[RELEASE] bravo.exe 0.2.3 RC2 | Automatic Captions and Censors

[theredreaper]

EDIT: Download link has been updated with the latest bug fixes/feature requests: Current version 0.2.3 RC2

As promised, here is my re-release of the beta.exe program recently released here by bonsec. All the code involved is original, as I started from scratch, and took a bit of a different approach to reproduce the functionality without the security issue, and took some liberties with various features along the way. This is more or less "feature-complete" in terms of doing the same thing as the original, with some notable differences.

1. It's significantly more lightweight, 90Mb vs 193Mb. (There is a pre-trained AI model for doing the censoring bit distributed with the application, so the download is larger, but the app itself is 90Mb) I think I can reduce this further with some more effort, but I wanted to get a release candidate version out there and see what if any user feedback I receive before really going in depth with trimming the fat.

2. No web server. It runs in a GUI on the device you launch it from.

3. Only displays one image at a time. Two images side by side with captions was just a little too busy for my tastes.

4. Everything is user configurable. You provide your own images, .jpeg, .jpg, .png and .gif are currently supported, though .gif hasn't been tested now that I think of it. Simply point the program to a directory containing images from the Load Images button in the UI. You provide your own captions through a simple text file, which you point the program at from the Load Captions button in the UI. More on this text file in a bit. The censoring can be enabled/disabled at will with a simple checkbox in the UI, appropriately labeled "Censoring."

5. Additional features:
- As mentioned, you provide your own captions now. I have provided an example 'captions.txt' file in the archive with the program, but it's not going to be everyone's cup of tea. It is included just for demonstration purposes. The captions will be selected at random from the file, and you should put each separate caption on a new line. You've got a good bit of space to work with if you want to get wordy, with one limitation...
- The program can you command you to edge, through "specially coded" captions in the captions file you provide. If you want a caption to trigger an edge, simply place a ! at the beginning of the line. Super special and tremendously difficult, right! For example:

Code: Select all

!Edge, beta boy!

will trigger the program to stop automatically changing slides and wait for you to edge. To signal the program that you're ready to continue, simply left-click the mouse anywhere inside the program window, at which point it will move on to the next slide. When the program is waiting for you to edge, the slide will not change and you will see the message "Click anywhere once you've edged" appear on screen to let you know the program is awaiting your input. For now, this messages appears in the middle of the "text area" on the right side of the screen, which could get stacked under the actual caption if the caption itself is fairly long-winded. Let me know if anyone finds this problematic.
- The amount of time the program waits before forwarding the slide is determined by the length of the caption, if it's not waiting for you to edge. I wouldn't mind getting some feedback on whether the time it allots is too much, not enough, etc.
- There is a counter up in the right hand corner that will show how many edges you have done so far in this session. It was a feature I wanted.
- Besides your photos and the captions file, there is one other file of importance. A 'config.cfg' file will be created the first time you exit the software, which simply records your selected photos directory, the path to your captions file and whether or not you have censoring enabled so that these values can be re-loaded the next time you start the software. If something in this file gets screwed up and you see errors related to it, you should be able to simply delete the file, re-run the program and it will be re-created the next time you exit from the program.

The program, along with an example captions file, can be downloaded from this link:
Edit: link updated for v0.2.3 release.
https://drive.google.com/file/d/1RNzqiX ... sp=sharing

It's not particularly aesthetically pleasing, but I'm a software engineer, not a UX/UI design guy. Making stuff pretty is not my strong suit. I may do some work to make it a little nicer looking in the near future.

I will most likely open-source this, but I need to investigate all the licenses on the various libraries I leveraged in the process of creating this to be certain which open-source license to release it under. Until that's sorted out, consider this an alpha release and the source closed at the present time. If it doesn't get much interest, I probably won't bother releasing the source. Alternatively, if there is a decent amount of interest and it doesn't prove to be a huge pain in the ass to maintain, I may go ahead and keep it to myself just to ensure that new features and whatnot follow the conventions and respect the design of the core. This simply results in cleaner code and more reliability over the long-term because no two programmers think alike, and it is quite likely and in fact, generally expected that any given programmer will quickly find themselves lost and confused in another programmer's code.

