁剪了吗？

您好，在预处理数据（create_db.py）时，我注意到你对原本数据集的图像进行了人脸识别，但好像并未对人脸部分进行裁剪。（若是看错，还请谅解）
当我利用MegaAsian的数据集进行训练时，我发现该数据集人脸外的背景并不小，担心“噪声”较大，这样训练的结果是否会对结果造成影响呢？（我裁剪后，对于小孩的判断有一定提升）
不知道您的看法如何？谢谢。

Training On IMBD

Hi,shamangary:
I use imdb data to train the model, use the parameter in your paper,but the training and val loss is very high,about 9.2,is that correct?

Inference time

Hi，shamangary：
how did you evaluate your cpu time？I use the following code to evaluate the inference time：
start_CPU = timeit.default_timer()
results = model.predict(faces)
end_CPU = timeit.default_timer()
in the TYY_demo_mtcnn.py code，but it takes about 20-30ms in cpu time，is it correct？

模型中的V=101是代表训练样本的最大年龄值吗

你好，模型中的V设置为101，是因为训练数据最大的年龄是101吗。如果自己的数据集，是不是也需要进行相应的调整

.h5 model to .pb model but output node is strange!

I transfer h5 model to .pb model but output node is strange, the output node is pred_a/mul_33, I can't find this node in the tensorflow graph, do you know the reason, thanks!

find pred_a/mul_33 node, but I don't know what the node means.

How to explain “netType1”“netType2” and “lambda_local”“lambda_local”?

hello! @shamangary

I want to understand what “netType1”“netType2” and “lambda_local”“lambda_local” stand for?

Thank you very much for your answer!

Morph2 dataset website is dead

Morph2 dataset website is dead. Do you happen to have the dataset for download?

wrong model folders

error in line 76 to 81 in master/training_and_testing/SSRNET_train.py
if db_name == "wiki":
weight_file = "wiki_models/"+save_name+"/"+save_name+".h5"
model.load_weights(weight_file)
elif db_name == "imdb":
weight_file = "imdb_models/"+save_name+"/"+save_name+".h5"
model.load_weights(weight_file)
elif db_name == "morph":
weight_file = "morph_models/"+save_name+"/"+save_name+".h5"
model.load_weights(weight_file)

another question

HI, @shamangary can I use this net to train age and gender at the same time? and if gender has 3 classes ,ie (0: other, 1: female, 2: male) , in merge_age function V should be set V=1 or V=3?

MORPH2 dataset

hello，
How to get the MORPH dataset ? If I want to use it as academic purpose ,can I apply the dataset by sending an official Email or I HAVE TO pay for it ?
thank you .

Why is the probability value not processed by softmax？

Sorry to bother you ,why pred_xxx do not processed by softmax？
If it is not processed by softmax, is it probabilistic?

Do we have to use the GTX1080 platform？

Do we have to use the GTX1080 platform？is RTX2060 OK？

Demo with an input image version

Hey, Is it possible to make a demo which runs on input images rather than gifs and mp4 videos?

Thank you

using Megaage-Asian train mae is 3.8,is this normal？

i train ssrnet with Megaage-Asian and get the mae is 3.8,is this normal？ what's your result about Megaage-Asian. Thanks

negative age

Hi，shamangary：
In the age prediction process，the age is negative，is there a way to avoid this？
Thanks。

转caffe模型得到结果不太一直，想向您确认一下预处理形式

模型转换后得到的[pred_a_s1,pred_a_s2,pred_a_s3,delta_s1,delta_s2,delta_s3, local_s1, local_s2, local_s3]不太一致。 根据你的demo, 我把格式设成了一致的，mean设为0，scale设为1，图像类型为BGR，未做对齐。想向您确认一下预处理是否不当？

Do you have plan to use tensorflow to train it directly?I need to use c++ and use opencv to load a tensorflow model.

how to implement model.predict() function in TYY_demo_mtcnn.py using c++

I need to engineer ssrnet. How do I plan my final age by getting 9 output nodes（ pred_age_stage1、 pred_age_stage2、 pred_age_stage3、delta_s1、delta_s2、delta_s3、local_delta_stage1、local_delta_stage2、local_delta_stage3）?

