ajime

What is this?

Still me goofing around.

Seriously what is this?

this is classifier trained on very few images scraped from duckduck go.

What does it classify?

• Potholes
• Dumping
• Accidents
• Flooded
• Bad drainage
• Construction

Yeah I know they don't make much sense but .....

Companion repo

• mrdvince/mime

This has all the "stitched up" code for the whole thing.

Training graphs and logs

A small dataset was used, more of a base line and less of perfomance and accuracy oriented.

Some training losses.

Inputs (train, valid)

Getting Started

Dependencies

To set up your python environment to run the code in this repository, follow the instructions below.

1. Create (and activate) a new environment with Python 3.6. - Linux or Mac: bash conda create --name py39 python=3.9.7 conda activate py39 alternatively, use virtual environments if you don't have Anaconda installed.

2. Clone the repository (if you haven't already!), and navigate to the anga folder. Then, install several dependencies.

git clone https://github.com/mrdvince/ajime
cd ajime

3. Requirements

Install requirememts in the rquirements.txt file

pip install -r requirements.txt

4. Images (Dataset)

cd into the data folder and open the images notebook

feel free to change the classes to whichever ones you like

executing the cells will download images from duckduckgo and store them in the data folder

5. Training

Modify the dataloader and the model classes and use a model definition of your choosing.

Update the config.json file and include the model and dataloader class names.

Finally run:

python train.py -c config.json

Folder structure

Head over to this repo to get an overview of the folder structure and file name descriptions.