If you use the software and find it enjoyable and feel like it's something you will keep around, please drop a comment so I know whether or not there is value in putting any additional time into this.

Credit goes to bonsec from these forums for the original concept, any resemblance to his beta.exe software is roughly intentional, though I have departed from the original design, I believe the core concept is still intact, so he should be credited for that.

[Augustulus]

Hey ,

well, so far it looks pretty decent, BUT as soon as I want to censor the images I get the following error:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 229, in <module>
  File "main.py", line 151, in next_slide
  File "main.py", line 170, in cens
  File "nudenet\detector.py", line 57, in __init__
  File "onnxruntime\capi\onnxruntime_inference_collection.py", line 335, in __init__
  File "onnxruntime\capi\onnxruntime_inference_collection.py", line 370, in _create_inference_session
onnxruntime.capi.onnxruntime_pybind11_state.InvalidProtobuf: [ONNXRuntimeError] : 7 : INVALID_PROTOBUF : Load model from <Systempath>.NudeNet/detector_v2_default_checkpoint.onnx failed:Protobuf parsing failed.

Any idea what's going on? Btw. I censored the systempath in the error message.

Additionally you get an error if you don't loaded any captions and start a slideshow:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 218, in <module>
FileNotFoundError: [Errno 2] No such file or directory: '' 

I haven't tested it, but I guess the same happens, if you don't load images.
I guess in general it would be wise to catch those "user-errors" with a warning, so that the program doesn't close with an error. Or even better only enable the start slideshow button after the user has loaded images and captions.

Augustulus

[theredreaper]

Any idea what's going on? Btw. I censored the systempath in the error message.

Absolutely. My stupid ass forgot to include the pre-trained neural network model with the download, so it doesn't know how to do the censoring part.

I guess in general it would be wise to catch those "user-errors" with a warning, so that the program doesn't close with an error. Or even better only enable the start slideshow button after the user has loaded images and captions.

You are quite correct. An oversight on my part not to catch and handle these errors. I will get this patched right now and upload a new version and update the first post with the new link. This is why we do alpha/beta testing. Thanks for your feedback! Look for the new version here in a few minutes.

EDIT: Actually it looks like I was attempting to check to make sure those variables were populated, but ended up checking if they existed, rather than if they held values, because not enough coffee or something. I've fixed this one now. Additionally, the 'Start Show' button is disabled entirely if either an image directory or a caption file is not selected. Also, the button now toggles between 'Start Show' and 'Stop Show' respective to the state of the slideshow itself. Thank you for the suggestion about disabling the button, I agree it's a good idea.

[Augustulus]

Absolutely. My stupid ass forgot to include the pre-trained neural network model with the download, so it doesn't know how to do the censoring part.

Nice, forgetting about one of the most important parts. Well, I guess the small size of the project is now irrelevant with those big pre-trained net included.
It's working for me now after I copied the net to the correct location. Because at least under Windows 10 the program searches for the net inside a system path and not the application path. I as a computer scientist myself don't have a problem fixing this, but I guess you should handle this also for all the non experienced users .

You are quite correct. An oversight on my part not to catch and handle these errors. I will get this patched right now and upload a new version and update the first post with the new link. This is why we do alpha/beta testing. Thanks for your feedback! Look for the new version here in a few minutes.

EDIT: Actually it looks like I was attempting to check to make sure those variables were populated, but ended up checking if they existed, rather than if they held values, because not enough coffee or something. I've fixed this one now. Additionally, the 'Start Show' button is disabled entirely if either an image directory or a caption file is not selected. Also, the button now toggles between 'Start Show' and 'Stop Show' respective to the state of the slideshow itself. Thank you for the suggestion about disabling the button, I agree it's a good idea.

Great! This is pretty neat now.

One question about exiting the program while in a slideshow: Does the program wait for the next image to be loaded before really exiting? Because I saw it one time that the program froze for a couple of seconds before finally closing. I wonder if this has something to do with waiting for some background task to be finished. If so, maybe you can cancel the execution?

