TECHNICAL DOCUMENTATION

Technical overview of the framework

Discover architectures, loss functions, metrics ready-to-use in openmlcore.

Model Architectures Availables

OpenMLCore provides ready-to-uses machine learning tool, written in PyTorch for computer vision tasks such as segmentation or classification. Click on any cards to learn more.

openmlcore.models Stable

UNet

Semantic Segmentation Pixel by Pixel

CNN Architecture based on an encoder, a bottleneck and a decoder. The U-Net uses skip connection to provide spatial infos to the decoder to create the segmented image.

Input: (B, C_in, H, W)

Output: (B, nb_class, H, W)

Params: input_channels, output_channels, depth, initial_filters Details

openmlcore.models Stable

Attention Unet

Advanced CNN architecture

Based on the Unet, the Att-U-Net uses soft attenntion gates to have region of interest in the images and make the prediction easier.

Input: (B, C_in, H, W)

Output: (B, nb_class, H, W)

Params: input_channels, output_channels, depth, initial_filters Détails

openmlcore.models Stable

UNet++

Semantic Segmentation Pixel by Pixel

CNN Architecture based on an encoder, a bottleneck and a decoder. The U-Net++ is an complex CNN Architectures to do image segmentation.

Input: (B, C_in, H, W)

Output: (B, nb_class, H, W)

Params: input_channels, output_channels, depth, initial_filters Details

openmlcore.models Stable

SegFormer

Vision Transformers

Vision Transformer Architecture based on an encoder - decoder. Note that SegFormer is hard to converge, impossible to converge on small dataset. To avoid that, please use data augmentation and drop out.

Input: (B, C_in, H, W)

Output: (B, nb_class, H, W)

Params: input_channels, output_channels, depth, initial_filters Details

openmlcore.models Stable

MobileUnetR

Vision Transformers

Vision Transformer Architecture based on an Unet architecture, with little weight Transformers. This model has the best performance on Cell dataset for now.

Input: (B, C_in, H, W)

Output: (B, nb_class, H, W)

Params: input_channels, output_channels, depth, initial_filters Details

openmlcore.models Soon

Sam2 & QLoRA

Vision Transformers

SoTA from Meta for the task of segmentation.

Input: (B, C_in, H, W)

Output: (B, nb_class, H, W)

Params: input_channels, output_channels, depth, initial_filters Details

Loss Functions

Ready-to-use loss function, and the possibility to create custom loss based on classical one. Available in the sub-package openmlcore.loss. Compatible with PyTorch.

CustomLoss

Composite

Stable

With CustomLoss, you can ponderate classical loss to imporve your finetuning!

Example : Combinaison Dice + BCE (0.7 * Dice + 0.3 * BCE) :

from openmlcore.loss import CustomLoss, DiceLoss, BCEWithLogitsLoss

weights_for_class = torch.tensor([4.0, 1.0, 4.0, 3.5, 2.0])
weights_type_loss = [0.7, 0.3]

myLoss = CustomLoss(loss_fcns=[ DiceLoss(weights=weights_for_class), BCEWithLogitsLoss(weights=weights_for_class)], name="CustomLoss", coefficients=weights_type_loss)

DiceLoss

Overlap-based

Stable

Based on the Sørensen–Dice coefficient, it prioritizes the overlap between the ground truth and the prediction. Unlike BCE, Dice Loss operates at the region level. It is very effective for small objects or class imbalance.

Example (weights for multiclass):

from openmlcore.loss import DiceLoss

weights_for_class = torch.tensor([4.0, 1.0, 4.0, 3.5, 2.0])
myLoss = DiceLoss(weights=weights_for_class)

Key Parameters:

weights (Tensor=None): weights for multiclass

BCEWithLogitsLoss

Entropy-based

Stable

BCE is based on a logarithmic difference and works on a pixel-by-pixel basis. It penalizes incorrect predictions more than it rewards correct ones.

Example for multiclass segmentation: :

from openmlcore.loss import BCEWithLogitsLoss

weights_for_class = torch.tensor([4.0, 1.0, 4.0, 3.5, 2.0])
myLoss = BCEWithLogitsLoss(weights=weights_for_class),

Paramètres clés :

Weights (Tensor): Weights for multiclass segmentation

MSE

Classic Loss

Stable

TODO

Example for multiclass segmentation: :

from openmlcore.loss import BCEWithLogitsLoss

weights_for_class = torch.tensor([4.0, 1.0, 4.0, 3.5, 2.0])
myLoss = BCEWithLogitsLoss(weights=weights_for_class),

Paramètres clés :

Weights (Tensor): Weights for multiclass segmentation

Technical overview of the framework

Model Architectures Availables

UNet

Attention Unet

UNet++

SegFormer

MobileUnetR

Sam2 & QLoRA

Loss Functions

CustomLoss

DiceLoss

BCEWithLogitsLoss

MSE

Performance metrics

accuracy

iou

Precision

Recall

f1_score