YOLO trained models

This is weights of trained YOLOv8 models in three variations:

• nano
• middle
• large

It was trained and further used in "equal" data organization, all images were squared crops size of 384 pixels.

More sustainable model to different scale of spike on crop

This model have the same characteristics as model v1.0.0 except dataset.
Images were augmented by different scale of spike on crop to make model more sustainable to different crops:

cropping.py file: https://github.com/kiteiru/nsu-diploma-wheat/tree/main/notebooks/images_cropping

Geometry of spikelets: circles

Final model as a result of graduation work

This model was achieved after series of experiments.
Model is described with its parts and training hyperparameters (for reproducing results also need to mention images and augmentation parameters), namely:

Architecture: Unet
Encoder/backbone: EffecientNet-b4
Loss: BinaryCrossEntropy
Optimizer: RMSProp(lr = 8.57e-4, eps = 8.14e-6, mu = 0.479)
Batchsize: 8
Cropsize: 384 px
Augmentation: HorizontalFlip(probability = 0.736), VerticalFlip(probability = 0.277), Rotate(limit=30, probability = 0.735), ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.2, probability = 0.703)
Markup: Circles and Ellipses separately

This models are trained, validated and tested on data organization "equal" due to the higher variability of data in sets in comparison with "certain" and "random" data organizations.

Best model in that very moment after series of experiments

This is model was received after series of experiments with different topologies and backbones, its trained, validated and tested on equal data organization, its path: "project/smp/data_organization/equal.json"
Pipeline code for using this model: ''project/smp/pipeline"

Architecture: Unet
Encoder/backbone: EffecientNet-b4
Loss: BCEwithLogits
Optimizer: Adam
Learning rate: 1e-3

Old models giving circle and ellipsoid binary masks as output

There are models trained, validated and tested on 3 types of data organization
Data organization jsons can be found by path: "project/smp/data_organization"

Architecture: Unet
Encoder/backbone: EffecientNet-b2
Loss: BCEwithLogits
Optimizer: Adam
Learning rate: 1e-3

Model name is "Geometry-of-output-masks + Name-of-data-organization"