Also I found out that you get an index out of range error once you reach the last image of the slideshow:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 248, in <module>
  File "main.py", line 155, in next_slide
IndexError: list index out of range

And lastly some features that would be really nice to have:

1. Variable speed for the slideshow. In my opinion it is a little bit too fast, but I also know that there are times when I want it to go even faster.
2. Options on what to censor and how. Nudenet has parameters for this, so it would be nice to be able to manipulate them to your personal preferences.

Also I have to say that the censoring AI is as horrible as ever and only censoring around 40-60% of the things it should (results differ on the image sets you use, mine are not really suited for the AI as it looks). Doesn't look as if they made much progress improving it over the last year. Well, that's of course something out of your scope .

Edit: I tried with gifs and it is absolutely not working at the moment. You immediately get an error. I've send you a pm with the details .

[theredreaper]

It's working for me now after I copied the net to the correct location. Because at least under Windows 10 the program searches for the net inside a system path and not the application path. I as a computer scientist myself don't have a problem fixing this, but I guess you should handle this also for all the non experienced users .

That shouldn't be necessary, I actually built in logic to have it install those files in the correct system path on the first execution of the program if it doesn't detect that directory. You didn't by chance have an existing .Nudenet directory at ~/.Nudenet did you? That would have caused it to fail the check I'm making and not copy the files out. Barring that, permissions problem, maybe?

I ran into the list index out of range myself tonight and have already patched it. Another obvious one that I should have thought of before.

You are correct in that it's fairly trivial to implement the censoring options. I just didn't do it because I wasn't sure if it was a feature that people were actually going to want. Look for it to be in the next release now.

As far as the effectiveness of the model goes, part of the reason for the large size is that I'm using a model that is supposedly trained better than the default ones distributed by the library author. It's over 8 times the file size, so it would seem like there ought to be something to the claim. In my own testing, it did seem marginally better than the default, but yeah, it still missed a lot. Not much I can do about that one just yet, but I have thought about taking a fresh model and running my own set of training data against it and seeing if I can get something that performs better at least on the pictures that I use. If you're in the field, you can probably handle doing the same thing for your own dataset.

One question about exiting the program while in a slideshow: Does the program wait for the next image to be loaded before really exiting? Because I saw it one time that the program froze for a couple of seconds before finally closing. I wonder if this has something to do with waiting for some background task to be finished. If so, maybe you can cancel the execution?

Well the advancement of the slideshow is handled within the main loop of the program, which at this point, probably could do with having some fat trimmed off of it. The call to advance the slide is the 2nd step of the loop, with the first one being the break that happens when you click exit. I could quash the behavior you're seeing by moving the break step to the end of the loop. What you're running into is basically the break step had been passed by and the image advancement had already been executed, and during the rest of the loop running, your button press event was registered, which couldn't execute until it hit that step in the loop on the next iteration.

[Augustulus]

That shouldn't be necessary, I actually built in logic to have it install those files in the correct system path on the first execution of the program if it doesn't detect that directory. You didn't by chance have an existing .Nudenet directory at ~/.Nudenet did you? That would have caused it to fail the check I'm making and not copy the files out. Barring that, permissions problem, maybe?

I definitely had already a .nudenet directory, but with an empty net in it. So, this explains the problem. Maybe it was created as I first ran the software? Maybe it's still from an old nudenet installation. I don't know exactly.

I ran into the list index out of range myself tonight and have already patched it. Another obvious one that I should have thought of before.

You are correct in that it's fairly trivial to implement the censoring options. I just didn't do it because I wasn't sure if it was a feature that people were actually going to want. Look for it to be in the next release now.

Both is great !

As far as the effectiveness of the model goes, part of the reason for the large size is that I'm using a model that is supposedly trained better than the default ones distributed by the library author. It's over 8 times the file size, so it would seem like there ought to be something to the claim. In my own testing, it did seem marginally better than the default, but yeah, it still missed a lot. Not much I can do about that one just yet, but I have thought about taking a fresh model and running my own set of training data against it and seeing if I can get something that performs better at least on the pictures that I use. If you're in the field, you can probably handle doing the same thing for your own dataset.

