retinal_vessel_segmentation

Data

Augmentation

Training (fucked color channel)

Test example

Installation

$ pip install -r requirements.txt
$ python prepare_dataset.py
$ python main.py

TODO: about all scripts

Features

• Monitoring with W&B
• HPO
• Running of multiple configs with hydra ...TODO

TODO

• suppersampling or upsampling back resolution (GuidedConvFilter add to the architecture)
• stable diffusion for data augmentation (ControlNet)
• U-net squarred
• hydra
• interpretebility (gradcam, saliency maps ,captum)
• distributed training
• MaskFormer, FocalNet
• use torch.timm or smp backbones
• hyperopt (hpo)
• multi-layer loss
• DiceLoss + FocalLoss https://gitlab.giraffe360-mimosa.com/machine-learning/training/mirror-segmentation-trainer/-/blob/main/scripts/model.py?ref_type=heads
• Tversky loss (check smp libraries loss and metrics)
• add reqs (mamba, miniconda, devcontainer)
• add confusion matrix to wandb
• add modelcheckpoint for saving weights
• self-supervised learning (since masks are not perfect)
• advanced train set augmentation (ideas :get new images with diffusion -> segment -> retrain and compare, maybe filter out bad masks visually or which are out of distribution)
• try to inference with this model bigger dataset on Kaggle and compare
• numba and torch.jit
• model pruning and distillation
• openvino inference

Fix

• show augmented training image masks
• wandb don't delete local runs which were removed on website
• wandb multiple colors on segmentation
• wandb image plot steps don't match epoch
• inverse mask
• inference not working (+ test trainer.test)

pip install hydra-core

You are using a CUDA device ('NVIDIA RTX A4000') that has Tensor Cores. To properly utilize them, you should set torch.set_float32_matmul_precision('medium' | 'high') which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision