Notebook: deepface

[msramalho]

Tool: https://github.com/serengil/deepface Goals:

• explain what the tool does and it's caveats (memory requirements?)
• install and run a very simple example of the tool: comparing 2 faces
  • instructions on how to upload are needed. This is easy in Colab, but will it also work in Binder?
• Given a folder of images find all faces and then find all similar faces
• Run the analyze method on an uploaded face to get predicted face features like age

The tool has a python client but also a command line interface (see the bottom of the readme), ideally the CLI should be given priority but if there are limitations python should be used. Optionally both: if there's an educational value to that.

[amithr]

Hi Miguel,

Based on my research, In order to analyze a folder of images using DeepFace, they each need to be vectorized and then pickled. Then, we can compare a single image against this collection of pickled images (a database?)

The process is significantly more complex than what we've included in Notebooks before. If you access this link, the process is outlined in the section "The hacker's way."

Is this still something we should include?

Or, alternatively, is there a more straightforward way of doing this?

[msramalho]

Hi Amith,

Indeed this will be a more technically challenge one.

Based on this readme code (similarity section) I assumed the library could take care of that, I imagine it may be slower the first time around if it vectorizes things, but this would prevent the need to have all that code in the notebook, assuming it works:

#face recognition
dfs = DeepFace.find(img_path = "img1.jpg", 
          db_path = "C:/workspace/my_db", 
          distance_metric = metrics[2]
)

At some point it may reach a case where the code is not easy for someone who does not understand it, but it they can still specify the input parameters then it will still be impactful.

A possible path to explore here is having a section that describes how to use google drive connection

[amithr]

Ok, I understand.

Based on the article I read, I made the incorrect assumption that all the images would have to be vectorized and stored in a database manually.

[msramalho]

I'm glad there's an easier way to it!

[amithr]

Just a quick update: I had to take a quick break to get started on a freelance project I'm working on. I'll be back to working on this and I hope to get the deepface Notebook done by this next weekend.

[msramalho]

Thanks for letting me know!

[amithr]

Hi Miguel,

I don't have a lot of experience in Binder, but is there any way to host and run a Jupyter Notebook without it already being in a Github repository? Or is that kind of the whole point of Binder?

[msramalho]

I think that's the point of it, they have a few alternatives but all in the same "link your public code" mindset

[amithr]

That being said, would it be necessary to include instructions for uploading to Binder? I suppose I could see users using the notebook directly from this repo (once it's public). However, the instructions for uploading to Google Colab were fairly long and it might be a little inefficient for half the notebook to be instructions for uploading files.

[amithr]

Hi Miguel,

I'm kind of at a stage where I'm not really sure what else to add, particularly to the FAQ - I could really use some feedback on the notebook as a whole. Here's a link.

[msramalho]

Hi Amith,

I've had a look and it's looking good, I feel you should provide with some example files for people to easily download, I can host them on google drive. I've also added a "research note" about the limitations of using face comparison methods blindly. There's an empty Binder section, is that intentional?

A suggestion, though it's specific for colab, is showing how to easily receive locally uploaded files: from google.colab import files

uploaded = files.upload()

# this part being optional, maybe just calling `files.upload()` is enough.
for fn in uploaded.keys():
  print('User uploaded file "{name}" with length {length} bytes'.format(
      name=fn, length=len(uploaded[fn])))

in the FAQ maybe point to a tool or two where people can convert images from one format to jpg for example. one such tool is ffmpeg, we could do a notebook about that one too since I've found it to be useful for converting files, extracting frames etc. :smile: ffmpeg -i input.tif output.jpg (but let's hold on that too)

[amithr]

Hi Miguel,

Thanks for the feedback. I'll address all of this.

Do you think that the whole section explaining how to upload images to Google Drive is necessary? I'm not really sure why I did it that when it seems that users could simply upload their images locally (within the Notebook's own allocated storage).

[msramalho]

That's a fair point that also eluded me, also learning! If this way works across notebook environments then it's preferable.

[msramalho]

It seems I didn't have permission to modify the notebook so my changes didn't apply.

I added

- **Important research note**: Comparing faces is a a task that leads to false positives: face of individual X matches Y even though they are different people, so it should only be used to verify cases where there's additional reasons to believe that X and Y are the same person, so take it with a grain of salt and not apply the tool blindly. You can also use other tools to validate a positive result, such as [AWS Rekogintion](https://aws.amazon.com/getting-started/hands-on/detect-analyze-compare-faces-rekognition/). 

under the limitations

[amithr]

I added your text and also changed the limitations so that you should have edit permissions.

[amithr]

Hi Miguel,

I completed the following tasks:

• I added a section on converting images between file formats using the pillow (PIL) library to the FAQ.
• I added your code snippet for uploading images but did so in the FAQ. I think that for non-technical users using Google Colab, the GUI-based approach involving Google Drive might be more straightforward? Let me know what you think.
• I created a folder consisting of images that would match Angelina Jolie along with 1 that doesn't. Here's a link.

[amithr]

Do you think there are any other significant changes to make?

[msramalho]

Hi Amith sorry for the delay in feedback, the notebook looks good to commit. Thanks for including both upload options. I think, depending on where people run the notebook, there will always be some way or another of achieving the file upload.