I could train it myself, but I don't think that I have the capacities to run a training besides my usual duties.
And as I already said: that's basically out of your scope. It was just an interesting discovery I had made .

Interesting that it already performs even better than the provided one. As far as I can see it's indeed working great on obvious parts, but as soon as there is something slightly hidden it fails horribly. Same goes for slightly angled views in my opinion.

Well the advancement of the slideshow is handled within the main loop of the program, which at this point, probably could do with having some fat trimmed off of it. The call to advance the slide is the 2nd step of the loop, with the first one being the break that happens when you click exit. I could quash the behavior you're seeing by moving the break step to the end of the loop. What you're running into is basically the break step had been passed by and the image advancement had already been executed, and during the rest of the loop running, your button press event was registered, which couldn't execute until it hit that step in the loop on the next iteration.

Yeah, exactly what I thought of. I don't think that it makes some improvement to change the position. It just shifts the problem from a certain timestep to another.
As long as there's no immediate interrupt, I don't think it is worth spending time on this.
Personally I hate using python, so my knowledge about the language is only rudimentary, thus I can't be more precise about solutions to this .

[theredreaper]

the issue is partially due to me being lazy and wanting to get this out rather quickly due to glaring security issues with its predecessor. i'm using a gui library that is quite simple to mock something up in, but it is itself another layer on top of tkinter, which I believe is listening for click events through the windows API, so an event has to propagate up through the windows API, through tkinter, and then through the gui framework I'm using that effectively layers on top of tkinter, so it's not the most efficient thing in the world in terms of latency, but it is quite efficient in terms of saving MY time, lol. this entire program, not including the libraries it links from the python standard libraries or third party open source libraries, is 118 lines of code. And it fully adheres to all of python's best practice conventions, like 2 blank lines before and after every function, multiple lines of comments in every function to document it's parameters and returns, etc. If i cut out every line that isn't actually doing something, it's probably more like 60-70 lines of code. That's why I chose to use a GUI framework that is kind of slow and has an odd way of handling its event loop. I could probably hook the Windows API directly to intercept the click event on exit if I was really determined to eliminate that small bit of lag, but it's enough for me to know that the problem could be solved, so I will stop short of actually putting the time into it that it would require. :) I have also encountered a strange bit of behavior occasionally on startup where the first slide will display for less than a second before changing, and I haven't quite figured that one out just yet. It's similarly easy to just overlook, but it does annoy me because I can't fathom what could be causing it.

[basim]

I've the same problem with the censoring. I don't know if it is from my very first test of your first version.

I delete the nudenet folder under user/myuser and all the files of the Zip file. Than extracted the latest Zip once again and started it. A nudenet folder got created under my user again. But the file in it is 0 KB. The nudenet "plugin" then tries to reach out for a server.

Is there any other saved data that I have to delete first?

*Edit

Manually copying the detector & classes file over to the nudenet folder does help.

*Edit 2.0

Btw. it doesn't save the last choosen folder path and caption file. Even after running a session successfully and exit the program by the button. After loading it again I have to select both again. You mentioned that it should remember that..

[theredreaper]

I've the same problem with the censoring. I don't know if it is from my very first test of your first version.

I delete the nudenet folder under user/myuser and all the files of the Zip file. Than extracted the latest Zip once again and started it. A nudenet folder got created under my user again. But the file in it is 0 KB. The nudenet "plugin" then tries to reach out for a server.

Is there any other saved data that I have to delete first?

*Edit

Manually copying the detector & classes file over to the nudenet folder does help.

*Edit 2.0

Btw. it doesn't save the last choosen folder path and caption file. Even after running a session successfully and exit the program by the button. After loading it again I have to select both again. You mentioned that it should remember that..

