tf-yolov4-pipeline

Modular image processing pipeline implementation of [tensorflow-yolov4-tflite] (https://github.com/hunglc007/tensorflow-yolov4-tflite).

Setup environment

This repo implements another repo: [tfpipe] (https://github.com/jfw225/tfpipe). To make sure that it is installed, do the following:

$ git submodule update --init --recursive

Next, use your preferred package manager to install the packages listed in requirements.txt.

Using pip:

$ python -m pip install -r requirements.txt

Command Line Arguments and Config

Each script (save_model.py and process_images.py) has a set of command line arguments. You can see a full list of these arguments by running:

$ python save_model.py --help
$ python process_images.py --help

However, each of these parameters can also be changed in the config.py file.

Converting from Darknet

First, change the config or set the parameters for your .weights file. Next, execute:

$ python save_model.py

Basic Usage

Install the appropriate model, config, data, and class files. You can then use process_images.py to process images.

Run the following command to see all of the available options:

$ python process_images.py --help

The -i or --input arguments are used to specify an input file path. The file path can either be a directory or a file, and the types of files that can be read are specified in config.py.

Here is a demo call using the the COCO model from the repo linked above:

$ python process_images.py -w checkpoints/yolov4-416 -i data/

Redis

To use Redis, you must specify the --redis argument. The host and port of the Redis server can be specified by --redis-host and --redis-port arguments.

To submit images to the processing queue, open your redis-cli and publish a file path to the frames channel:

$ publish frames <image/directory file path>

To read the output of the model, have a seperate redis-cli subscribe to the bbox channel:

$ subscribe bbox

TODO

add git clone --recursive add libdevice.10.bc use cpu for non-predict functions in async cpus corresponding to gpus wrap more things in tf.function add use all gpus

figure out appropriate vram create multiple virtual gpus make sure the only gpu operations are in prediction get rid of superfluous keys in data

create virtual cpus

change all to use tf functions (update redis)

abstract bounding box iterators

abstract redis input and image input

write tflite conversion

for async in realtime, constantly have input add to fixed size queue. when gpu ready, popright.