check-face / checkface Goto Github PK

On small screens, the home button on the new layout for the grid template is weird.

Expected the buttons to flow to multiple rows.

Hi,

We are trying to run checkface locally and want to run the encoder4editing service locally as well.
Unfortunately the image, as referenced in your docker-compose.yaml is not available anymore on docker hub (olicoad/encoder4editing:latest). There are other images with the same name available but they don't seem to run an HTTP service and exit immediately after starting.

Any advice on how to overcome this problem?

Thanks in advance!

Per your docs, you mention an internal api for image encoding. Would the features described here be suitable?

Thanks!!!

Change implementation of the seed logic throughout the application so that we call some methods on a specific implementation of LatentProxy but for each of the individual seed guid and hash as per: https://checkface.ml/api#apiface

Using Travis CI free tier for open source, create a GitHub release with build artefacts including the MSI and other Windows installers, and the electron installer if we get it working

On the clients, the value is not trimmed, allowing for values such as "c " to be requested. However, on the website, it does not include trailing spaces in the url query parameter even though it makes the api request with the trailing spaces.

Current Behaviour

After typing in "c " in checkface.ml, the url is updated to https://checkface.ml/?value=c instead of https://checkface.ml/?value=c%20%20%20%20.

Expected behaviour

Either the worker trims the input so that both requests are the same, or the website should include trailing spaces in the query parameter.

When doing a morph from A to B using the /api/mp4/ or /api/gif/ endpoints, the resulting video does not start at A.

Expected behaviour

Starts at A, morphs to B, then morphs back to A (-1 frame)

Actual behaviour

Starts halfway between B and A, then goes to A, morphs to B then half way back towards A.

Example:
https://api.checkface.ml/api/mp4/?from_value=asdf&to_value=fdsa

In order to give better social media cards when linking to particular searches from the website, the meta tags og:image and og:title need to be server side rendered based on the value query parameter.

For example, https://checkface.ml/?value=qwert should have

<meta property="og:image" content="https://api.checkface.ml/api/qwert?dim=1000" />
<meta property="og:title" content="Check Face - qwert" />

At this point it would be good to keep it simple and could be done with a template engine such as php or https://github.com/giraffe-fsharp/Giraffe.DotLiquid

In preparation for people doing projects using the api, it is important to add easily accessible documentation for

• how to generate a face using seed/hashvalue/latents
• how to register latents
• how to make morphing gifs and mp4s

Expected:
Scaling by any dimension doesn't change the presence of GAN artefacts, and the image would look exactly the same at any dimension.

Actual result:

https://api.checkface.ml/api/6938e4ed47a23f17ceb6dbd880ab18d2787e7946 and https://api.checkface.ml/api/6938e4ed47a23f17ceb6dbd880ab18d2787e7946?dim=300 is different, as shown

The checkface by ip page needs a short description of what a checkface is and why it is useful for IPs.

Training StyleGAN on multiple GPUs requires Nccl, which is not included on windows.
There is some custom way of reducing and updating all of the gradients across the devices which is not similar to the api's exposed by tensorflow.

This causes an error like:

tensorflow.python.framework.errors_impl.InvalidArgumentError: No OpKernel was registered to support Op 'NcclAllReduce' used by node TrainD/SumAcrossGPUs/NcclAllReduce (defined at D:\data\oliver-train-checkface\fflowhq\00005-sgan-flower-1gpu\src\dnnlib\tflib\optimizer.py:135) with these attrs: [reduction="sum", shared_name="c124", T=DT_FLOAT, num_devices=2]

There is no drop in replacement that has been found, because the api for tf generic operations like a HierachicalAllReduce which is used in Keras like in: tensorflow/tensorflow#21470
is not compatible with the nccl_ops.py interface https://github.com/tensorflow/tensorflow/blob/r1.14/tensorflow/python/ops/nccl_ops.py

Perhaps even more surprising is the fact that other ops, like: collective_ops.py
https://github.com/tensorflow/tensorflow/blob/r1.14/tensorflow/python/ops/collective_ops.py
do not provide drop in replacements. These ops seem to have completely different use cases as is made clear by their use in tests:
https://github.com/tensorflow/tensorflow/blob/r1.14/tensorflow/python/ops/nccl_ops_test.py
https://github.com/tensorflow/tensorflow/blob/r1.14/tensorflow/python/ops/collective_ops_test.py

The line that needs to be updated or removed seems to be the following:

@olivercoad

Mobile doesn't have the github link

Also Nvidia and AWS etc aren't shown

As per the following pseudocode, implement the endpoint

# @app.route('/api/recoverlatent/cropface/', methods=['GET'])
# def recoverlatent(imgguid):
#     get image
#     return croppedimgguid

Make workers available at https://checkface.ml Optionally put multiple workers behind load balancer. To be done after GitLab#9 (closed)

from GitLab#10

We wish to change the current way of rendering social previews using php in the docs/index.php to use a lambda to do this for much cheaper. The costs aren't trivial, though they're not big either.

