ahmedhosny / 3d-unetplusplus

3d-unetplusplus

docker build -f dkr-model -t dkr-model . docker run -it --gpus all -v$PWD/files:/files -v /mnt:/mnt dkr-model /bin/bash

docker run -it --gpus all -v$PWD/files:/files -v /mnt:/mnt -p 8886:8888 dkr-model /bin/bash

to go into running container

docker exec -it /bin/bash

notebook

jupyter notebook --allow-root --ip=0.0.0.0

train

20-24 2d multi model

32 2d single model (lung first)

34 2d single model (lung middle, same order as planned)

only lung has ones

36 2d single model (no lung, removed from middle)

only cord has ones

---

37 3d single model for heart - dice

38 tversky loss - heart

39 tversky focal loss - heart

40 tversky focal loss 1/gamma - heart

---

41 dice ctv

42 tversky ctv ---x

43 ctv tversky focal

44 ctv tversky focal loss 1/gamma

---

45 ctv -tversky ---x

46 heart - tversky

47 lung - tversky

48 esophagus - tversky

49 cord - tversky

---

50 gtv- tversky

51 gtv - dice

52 gtv - tversky focal

53 gtv - tversky focal loss 1/gamma

54 gtv - bbox distance loss

---

55 gtv maastro without generator, weighted_dice_coefficient_loss

56 gtv maastro with crop/rotate, weighted_dice_coefficient_loss

57 gtv maastro with crop/rotate/blur, weighted_dice_coefficient_loss

---

#60-89 different loss functions at 3 lrs

---

testst for LR schedule

103 - focal-tversky-loss-0.0005

105 - focal-tversky-loss-lr (bad) (0.5, 5)

106 - focal-tversky-loss-lr-tight (0.8, 20)

107 - focal-tversky-loss-lr-tight-2 (0.8, 12)

108 - focal-tversky-loss-lr-tight-2-affine (0.8, 12) adds affine trans

109 - focal-tversky-loss-0.0005-augment (like 103 with affine and elastic, and more epochs) - multi(average) and single

110 - like 109 but with train on harvard-rt only - multi(average) and single

111 - like 109 but with train on maastro only - multi(average) and single (only did single for now, need plotting)

112, take model from 111 and train all layes on harvard-rt (only did single for now)

---

test

23 - 37~40

NOTEs: compare 45 to 50 then 50 to 51