I will look into this. Copying the files into where they're supposed to be works on my system. If it doesn't find the model in the folder where it's looking for it, it will reach out to github and try to pull the default model from the repo of the Nudenet library. However, this file will cause a crash of the program due to the opencv version included. Verify that the detector model file is around 139Mb. The other file should just be a small text file called CLASSES.

As far as the settings not being saved, look in the folder where you're running bravo.exe from for a file called config.cfg, if it's there, open it up and see what it contains.

[basim]

I will look into this. Copying the files into where they're supposed to be works on my system. If it doesn't find the model in the folder where it's looking for it, it will reach out to github and try to pull the default model from the repo of the Nudenet library. However, this file will cause a crash of the program due to the opencv version included. Verify that the detector model file is around 139Mb. The other file should just be a small text file called CLASSES.

As far as the settings not being saved, look in the folder where you're running bravo.exe from for a file called config.cfg, if it's there, open it up and see what it contains.

Nope, there is no config file.

[Augustulus]

I will look into this. Copying the files into where they're supposed to be works on my system. If it doesn't find the model in the folder where it's looking for it, it will reach out to github and try to pull the default model from the repo of the Nudenet library. However, this file will cause a crash of the program due to the opencv version included. Verify that the detector model file is around 139Mb. The other file should just be a small text file called CLASSES.

As far as the settings not being saved, look in the folder where you're running bravo.exe from for a file called config.cfg, if it's there, open it up and see what it contains.

Nope, there is no config file.

I second this. No config file there for me, too

the issue is partially due to me being lazy and wanting to get this out rather quickly due to glaring security issues with its predecessor. i'm using a gui library that is quite simple to mock something up in, but it is itself another layer on top of tkinter, which I believe is listening for click events through the windows API, so an event has to propagate up through the windows API, through tkinter, and then through the gui framework I'm using that effectively layers on top of tkinter, so it's not the most efficient thing in the world in terms of latency, but it is quite efficient in terms of saving MY time, lol. this entire program, not including the libraries it links from the python standard libraries or third party open source libraries, is 118 lines of code. And it fully adheres to all of python's best practice conventions, like 2 blank lines before and after every function, multiple lines of comments in every function to document it's parameters and returns, etc. If i cut out every line that isn't actually doing something, it's probably more like 60-70 lines of code. That's why I chose to use a GUI framework that is kind of slow and has an odd way of handling its event loop. I could probably hook the Windows API directly to intercept the click event on exit if I was really determined to eliminate that small bit of lag, but it's enough for me to know that the problem could be solved, so I will stop short of actually putting the time into it that it would require. :) I have also encountered a strange bit of behavior occasionally on startup where the first slide will display for less than a second before changing, and I haven't quite figured that one out just yet. It's similarly easy to just overlook, but it does annoy me because I can't fathom what could be causing it.

Yeah, as I said: if it's too much time to spend on, then it's not worth looking into it.
Regarding the first slide being faster than the other ones: maybe the time of loading and displaying the first image does a lot longer than the others? I also think that nudenet needs some time to load the first time it is used.

[wqslave]

I applaud you and anyone making content and programs in this very niche space, thank you!

I am having many of the same issues reported here already (no config file, failure to execute the script main if I check the censored box). Tried the usual like turning off anti virus before running it, still no config.

But also wanted to mention a few things that make this different from, and still arguably not matching the features of, the hacky-beta.exe program:

- This does not seem to do video. The beta.exe program does video the best and fastest of any "easy" solution I've seen. IE this one is promising but I can't seem to get it to work: https://github.com/8TBHomework/AutoCensorMPV

- Captions outside of image body is imo less interesting. I like the sensory overload (of the dual images too tbh).

- Caption system is much less robust, needing every variation to be its own line. beta.exe spits out near gibberish if configured wrong, but I do like the dynamic nature of it.

- I haven't been able to see the censoring in action because crashes, but I'm guessing this does not do highlighting.

- It appears you can only load one directory at a time.

So at least for me, this is unfortunately not a substitute, but with the other program kinda hacky in the implementation and security that's a shame /: Excited to see how you develop this if you are able to spend more time on it!