Who can help me? Thank you

Could you please tell me what does 'lambda-local' and 'lambda_d' mean in 'SSR_net'? Thank you!

Thank you for providing such an extrodinary image-processing solution, and I have already run the code and read the paper related to this project.

I have a question about parameters 'lambda_d' and 'lambda_local'. I can't understand their function and meaning. ( I noticed that you set them both as 1.0 in your demo).

性别识别

请问得到的性别识别的准确率有多高呢？

mtcnn

hello, I find you use mtcnn to detect face in video demo code, and install mtcnn with pip. So I want to ask you, this mtcnn whether or not Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks(MTCNN)？
thank you!

Does it have plans to support C++?

Hi
Does it have plans to support C++?

Network architicture

It seems like that the network architicture in the code of SSR_Net.py is different from paper.For example only one pooling layer in every stream of stage 1 but two in paper @shamangary

How about the performance without pretrained weight?

How about the performance without the pretrained weight (that is, training from scratch)?

about the results of MegaAgeAsian

In your paper, you got this result, is that on the training datasets or validation datasets? Cause I reimplement the SSRNet in pytorch, but the best results of 90 epochs is like :

train Loss: 22.0870 CA_3: 0.5108, CA_5: 0.7329
val Loss: 44.7439 CA_3: 0.4268, CA_5: 0.6225

the parameters are like:

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.MSELoss()
lr_scheduler = optim.lr_scheduler.StepLR(optimizer_ft, step_size=30, gamma=0.1)

Thanks in advance~

Do you have plan to add gender classification?

Hello

it is amazing model.

but it doesn't give gender classification.

I know

https://github.com/shamangary/Keras-real-time-age-gender-estimation-demo

give gender but SSR-net is more good accuracy.

do you have plan to add gender classification?

a Pytorch reimplementation of SSRNet.

@shamangary 你好，求大佬在readme的third party implementation里多加个链接哈哈～～：

Pytorch: https://github.com/oukohou/SSR_Net_Pytorch

让我也能蹭一波大佬的热度～～～谢谢～～

the result is very poor

it is amazing work, thank for u work. But the result is very poor except for gender . how can i improve the model

MegaAge-Asian Testing Script

Hi, in your paper, you have tested SSR-Net model on MegaAge-Asian dataset and got result [CA(3): 0.549, CA(5):0.741].
I modified the script TYY_demo_mtcnn.py, use your "pre-trained/megaface_asian/ssrnet_3_3_3_64_1.0_1.0/ssrnet_3_3_3_64_1.0_1.0.h5" and read MegaAge-Asian images as input to test.
But I can't get the approximate testing result, can you provide the testing script? Thank you!

how to test the model to get accuracy

Hello, how to test the model to get accuracy?

Error when training the dataset. Please help!

Could someone please help me. Im getting the error and I cannot understand why. I get it when I run the code:
python3 create_db.py --output data/imdb_db.mat --db imdb --img_size 64

Is it something to do with a dependency not installed?
Please help me!

Code Error in full;

(EnVee) D:\FYP\FYP Platforms\SSR-Net-master\data>python TYY_IMDBWIKI_create_db.py --db imdb --output imdb.npz
Traceback (most recent call last):
File "TYY_IMDBWIKI_create_db.py", line 61, in
main()
File "TYY_IMDBWIKI_create_db.py", line 34, in main
full_path, dob, gender, photo_taken, face_score, second_face_score, age = get_meta(mat_path, db)
File "D:\FYP\FYP Platforms\SSR-Net-master\data\TYY_utils.py", line 17, in get_meta
meta = loadmat(mat_path)
File "C:\Users\VenuraKetipearachchi\Anaconda3\envs\EnVee\lib\site-packages\scipy\io\matlab\mio.py", line 142, in loadmat
matfile_dict = MR.get_variables(variable_names)
File "C:\Users\VenuraKetipearachchi\Anaconda3\envs\EnVee\lib\site-packages\scipy\io\matlab\mio5.py", line 272, in get_variables
hdr, next_position = self.read_var_header()
File "C:\Users\VenuraKetipearachchi\Anaconda3\envs\EnVee\lib\site-packages\scipy\io\matlab\mio5.py", line 226, in read_var_header
mdtype, byte_count = self._matrix_reader.read_full_tag()
File "mio5_utils.pyx", line 548, in scipy.io.matlab.mio5_utils.VarReader5.read_full_tag
File "mio5_utils.pyx", line 556, in scipy.io.matlab.mio5_utils.VarReader5.cread_full_tag
File "streams.pyx", line 171, in scipy.io.matlab.streams.ZlibInputStream.read_into
File "streams.pyx", line 158, in scipy.io.matlab.streams.ZlibInputStream._fill_buffer
zlib.error: Error -3 while decompressing data: invalid code lengths set

