rezoo / illustration2vec Goto Github PK

A simple deep learning library for estimating a set of tags and extracting semantic feature vectors from given illustrations.

License: MIT License

Python 94.90% Shell 5.10%

illustration2vec's People

Contributors

Stargazers

Watchers

Forkers

joyhuang-machine-learning xuanhan863 lebronyxm lai-bluejay ml-lab boodoo hit-computer barneyeldinosaurio jsonbao piperun nieled wlads rmmilewi hatsunearu yasumo nagyist anubhav722 codeaudit hunny1389 aparna-b varun-suresh tamanobi ghostviper kktsubota blacktanktop mitliucak mangaki dreadlord1984 samuel2015 naafx8 cyxuanwater mosmeh tongsong91 ktolstikhin johndpope moeoverflow calculatedcontent kixiang halkliff xiangchenchao akitchin fo40225 risusanto paul0m katsuya-ishiyama forsskieken baifengbai togheppi cirq phdkiran tracycuiq alexvonduar tapdo shafiahmed rachmadaniharyono anirband afcarl hellopain kentashoji wilkinson1905 netdex auserj hiroshiba abbyyangbb ewrfcas lychees miracleyoo twtravel faruba manal-almodala justinlion83 hierifer carlosx7 kurodenjiro hakanaku1234 littlebuzi yw0nam frdid kivantium nat-chan thangmai sean16sysu xinwang-hnu savras takayama-opu 6gsn kiminh msksk cceyda stephy briangrant-data jjshisajia cstichbury ymzlygw zhy921375382 ototoi ssafeluck shaozhenged teufel1516 5l1v3r1

illustration2vec's Issues

Models missing

curl -vvvv http://illustration2vec.net/models/illust2vec_tag.prototxt
*   Trying 130.34.54.33...
* Connected to illustration2vec.net (130.34.54.33) port 80 (#0)
> GET /models/illust2vec_tag.prototxt HTTP/1.1
> Host: illustration2vec.net
> User-Agent: curl/7.48.0
> Accept: */*
>
< HTTP/1.1 502 Proxy Error
< Date: Sun, 01 May 2016 23:18:58 GMT
< Server: Apache/2.4.7 (Ubuntu)
< Content-Length: 524
< Content-Type: text/html; charset=iso-8859-1
<
<!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML 2.0//EN">
<html><head>
<title>502 Proxy Error</title>
</head><body>
<h1>Proxy Error</h1>
<p>The proxy server received an invalid
response from an upstream server.<br />
The proxy server could not handle the request <em><a href="/models/illust2vec_tag.prototxt">GET&nbsp;/models/illust2vec_tag.prototxt</a></em>.<p>
Reason: <strong>Error reading from remote server</strong></p></p>
<hr>
<address>Apache/2.4.7 (Ubuntu) Server at illustration2vec.net Port 80</address>
</body></html>
* Connection #0 to host illustration2vec.net left intact

Dockerized version available

Hello, and congratulation for this software, it's amazing!

I've created a ready to use Docker image containing all the dependencies, the software and the models, hope you find it useful.

It's at https://github.com/jacopofar/illustration2vec

Model design benchmarks

Is it possible to compare the VGG model in the paper with other Architectures like Inception, ResNet, NASNet or others? Photos for reference

ValueError: train argument is not supported anymore. Use chainer.using_config

When I run the end sentence of the example1, this error occurs, and I don't know how to fix it.

Create your own models

Is there a way to train your own models so that someone can add new tags?

I am trying to visit the demo website, but the site is inaccessible and returns 'The server is temporarily unable to service your request due to maintenance downtime or capacity problems'.
Is there anything wrong with the site?

pypi

hi, i accidentally upload the repo to pypi (instead of test.pypi) https://pypi.org/project/illustration2vec/

i thought i will delete it, but after some thought i will keep it if anybody need it or remove it when rezoo don't want it to be uploaded.

the version that i uploaded include

#14 from @Katsuya-Ishiyama
a cli command, untested and not documented
required package using chainer instead caffe
project data that i take from repo and rezoo github page
e: and the wrong version. should be something like 2.0.1
e2: broken image link

Checksums to verify downloaded models

Hello. Is it possible to place checksums for models from illustration2vec.net somewhere on the site?
The models are up to ~900 Mb, and redownloading them for verification would be a waste of time and traffic.

MD5 or SHA-1, SHA-256 etc.
Thanks in advance.

illustration2vec website down

I can't fetch the models anymore!

ImportError: i2v requires caffe or chainer package

I am trying to run sample code after downloading the models. I have installed all the dependencies, but i always get the error ImportError: i2v requires caffe or chainer package. It have tried the same on win 10, ubuntu 16.04 and docker.

Why does the probability of the label of the same picture change?

I found that every time I estimate the label probability of the same image, the result is different.

That is, for the same picture (for example, the picture you provided miku.jpg), I ran the program twice and the results were

Is this phenomenon normal? And why this phenomenon occurs? (maybe I was not careful enough and did not find the random function in the program)

Thanks for this project.

Train network

A basic view of how the network was originally trained would be very helpful. I have a larger dataset than was originally used and I'm hoping to evaluate your network and compare the results to Google's Inception network.

Without knowing what kinds of preprocessing were used (this can be partially inferred from the code provided) and how exactly the network was trained, this is difficult though.

Adding the solver prototxt that was used to train may be enough.

KeyError: 'encode1'

I can run the estimate_plausible_tags function just fine- the model seems to be working, but calling extract_feature or extract_binary_feature returns KeyError: 'encode1'

looking in to it a little, it seems the error was generated as the output of :
feature = self._extract(imgs, layername='encode1')
and in the illust2vec model, the layers are only iterations of convX_Y, reluX_Y, and poolX.
Passing actual layer names doesn't seem to get rid of the error or produce any results so I'm lost with this...
On windows, chainer is v5.2

$DIR with spaces for get_models.sh

You should separate $DIR with quotes in 5 row of get_models.sh so it can work when $DIR contains spaces.

Some advice about license compliance

Hello, such a nice repository benefits me a lot and so kind of you to make it open source!

Question
There’s some possible legal issues on the license of your repository when you combine numerous third-party packages.
For instance, numpy and scipy you imported are licensed with BSD License and BSD License, respectively.
However, the MIT License of your repository are less strict than above package licenses, which has violated the whole license compatibility in your repository and may bring legal and financial risks.

Advice
You can select another proper license for your repository, or write a custom license with license exception if some license terms couldn’t be summed up consistently.

Best wishes!

Pre-processing images

Before using the Illustration2Vec model on a dataset of images, what sort of pre-processing should I do to match your methods? In the appendix, it seems you subtracted the mean of the images. Is that it? Did you scale the data to fit in [-1, 1], for example?

big CPU cost

when i import i2v, the CPU cost climbs up to 10G slowly and then drops to 2G.
Any ideas about what happens?
Can this be related to GPU? how to use GPU in i2v?

KeyError: 'conv6_4'

Hi,

On Mac, I downloaded the models, but I can't load them. Seems like the loaded model does not have any layers, even if I use an old version of chainer (< 2.0).

Is it a Mac issue?