My crash traceback if it's useful at all:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 248, in <module>
  File "main.py", line 162, in next_slide
  File "main.py", line 181, in cens
  File "nudenet\detector.py", line 57, in __init__
  File "onnxruntime\capi\onnxruntime_inference_collection.py", line 335, in __init__
  File "onnxruntime\capi\onnxruntime_inference_collection.py", line 370, in _create_inference_session
onnxruntime.capi.onnxruntime_pybind11_state.InvalidProtobuf: [ONNXRuntimeError] : 7 : INVALID_PROTOBUF : Load model from C:\Users\-----\.NudeNet/detector_v2_default_checkpoint.onnx failed:Protobuf parsing failed.

^ I was able to make the above stop crashing/occuring by deleting my nudenet directory, running it again (it crashed again,) then manually copying the included model into the new .nudenet directory as the created model was only 32 kb, but now it doesn't seem to actually censor anything, then gave a new error after a few slides:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 248, in <module>
  File "main.py", line 155, in next_slide
IndexError: list index out of range

[Augustulus]

My crash traceback if it's useful at all:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 248, in <module>
  File "main.py", line 162, in next_slide
  File "main.py", line 181, in cens
  File "nudenet\detector.py", line 57, in __init__
  File "onnxruntime\capi\onnxruntime_inference_collection.py", line 335, in __init__
  File "onnxruntime\capi\onnxruntime_inference_collection.py", line 370, in _create_inference_session
onnxruntime.capi.onnxruntime_pybind11_state.InvalidProtobuf: [ONNXRuntimeError] : 7 : INVALID_PROTOBUF : Load model from C:\Users\-----\.NudeNet/detector_v2_default_checkpoint.onnx failed:Protobuf parsing failed.

Great! I just wanted to mention you to copy the files by hand into this directory to get it working.

^ I was able to make the above stop crashing/occuring by deleting my nudenet directory, running it again (it crashed again,) then manually copying the included model into the .nudenet directory, but now it doesn't seem to actually censor anything, then gave a new error after a few slides:

Code: Select all

Traceback (most recent call last):
  File "main.py", line 248, in <module>
  File "main.py", line 155, in next_slide
IndexError: list index out of range

This error is also known already. You simply got to the end of your slideshow. Have you tested it with a directory containing only one or two images? If so try again with a directory containing more images to verify that the censoring is now working. Also do you have copied both files the detector_v2_default_checkpoint and the classes?
They are both required to make the censors work .

[Mr-Elusive]

Nice little piece of software. I haven't seen any of the other versions, but really cool idea.

[theredreaper]

Back with another update. Thanks to all of you for all of the feedback so far. Here are the patch notes for this release:

Code: Select all

0.1.2 RC1 - 
    - the application now ensures that the working directory is the same as the directory it is being run from.  this shouldn't be necessary, but I've been seeing some weird behavior when trying to create or access files that should be available from the application's directory.
    - addressed an issue where the slideshow could advance on it's own when the user has been instructed to edge instead of waiting for the user's input
    - addressed a small issue that could cause the very first slide of a new slideshow to advance too quickly
    - the application will now raise an exception if the config file can't be written
    - the application will now raise an exception if the detector model and classes file aren't installed correctly, and provide the user with instructions to get them properly installed.
    - the checking of image file extensions is no longer case sensitive
    - the application will no longer get caught in a loop of popping up error messages if the user somehow manages to start a slideshow and no images can be loaded from the specified folder for some reason.
    - the slideshow should now loop after playing the last image in the list.
    - support for .gif files has been dropped
    - a new 'Options' button has been added to the UI, which exposes some configurable slideshow options to the user

The reason this took awhile to get done was that this update added quite a lot of new logic, and therefore quite a lot of code to the project. It seems stable on my machine in its current form, and I wanted to get a release out with these bugfixes and the requested features so far, but I will be moving to an RC2 version very soon, as the code is now in dire need of some refactoring. Let me know how you all get by with this version.

https://drive.google.com/file/d/1A7i1Jx ... sp=sharing