change input size to 112X112

您好，请问下如果需要将输入图片修改为112X112，是否可以在第一个卷积前面加一个正常的3*3 ，stride为2的卷积就可以了。效果会比64X64的应该也会好吧。

Accuracy question about gender estimation

hello, I use SSR-Net to estimate gender with wiki pre-model, but the value obtained does not seem to be a confidence， the values of males are greater than 1, and the values of females is not less than 0.5, mostly is 0.9999.... , so there are many errors with 0.5 as threshold.
Is there a problem with my test?
Thank you!

how to use the gender model ?

hi,
In the pic
I find the gender prediction, so could you tell me where it is in the demo ?
I want to put it in the demo,
thx

one train question

here , SSR-NET use MAE to be loss function, I try to write code using tensorflow ,
and the error is like that:
ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients,

and the loss code like that:
logits_loss_age = tf.reduce_mean(tf.losses.absolute_difference(tf.cast(self.classifier_logits,tf.int32),age_labels))
optimizer = tf.train.AdamOptimizer(learning_rate,epsilon=1e-4)
train_op = optimizer.minimize(all_loss)

how can I resolve this quesiton?

Test image have margin can improve accuracy！

Hello!
I used megaage asian to train model, and use my images to test. My images was detected faces with MTCNN and saved with margin! When I use the trained model to test the gender with MTCNN detecting face again, and the results bad, but use images with margin (don't detect face with MTCNN), the results very good!
Whether it is related to the training set containing boundaries？
Is it reasonable for me to test like this?
Thank you!

Do you use the alignment face?

@shamangary In your demo code ,you dont use the alignment face ,Why?

Positive mean errors for ages >50yrs

I have trained the SSR-Net on my private dataset and the best model has MAE of ~4yrs
Apart from heteroscedasticity, while the mean MAE for smaller ages is zero, for ages > 50 I am seeing large positive mean of errors. Is there any way to resolve this?

预测值没有年龄15岁以下的

拿模型predict的时候，发现小孩儿and 婴儿，全部预测15+岁。我看那个megaage-asian数据集是0-70,所以为什么我预测没有小孩那一段的，是模型本身的局限还是我哪里没搞对？？

检测的蓝色的框外面的红框有特殊的含义吗？

检测的蓝色的框外面的红框有特殊的含义吗？

Question about results

How accurate is your model with predicting older people (60+)? Does it predict them 95% always younger than their real age?

What are `lambda_local` and `lambda_d` before $\eta$ and $\Delta_k$ respectively?

Hello. In paper 3.3 Dymamic Range part,
formula (4)
$$\bar s_k = s_k (1 + \Delta_k) \tag{4}$$
formula (6)
$$\bar i = i + \eta_i^{(k)} \tag{6}$$

but there is sort of different in the code, the $\Delta_k$ and $\eta$ are multiplied by lambda_local and lambda_d respectively which (lambda_local and lambda_d) are assigned [0, 0.25, 0.5, 0.75, 1] in the SSRNET_train.py.

Why are they(lambda_local and lambda_d) exist?
Thank you.

could TYY_demo_mtcnn.py load age and gender model at the same time?

could TYY_demo_mtcnn.py load age and gender model at the same time?

memory leak?

Hello, I am running the program on 1080Ti, but the memory is very large, about 10 G, is it a memory leak? Does this happen to you? thank you very much