Adding prostests analyser

[Smoltbob]

I added a folder and tried to replicate the structure of KerasPretrained. There is also a temporary script test.py that allows to test the model outside of mtriage.

[Smoltbob]

This is built using the trained model from the authors of the following repository. https://github.com/wondonghyeon/protest-detection-violence-estimation

[breezykermo]

Hi @Smoltbob, sorry for the delay. One basic update that is needed here to get the build passing is to lint the code using black. If you've pip installed the outer requirements.txt, there's a script to easily do this for all relevant files:

sh scripts/lint.sh

[Smoltbob]

Thank you,

• I wrote this prototype requirement.txt and forgot to make it proper, sorry!
• I'll re-lint the script, I had done it at some point but I must not have redone it before pushing

[breezykermo]

Fantastic, thanks. I think this build may fail regardless due to something I need to fix in the core mtriage runtime. I'm going to post a long explanation here now about how you can test ProtestsPretrained on real data with a custom build.

[breezykermo]

Downloading some example data

I'm going to create an mtriage config that selects some video frames from Youtube, so that we can run the ProtestsPretrained analyser on them and simulate a real-world mtriage run. Running a generic search for "tear gas + mexico" in a span of 15 days gives me ~460 results, which running on my 4 CPU computer takes about 15 minutes to download using mtriage, and amounts to ~1.5gb of videos (mtriage downloads them all very low res by default).

If you don't have enough space or want to try on a smaller sample, just tweak the uploaded_after parameter to narrow down the range of dates.

I'm going to split my config into two parts, so that I don't have to redownload videos on every run while I'm debugging the ProtestsPretrained analyser.

Let's put this is 'data/download_vids.yml':

folder: media/test_protests
select:
  name: Youtube
  config:
    search_term: tear gas + mexico
    uploaded_before: "2018-11-30T00:00:00Z"
    uploaded_after: "2018-11-15T00:00:00Z"
analyse:
  - name: Frames

Make sure you've followed the youtube selector setup properly as well.

Now I can run this config using the regular build of mtriage (to do Youtube selecting and Frames) with the following command.

./mtriage run data/download_vids.yml --dev

This will download all the videos, and split them up into frames.

Testing ProtestsPretrained with a custom build

Create a 'whitelist.txt' file with the following lines:

ProtestsPretrained

This will allow us to build a stripped down version of mtriage with only the core library, and the dependencies for this component in development.

We can now build that with this command:

./mtriage dev build --tag protests --whitelist whitelist.txt

Now I have a build of mtriage with the ProtestsPretrained dependencies that I can experimentally run and debug.

When we run mtriage with this analyser, we want to use our custom build, to ensure that all of the dependencies specified in the ProtestsPretrained analyser are sufficient and correct:

Let's create a yml that operates on the data we downloaded in the previous step in 'data/test_protests.yml':

folder: media/test_protests
elements_in:
  - Youtube/Frames
analyse:
  - name: ProtestsPretrained
    config:
      labels:
        - protest

Now we can test a build of mtriage that only has the ProtestsPretrained requirements installed with the following command:

./mtriage run data/test_protests.yml --tag protests --dev

[breezykermo]

I've walked through this setup, and I think at least more dependencies- pandas and numpy- are required.

[Smoltbob]

Thank you ! Let's test all this then ! Yes there is a numpy dependency, I hadn't included it thinking it was standard. I'll track the pandas, I think I can remove this dependency.

[Smoltbob]

I've tested on a full run with no issues!

[breezykermo]

Awesome thanks. Sorry for the slow pickup here. It's in the works

[breezykermo]

First minor thing that is my fault: could you rebase these changes on top of the latest release? There's a small fix to the build process that I think you found your way around there.

[breezykermo]

Just running a clean build and set of tests locally to confirm this all works as expected, then will merge if all goes well. Thanks again!

[breezykermo]

@Smoltbob how long would you expect inference to take on a CPU? I'm currently testing with ~10,000 frames on Intel I7 and nothing is produced yet after 20 minutes of running.

Realising if inference for one frame takes closer to 0.5s than 0.1s this could be running for a while...

[Smoltbob]

Hi, so sorry for the delay! It takes 0.15sec on my i5-5200U, so that should be ok I think ? What I did notice when testing (including KerasPretrained) is that CPU in parallel doesn't seem to work. Could that be related ?

[breezykermo]

Ah I see. Yes I think you're right- I'll investigate the CPU in parallel problem.

[breezykermo]

Ah sorry- one other thing I just realised. Could you please PR to the 'release' branch? That way I can debug things before the changes flush through to master if something is up. Thanks so much, I'll merge as soon as the tests pass on the PR to release!