Additionally, an internal reverse proxy to our on-prem server would mean we can still have an autoscaling group on an ALB but not have to pay for the t2.micro instance running nginx currently.

This blog post has details on how to return images, which proved quite difficult: https://aws.amazon.com/tw/blogs/compute/microservices-without-the-servers/?sc_campaign=blog&sc_category=compute&sc_channel=SM&sc_content=serverless&sc_detail=std&sc_medium=aws&sc_publisher=fb_ln&adbsc=social_blogs_20150909_52150936&adbid=882627328441300&adbpl=fb&adbpr=153063591397681

It should ideally support auto-updating with Squirrel, and to register as a right click menu for files.

Currently, the pages load ~6MB of images, they only need to be 130px max.
Simply resize and commit smaller images in images, linked by index.html

One time got corrupted video when loading same mp4 from multiple places at the same time.
I'm speculating that the second load loaded the existing mp4 because the file existed even though it was only partly created.

It should probly wait to make sure the video file is actually finished generating before serving it.

Current Behavior:
When a question mark is included in the hash, the default dimension of 300 is induced

Expected behavior:
Question marks have no impact on size and the ?dim=500 argument persists regardless of textbox entry

Proposal:
Check how the client side box is sending ? chars and see if it's interfering with the GET request url and causing it to be malformed

In order to handle any user input correctly (see #4), it has to be URL encoded before placing in the URL.
For example, in javascript, encodeURIComponent("asdf ") returns asdf%20%20.
This generates the following face.

Unfortunately, that's not the only way spaces (and probably other things) can be URL encoded.
For example, an HTML form request or System.Web.HttpUtility.UrlEncode will encode spaces as + rather than %20: System.Web.HttpUtility.UrlEncode("asdf ") returns asdf++.
This generates the following face, noticably different to above.

Furthermore, it returns the same face for asdf%2B%2B as asdf++; %2B is the url encoding of +.

I expect the problem is to do with how flask automatically decodes parts of the path to pass as the hash parameter

Expected Behaviour:
The generated face for the url encodings as asdf%20%20 and asdf++ should both be the same (should both be the first image), corresponding to the user input of asdf .
The generated face for the url encoding as asdf%2B%2B should be different (it should be the second image), and correspond to the user input of asdf++.

Because people are human, some may think it is a good idea to put their passwords in for some reason.
Unfortunately, since this sends an api request to the server, which keeps logs, we could unintentionally end up with lists of passwords.

A solution could be to pre-hash all values on all clients before sending them to the server, so that no logs of user input is kept while still being able to cache requests.
This would mean that when checking a file, there are three hashing steps:

• SHA256 (or whatever method the user clicks) of file
• Send request using hash of that
• Server received request and takes hash of it to generate seed for the rng

The layout of the landing page isn't very good for mobile at the moment.

Ideally, one should be able to enter values to the text box, click submit, and view the entire face all without scrolling.

from GitLab#13

When using the back button to go to previous values, the preview image does not go back and the text in the search box does reflect the old value.

Note: this issue does not replicate on the very first search after a reload because that uses a full form request and subsequent searches use javascript and history api.

See #18 for context.

If a user inputs asdf%2B%2B *(not as the first search on that tab), it is, as expected, URL encoded as asdf%252B%252B and generates the following face.

However, if you navigate to https://checkface.ml/?value=asdf%252B%252B, it shows the input box and checkface for asdf++ instead, which normally has the url https://checkface.ml/?value=asdf%2B%2B.

*Although this is not caused by a18c70e, it is complicated as it makes the first search one does after navigating to checkface do a full page navigation (whereas subsequent requests are done with the original behaviour using javascript like a SPA).

The cloudflare worker is getting the face on all requests even on https://checkface.ml/ where no value is even specified.

When the value is not specified, the page should not be transformed.

The position of text in the popups and hover tips with the extension is inconsistent. They should all be above and centred, but are sometimes to the left of the image.

Hello! This is a really cool project. Thank you very much for making it open source. You mention in the README,

If you don't have a suitable GPU, you canrun the backend with an AWS p3 instance on ECS, or g3s.xlarge via docker-machine for testing.

I am trying to deploy this to an AWS instance, but am facing some issues. I'm looking at the following tutorial as a guide for deploying this. Would this be the correct strategy? Have you found that your AWS account limits needed to be lifted for running the necessary P3 or g3s.xlarge specs? Also, I'm curious what you've found to be the costs.

Apologies for all the questions, which are more related to devops and cloud than this repo specifically, but thank you very much in advance if you could provide any additional guidance. Thank you!!!

As GitHub is better for visibility, we may as well use it for active development and make use of issues here