Pytoch >= 1.6
Following guidelines should be followed to correct performance:
- Datasets have to return dictionaries with:
- Training dataloader at least 1.'image' and 2.'label' and 3.'original_mask' entries and 'num_classes' and 'class_to_cat' attributes
- num_classes should be 1 (no background) for single class segmentation or the number of classes + 1 (background)
- when multiclass and average metrics and a class at class_to_cat named 'Mean' or as you prefer
- Validation dataloader at least 1.'image', 2.'original_img', 3.'original_mask', 4.'img_id' entries
- If you want to load checkpoint unfreezed set defrost_epoch param to 0
- In segmentation background class is equals to label 0
- You can use --notify to send you a slack message to 'experiments' channel. Set envionment variable with slack token. How can create slack token here:
export SLACK_TOKEN='you_slack_token'
MMs dataset naming:
_full
Get all volumes (not only segmented 'ED' and 'ES' phases volumes)._unlabeled
Get only unlabeled volumes (for 'ED' and 'ES' phases)_centre*xyz*
Get volumes (for 'ED' and 'ES' phases) for selected centres. Example_centre1
,_centre13
. Last one picks centres 1 and 3. Available Centres from 1 to 5._vendor*jkl*
Get volumes (for 'ED' and 'ES' phases) for selected vendors. Example_centreC
,_vendorAB
. Last one picks vendors A and B. Available Vendors 'A', 'B', 'C', 'D'.
-
Valores de distancias infinitos (Hausdorff, ASSD)?
-
Ejemplos de uso: classification.sh
-
Redes: classification y segmentation
-
Método report para guardar metricas en csv de:
- Por pacientes par analisis de errores
-
Probar problemas classification multiclase y una clase