forensic_pathology

---

---

require

pytorch (cuda=11.7)
torchvision
torchmetrics 
timm
einops
cython
scikit-image
opencv-python 
scikit-learn 
transformers
tqdm

usage

train contrast learning

train contrast learning, you need eight NVIDIA GEFORCE RTX 3090 Graphics Cards
python -m torch.distributed.launch --nproc_per_node=8 --master_port=xxxx  train_main.py --epochs=100 --batch_size_pergpu=128 --obj_loss=True | tee xxx.log

evaluate (linear)

linear evaluation (percent or all! if all data train_percent=1)
CUDA_VISIBLE_DEVICES=xxxx python -m torch.distributed.launch --nproc_per_node=8 --master_port=xxxx   linear_percent.py --train_percent=xxx  --save_checkpoint=xxx --weights=freeze  | tee xxx.log

train multiple instance learning

when train multiple instance learning, you must have a backbone checkpoint, and also a small batch_size is required
python adaptive_pool_train.py --epochs=50 --checkpoint=xxx  --size=xxxx --batch_size=xxx

more visulasition

---

---

reference

• vicreg
• SupContrast
• LEWEL
• barlowtwins

Citation