paeti/ipneuronalenetze

Using neural networks for age and gender estimation.

ipneuronalenetze's Issues

write dataloader for IMDB-wiki dataset

write dataloader that loads the local dataset in a format the nn can use

usefull links:

• https://scikit-learn.org/stable/datasets/index.html

Add data and docs folder

I have created them before, but docker can't pull empty folders.
So place some empty files in there and it can be pulled.

Create docker image for the build test

Drone ci works with docker.
Each step of the pipeline creates a docker container.
So a container is needed which contains all dependencies for:

• tensorflow
• numpy
• mathplotlib
• opencv
• make

Check if VGG-16 is a good choice

I found that VGG-16 convolutional networks are good for age and gender estimation.
This needs to be checked.
research:

• which group of neural networks is best for our task
• which specific network would you choose
• are convolutional nn and VGG-16 a good choice

Link convolutional nn:
konvolutionsnetze
Link paper about VGG-16 for age estimation:
dex deep expection of apparent age from a single image
Link tutorial implementing VGG-16:
learning keras by implementing vgg from scratch

summarize the results and add them to our docs.
(Please use markdown)

prepare CACD dataset

preprocess the images

usefull links:

• http://bcsiriuschen.github.io/CARC/

Research on splitting our data and which strategy we should use to train

Research in how much parts we split our data, including validationset, trainingset, testset, crossvalidation
research on other strategies for training
http://cs231n.github.io/classification/

Create research_result_datasets.md. Document your results in there.

prepare MORPH dataset

preprocess the images

usefull links:

• https://ebill.uncw.edu/C20231_ustores/web/classic/store_main.jsp?STOREID=4
• https://www.garrettbingham.com/files/Cleaning%20MORPH-II%20Presentation.pdf
• http://www.faceaginggroup.com/morph/

Creating the projects Readme

For better contribution and a good overview the project should have a Readme.
Which contain an introduction to the project, contribution rules and an explanation how the infrastructure works.

Create Contributing.md

So interested people are able to know how to work in this project it should have a Contributing.md.
These file should contain:

• code styleguide
• git workflow
• git commit rules
• issue rules

Write MAKEFILE

For automated building and testing of a project a buildtool is needed.
A MAKEFILE is a good choice.
it should:

• built the whole project
• make all the test

write trainer script for training with IMDB-wiki (age model)

write trainer wich loads the imdb data, one set labeled with age, one with gender, and trains the models with it

Prepare dataset

Write methods to prepare the dataset

Write a method to extract the images out of the file, cut out the faces, scales the images and saves them in another folder.

Write a methode that splits the dataset in training, validation and test set.

Nice2Have:
Method that rotates the images and save multiple versions with differnet angles or in other ways creates a bit different versions off the picture

Build vgg-16

Build the vgg-16 convolutional nn.

Links for inspiration:
https://hackernoon.com/learning-keras-by-implementing-vgg16-from-scratch-d036733f2d5
https://keras.io/getting-started/sequential-model-guide/
https://keras.io/applications/

Fix Webcam View

check if the website is properly by the common browsers
check why thewebcam view isn't streamed

Write new build/test pipeline for drone

Now that the dockerfile is ready it can be used in building and testing.

• Write a new pipeline using the ippntensorflow image
• push the image to the docker hub or place it on the drone agent container

Add prototype webfrontend

To showcase our project we need some kind of frontend.
We decided to use a simple web ui
This prototype should be able to:

• connect from the browser to the camera
• take a picture
• send the picture via POST

write dataloader for CACD dataset

write dataloader that loads the local dataset in a format usable for nn

usefull links:

• https://scikit-learn.org/stable/datasets/index.html

prepare IMDB-wiki dataset

crop the faces out
combine picture with its related labels
bring it in a format that can easy be used

Shorten contributing link in Readme

Because the CONTRIBUTING.md is in the same folder you can shorten the Link in the README to CONTRIBUTING.md

prepare LAP dataset

preprocess the LAP dataset
this set is intended to be the set were the final test images are from
think about if you want to use the whole dataset for testing or just parts

usefull links:

• http://chalearnlap.cvc.uab.es/dataset/19/description/#

visualize the output of the convolution layers and filters

get the output of the convolution layers as picture so it can be seen what the filters extract from the original pictures

usefull links:

• https://github.com/InFoCusp/tf_cnnvis
• https://medium.com/@awjuliani/visualizing-neural-network-layer-activation-tensorflow-tutorial-d45f8bf7bbc4
• https://jacobgil.github.io/deeplearning/filter-visualizations
• https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html
• https://hackernoon.com/visualizing-parts-of-convolutional-neural-networks-using-keras-and-cats-5cc01b214e59
• keras-team/keras#4274

build model for age estimation

use classification
modify our model, so it predcicts age
use parameters from https://www.vision.ee.ethz.ch/publications/papers/articles/eth_biwi_01299.pdf

usefull links:

• https://keras.io/callbacks/

Expanded research on vgg-16/ split building nn in tasks

Expanded research on vgg-16, including convolution, pooling, filterkernel, vgg architectur
Think about in which task the coding could be split
http://cs231n.github.io/assignments2018/assignment2/
https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html
https://www.learnopencv.com/keras-tutorial-using-pre-trained-imagenet-models/
https://stackoverflow.com/questions/45383706/do-i-need-pretrained-weights-for-keras-vgg16
https://www.cs.toronto.edu/~frossard/post/vgg16/
http://agnesmustar.com/tag/vgg16/

Create research_results_extended_vgg_16.md. Document your results in there.

research on scikit learn (training)

get used to modell training with scikit learn
documentation not nessecerally needed, just experiment a bit

usefull links:

• wrapper class for keras modells:
  • https://keras.io/scikit-learn-api/
  • https://machinelearningmastery.com/use-keras-deep-learning-models-scikit-learn-python/
  • https://www.youtube.com/watch?v=AS1m2B86UbQ
  • https://www.youtube.com/watch?v=0AJECUm8gY0
• train model with scikitlearn:
  • https://scikit-learn.org/stable/tutorial/basic/tutorial.html
• use cross validation for evaluating models:
  • https://scikit-learn.org/stable/modules/model_evaluation.html
  • https://scikit-learn.org/stable/modules/cross_validation.html

Add bug and feature report templates

Add templates that help to create new issues
The feature template is used for al issus adding new stuff to the project.

write script to tune hyperparameters with scikit learn

use scikit to test and tune the hyperparameters of the nn's

usefull links:

• https://scikit-learn.org/stable/modules/grid_search.html
• https://www.youtube.com/watch?v=AS1m2B86UbQ

write dataloader for production

write dataloader that graps the image with given name out of the folder where the webcam images are safed

write test for FG-Net dataloader

usefull links:

• https://scikit-learn.org/stable/developers/tips.html
• https://scikit-learn.org/stable/developers/utilities.html

write test for LAP dataloader

usefull links:

• https://scikit-learn.org/stable/developers/tips.html
• https://scikit-learn.org/stable/developers/utilities.html

write script for cross validating the trained nets

write script that loads the validation dataset and uses it
to cross validate the trained models

usefull links:

https://scikit-learn.org/stable/modules/model_evaluation.html

write test for CACD dataloader

usefull links:

• https://scikit-learn.org/stable/developers/tips.html
• https://scikit-learn.org/stable/developers/utilities.html

prepare FG-Net dataset

preprocess the images

usefull links:

• http://yanweifu.github.io/FG_NET_data/

write dataloader for LAP dataset

write dataloader that loads the local dataset in a format usable for nn

usefull links:

• https://scikit-learn.org/stable/datasets/index.html

Create a small webserver with REST API

to connect to our frontend with tensorflow we should setup a small web api.
flask is a good framework for doing this.
the following steps are necessary:

• add a endpoint to deliver results from tf
• add a endpoint to fetch user data

train age model

use trainer method to train the model one the IMDB-wiki dataset
persist trained model

usefull links:

https://scikit-learn.org/stable/modules/model_persistence.html

write dataloader for MORPH dataset

write dataloader that loads the local dataset in a format usable for nn

usefull links:

• https://scikit-learn.org/stable/datasets/index.html

Fix Contributing link

In the Readme there is the wrong url in the contributing link.

research on scikit learn clustering

• research if the map format can be used for training in scikit learn
• research if/ how its possible to ignore some attributes of a dataset, otherwise we would need two preparations of the IMDB-wiki (one version with age and gender and two versions with just one of the attributes)
• research how to split datasets in scikit learn
• how to load your own dataset in scikit learn

usefull links:

• https://scikit-learn.org/stable/datasets/index.html
• https://medium.com/@contactsunny/how-to-split-your-dataset-to-train-and-test-datasets-using-scikit-learn-e7cf6eb5e0d
• https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html
• https://scikit-learn.org/stable/modules/preprocessing.html
• https://scikit-learn.org/stable/modules/classes.html#module-sklearn.feature_selection
• https://scikit-learn.org/stable/auto_examples/classification/plot_digits_classification.html

Get used to unit testing tensorflow and nn's

Unit testing tensorflow

inform you about how to test unit tensorflow, which libraries you need etc.
write a little test example(like for a test nn you programmed)

write tests for age model

write test methods

usefull links:

https://medium.com/@keeper6928/how-to-unit-test-machine-learning-code-57cf6fd81765

write dataloader for FG-Net dataset

write dataloader that loads the local dataset in a format usable for nn

usefull links:

• https://scikit-learn.org/stable/datasets/index.html

write test for MORPH dataloader

usefull links:

• https://scikit-learn.org/stable/developers/tips.html
• https://scikit-learn.org/stable/developers/utilities.html

build model gender estimation

use classification
modify our model, so it predcicts gender
use parameters from https://www.vision.ee.ethz.ch/publications/papers/articles/eth_biwi_01299.pdf

usefull links:

• https://keras.io/callbacks/

write trainer script for training with IMDB-wiki (gender model)

write trainer wich loads the imdb data and trains the model with it

Prototyp CI Test

Build small prototyp for ci build pipeline with drone.

write test for IMDB-wiki dataloader

usefull links:

• https://scikit-learn.org/stable/developers/tips.html
• https://scikit-learn.org/stable/developers/utilities.html

research how to split given data for classification

you could split with uniform age ranges or equally-distributed
https://www.vision.ee.ethz.ch/en/publications/papers/articles/eth_biwi_01299.pdf
here is written the last one gives better results
check if it can be done in our given time
you have to talk with the nn groups if they use regression or classification

usefull links:

• http://www.scikit-yb.org/en/develop/api/target/class_balance.html
• https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.StratifiedShuffleSplit.html
• https://imbalanced-learn.org/en/stable/index.html

train gender model

use trainer script to train the seperated models on their special version of the IMDB-wiki dataset
persist trained model

usefull links:

https://scikit-learn.org/stable/modules/model_persistence.html

write tests gender model

write test methods

usefull links:

https://medium.com/@keeper6928/how-to-unit-test-machine-learning-code-57cf6fd81765

write test for production dataloader

usefull links:

• https://scikit-learn.org/stable/developers/tips.html
• https://scikit-learn.org/stable/developers/utilities.html