Multi-label binary classification for multi-weather prediction with MLflow integration

This repository can be used to develop vision models for multi-label classification (One-vs-Rest) problems. As of now, model for prediction of multiple weather conditions from images of outdoor scenes (as shown in the below figure) has been developed as MLflow project.

General requirements

• OS : ubuntu 20.04 LTS
• CUDA version : 11.6
• Make

Environment setup & How to run

Specific CUDA version 11.6 is required if you use the following setup automation commands. If you build your own environment, you can skip the environment setup section.

General steps

• Download the required dataset and pretrained model file by running

make download-data

• There are two options to setup the working environments for this repository.

[Option 1] Setting up MLflow environment and running as a mlflow project

• Setup a working environment to run as a MLflow project

make setup-mlflow-env

• Then, you can simply run operations by

make run-as-mlflow-project OPERATION=<operation> CONFIG=<config_path>

• The value of operation can be one of { train, train_eval, evaluate, validate }

  • train : Simply train the model and saved best and last checkpoints
  • train_eval : Train the model and then evaluate the best checkpoint on the unseen test dataset
  • evaluate : Evaluate the model on the unseen test dataset
  • validate : Visualize the model prediction and compare with ground truth value on the unseen validation/test dataset. Sample

• The value of config_path is set default to config config.yml. So, you can leave it as optional.

• So, for example, if you want to train and evaluate the best checkpoint model, you can run

make run-as-mlflow-project OPERATION=train_eval

[Option 2] Setting up docker environment and running the operations

• If you do not want to setup MLflow environment, you can try this option. For building the docker image for the first time, simply run (you need to run just once),

make build-docker-env

• For setting up the working docker environment, please run

make setup-docker-env

• Once you enter inside working docker container, you can run

python main.py <operation>

• The value of operation can be one of { train, train_eval, evaluate, validate }.

• Additionally, you can use this demo script for inferencing a single image.

Backbone options

There are two available backbone options.

• Resnet50
• Inceptionv2

For changing the model backbone, You can change the related params under backbone in config.yml

Model Architecture

Overview of multi-label-binary-classifier architecture is as follow:

Dataset

The dataset used for this repository is a combined dataset of multi-label-weather-dataset and weathertime-classification-with-road-images from Kaggle.

Additional features

logging performance metrics, configuration params and the trained model as MLflow model format

• This helps you facilitate your development workflow and give you a standard model format to be used in a variety of many downstream tools.
• You can easily analyze and explore your past trainings including metrics and artifacts via simply

mlflow ui

Pretrained weights

Currently, I have trained the model using resnet50 and inceptionv2 backbones respectively and probably you have downloaded the pretrained models in previous section

Evaluation results of the pretrained model with Inceptionv2 backbone

\

Classification report for sunny weather classification model:

	precision	recall	f1-score	support
Not sunny	0.944409	0.965382	0.95478	1531
sunny	0.878161	0.814499	0.845133	469
macro avg	0.911285	0.889941	0.899957	2000
weighted avg	0.928874	0.93	0.929068	2000

Overall accuracy : 0.93

\

Classification report for cloudy weather classification model:

	precision	recall	f1-score	support
Not cloudy	0.856195	0.763314	0.807091	507
cloudy	0.922481	0.956463	0.939165	1493
macro avg	0.889338	0.859889	0.873128	2000
weighted avg	0.905677	0.9075	0.905684	2000

Overall accuracy : 0.9075

\

Classification report for foggy weather classification model:

	precision	recall	f1-score	support
Not foggy	0.952022	0.954639	0.953329	1455
foggy	0.878004	0.87156	0.87477	545
macro avg	0.915013	0.913099	0.914049	2000
weighted avg	0.931852	0.932	0.931921	2000

Overall accuracy : 0.932

\

Classification report for rainy weather classification model:

	precision	recall	f1-score	support
Not rainy	0.95534	0.952873	0.954105	1549
rainy	0.83956	0.847007	0.843267	451
macro avg	0.89745	0.89994	0.898686	2000
weighted avg	0.929232	0.929	0.929111	2000

Overall accuracy : 0.929

\

Classification report for snowy weather classification model:

	precision	recall	f1-score	support
Not snowy	0.966091	0.9924	0.979069	1579
snowy	0.968254	0.869359	0.916145	421
macro avg	0.967173	0.930879	0.947607	2000
weighted avg	0.966546	0.9665	0.965824	2000

Overall accuracy : 0.9665

\

Evaluation results of the pretrained model with Resnet50 backbone

\

Classification report for sunny weather classification model:

	precision	recall	f1-score	support
Not sunny	0.947607	0.956891	0.952226	1531
sunny	0.854626	0.827292	0.840737	469
macro avg	0.901116	0.892092	0.896481	2000
weighted avg	0.925803	0.9265	0.926082	2000

Overall accuracy : 0.9265

\

Classification report for cloudy weather classification model:

	precision	recall	f1-score	support
Not cloudy	0.820041	0.790927	0.805221	507
cloudy	0.929848	0.941058	0.935419	1493
macro avg	0.874944	0.865993	0.87032	2000
weighted avg	0.902012	0.903	0.902414	2000

Overall accuracy : 0.903

\

Classification report for foggy weather classification model:

	precision	recall	f1-score	support
Not foggy	0.937375	0.96701	0.951962	1455
foggy	0.903808	0.827523	0.863985	545
macro avg	0.920591	0.897267	0.907973	2000
weighted avg	0.928228	0.929	0.927988	2000

Overall accuracy : 0.929

\

Classification report for rainy weather classification model:

	precision	recall	f1-score	support
Not rainy	0.943289	0.96643	0.954719	1549
rainy	0.874092	0.800443	0.835648	451
macro avg	0.908691	0.883437	0.895184	2000
weighted avg	0.927685	0.929	0.927869	2000

Overall accuracy : 0.929

\

Classification report for snowy weather classification model:

	precision	recall	f1-score	support
Not snowy	0.96201	0.9943	0.977889	1579
snowy	0.975543	0.852732	0.910013	421
macro avg	0.968777	0.923516	0.943951	2000
weighted avg	0.964859	0.9645	0.963601	2000

Overall accuracy : 0.9645

References

Dataset

• /~https://github.com/wzgwzg/Multi-Label-Weather-Recognition
• https://www.kaggle.com/datasets/wjybuqi/weathertime-classification-with-road-images