search

MIC-DKFZ / medicaldetectiontoolkit

The Medical Detection Toolkit contains 2D + 3D implementations of prevalent object detectors such as Mask R-CNN, Retina Net, Retina U-Net, as well as a training and inference framework focused on dealing with medical images.
Apache License 2.0
1.31k stars 297 forks source link

seg_logits is zero and retina_unet learned nothing #104

Closed amberhappy closed 4 years ago

amberhappy commented 4 years ago

Hi,I train retina_Unet on LIDC dataset ,but got result like this:

tr. batch 101/200 (ep. 1) fw 2.07s / bw 2.82 s / total 4.89 s || loss: 1.57, class: 0.76, bbox: 0.31, seg dice: 0.998, seg ce: 0.012, mean pix. pr.: 0.00000
tr. batch 102/200 (ep. 1) fw 1.99s / bw 2.80 s / total 4.80 s || loss: 1.75, class: 0.83, bbox: 0.42, seg dice: 0.999, seg ce: 0.009, mean pix. pr.: 0.00000
tr. batch 103/200 (ep. 1) fw 2.81s / bw 2.89 s / total 5.70 s || loss: 1.30, class: 0.51, bbox: 0.29, seg dice: 1.000, seg ce: 0.006, mean pix. pr.: 0.00000
tr. batch 104/200 (ep. 1) fw 1.94s / bw 2.39 s / total 4.33 s || loss: 2.10, class: 0.85, bbox: 0.74, seg dice: 0.999, seg ce: 0.017, mean pix. pr.: 0.00000
tr. batch 105/200 (ep. 1) fw 2.19s / bw 2.62 s / total 4.80 s || loss: 2.17, class: 0.98, bbox: 0.69, seg dice: 0.998, seg ce: 0.011, mean pix. pr.: 0.00000
tr. batch 106/200 (ep. 1) fw 2.97s / bw 2.93 s / total 5.90 s || loss: 1.82, class: 1.00, bbox: 0.31, seg dice: 0.999, seg ce: 0.020, mean pix. pr.: 0.00000
tr. batch 107/200 (ep. 1) fw 2.10s / bw 1.82 s / total 3.92 s || loss: 1.08, class: 0.56, bbox: 0.01, seg dice: 1.000, seg ce: 0.004, mean pix. pr.: 0.00000
tr. batch 108/200 (ep. 1) fw 3.25s / bw 2.93 s / total 6.18 s || loss: 1.24, class: 0.56, bbox: 0.18, seg dice: 1.000, seg ce: 0.004, mean pix. pr.: 0.00000
tr. batch 109/200 (ep. 1) fw 1.66s / bw 2.55 s / total 4.21 s || loss: 1.87, class: 0.79, bbox: 0.56, seg dice: 0.998, seg ce: 0.021, mean pix. pr.: 0.00000
tr. batch 110/200 (ep. 1) fw 2.12s / bw 3.23 s / total 5.35 s || loss: 1.75, class: 0.82, bbox: 0.43, seg dice: 1.000, seg ce: 0.005, mean pix. pr.: 0.00000
tr. batch 111/200 (ep. 1) fw 2.96s / bw 3.24 s / total 6.20 s || loss: 1.82, class: 0.82, bbox: 0.50, seg dice: 0.999, seg ce: 0.010, mean pix. pr.: 0.00000
tr. batch 112/200 (ep. 1) fw 3.87s / bw 3.20 s / total 7.07 s || loss: 1.07, class: 0.55, bbox: 0.02, seg dice: 1.000, seg ce: 0.002, mean pix. pr.: 0.00000
tr. batch 113/200 (ep. 1) fw 2.96s / bw 3.09 s / total 6.05 s || loss: 1.89, class: 0.97, bbox: 0.40, seg dice: 0.998, seg ce: 0.028, mean pix. pr.: 0.00000
tr. batch 114/200 (ep. 1) fw 2.15s / bw 1.46 s / total 3.61 s || loss: 1.00, class: 0.43, bbox: 0.07, seg dice: 1.000, seg ce: 0.003, mean pix. pr.: 0.00000
tr. batch 115/200 (ep. 1) fw 3.00s / bw 2.34 s / total 5.34 s || loss: 1.50, class: 0.73, bbox: 0.26, seg dice: 0.999, seg ce: 0.011, mean pix. pr.: 0.00000
tr. batch 116/200 (ep. 1) fw 2.96s / bw 1.67 s / total 4.63 s || loss: 1.37, class: 0.65, bbox: 0.22, seg dice: 0.999, seg ce: 0.008, mean pix. pr.: 0.00000
tr. batch 117/200 (ep. 1) fw 2.92s / bw 3.18 s / total 6.10 s || loss: 1.27, class: 0.58, bbox: 0.18, seg dice: 1.000, seg ce: 0.007, mean pix. pr.: 0.00000
tr. batch 118/200 (ep. 1) fw 3.35s / bw 3.13 s / total 6.48 s || loss: 1.57, class: 0.83, bbox: 0.23, seg dice: 0.999, seg ce: 0.029, mean pix. pr.: 0.00000
tr. batch 119/200 (ep. 1) fw 3.48s / bw 3.14 s / total 6.62 s || loss: 1.71, class: 0.76, bbox: 0.44, seg dice: 0.998, seg ce: 0.020, mean pix. pr.: 0.00000
tr. batch 120/200 (ep. 1) fw 8.27s / bw 2.65 s / total 10.92 s || loss: 1.85, class: 0.91, bbox: 0.43, seg dice: 0.999, seg ce: 0.014, mean pix. pr.: 0.00000

mean pix. pr always be zero,retina_unet seems learn nothing. What should I do to solve this problem?

mehak-agr commented 4 years ago

Hi @amberhappy were you able to find the reason for this?

amberhappy commented 4 years ago

It happens only in epoch1, just let it go on and train more epoches.