lostminty/compx527_bag

COMPX527-21B A2:AWS CC App Service to detect suspicious behaviour on submitted CCTV footage

COMPX527_BAG (Behaviour Anomaly Group)

COMPX527-21B A2:AWS CC App Service to detect suspicious behaviour on submitted CCTV footage

Sample generation

Samples can be saved to file using predictor_src/create_json_sample.py

this takes 2-3 arguments

• location of dcsass dataset
• output file name
• optional 3rd parameter of the index of the sample to output

This script was used to generate the json string used in the demonstration of calling the API

Training on

• https://www.kaggle.com/mateohervas/dcsass-dataset (requires sign in) Has 13 classes and flags indicating a clip is of interest to the class or normal. so, 14 classes. could in theory look at it as 26 classes

Note:

for downloading the Dataset, the zip comes with a duplicate of the dataset.

Based on:

• https://github.com/PyTorchLightning/pytorch-lightning (pytorch helper module, automates a lot of 'boilerplate code'. Has some capacity to make code more portable: TODO!)
• https://github.com/YuxinZhaozyx/pytorch-VideoDataset (videodataset helper module, just uses the dataset.py and transforms.py. Expects them to be in the same dir: using only transforms.py)

most of what is in the my_prog.py is from the README.md

Some dependencies

• pytorch https://pytorch.org/ (will give you a conda command to run if you select appropriate for you system)
• pytorch lightning https://github.com/PyTorchLightning/pytorch-lightning
• CUDA toolkit https://developer.nvidia.com/cuda-downloads

Been running it in an anaconda env https://www.anaconda.com/products/individual

after installing, reload command prompt/bash (windows has a shortcut installed in the start menu for anaconda)

in my_prog.py are two vars:

• SEED = 1
• NUM_WORKERS = 8 (set to your number of CPU cores)

if you have multiple GPUs, edit the following line param gpus=x

line to edit

trainer = pl.Trainer(gpus=1)

Setup Instructions

1. Install nvidia cuda toolkit from link above

conda env create --file ./envs/video.yml

python my_prog.py /path/to/dcsass/videos/

Use - Generic

from predictor_src/scr.py

import predict
with open ("output.txt", "r") as myfile:
    data=myfile.readlines()

predict.predictor(data,"./example.ckpt")

Notes

Transforms

• VideoToTensor(frames_in_tensor=10,fps=2)
• Grayscale
• Resize(128,128)

Label Encoder

• Not well implemented: creates a sklearn labelencoder within the dataset class init then uses a helper method to create numpy arrays
• Labels in my_prog.py available as a global var list which the predict.py uses to relabel identify it's output predictions

Encoder

• Linear(in=video.xvideo.yframes_in_tensor,out=64)
• ReLU
• Linear(out=64, num_of_classes)

Decoder

• Linear(in=num_of_classes,out=64)
• ReLU
• Linear(in=64,out=video.xvideo.yframes_in_tensor)

Early Stop Callback

• generic callback, uses log_acc var. currently running tests
• Note: Doesn't produce a stable model

Forward Step

• Unpacks tuple. taking the frame representation only, ignoring any category
• Passes through the encoder and returns residual

TODO

• Optimizer description
• validation step & epoch_end desc

compx527_bag's Issues

Unknown labels

Some, mostly in one category, mp4s in the designated dataset are not referenced in the labels CSVs.

Is this because they don't have any noted activity?
should we include these as our baseline?

No validation step

The lightning module has no validation step. There is a dataloader passed to the trainer but it will not get used in the process.

• find out what the validation step does and decide if its useful
• find example implementations (might have seen some implementations that are part of pytorch/lightning)