issue in training

[manvirvirk]

INFO: Using device cuda
INFO: Network:
NestedUNet model
3 input channels
3 output channels (classes)
Bilinear upscaling
INFO: Creating dataset with 20 examples
training stops after this. can anyone help

[zonasw]

Could u give me a more detailed description of the error

[manvirvirk]

RuntimeError: CUDA error: device-side assert triggered

Traceback (most recent call last):
File "train.py", line 254, in
train_net(net=net, cfg=cfg)
File "train.py", line 142, in train_net
val_score = eval_net(net, val_loader, device, n_val, cfg)
File "train.py", line 210, in eval_net
sub_cross_entropy = F.cross_entropy(pred.unsqueeze(dim=0), true_mask.unsqueeze(dim=0).squeeze(1)).item()
RuntimeError: CUDA error: device-side assert triggered

[zonasw]

Your category number is 3, so your label map should only contain 0,1 and 2, but your label map contain something else. Check your label map.

[manvirvirk]

could not get it? Please give more details. Thanks

[zonasw]

Check that your label map contains only 0,1, and 2. You can complete the check by running the following program:

import os
import os.path as osp
from tqdm import tqdm
import cv2
import numpy as np


num_classes = 3
mask_dir = "masks"
mask_names = os.listdir(mask_dir)

for mask_name in tqdm(mask_names):
    mask_path = osp.join(mask_dir, mask_name)
    mask = cv2.imread(mask_path, 0)
    h, w = mask.shape[:2]
    pix = []
    for i in range(0, num_classes):
        pix.append(len(np.where(mask==i)[0]))
    if sum(pix) != h*w:
        print("error: " + mask_name)

[sunchuanxi]

when i set the model='Unet',cannot run,could you help to reply me ,thinks

follow is the erros:
Traceback (most recent call last):
File ".\train.py", line 255, in
train_net(net=net, cfg=cfg)
File ".\train.py", line 122, in train_net
loss += criterion(inference_mask, masks)
File "C:\Users\Administrator\anaconda3\lib\site-packages\torch\nn\modules\module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "E:\深度学习\github\unet多分类\unet-nested-multiple-classification-master\losses.py", line 23, in forward
loss = L.lovasz_softmax(out, target)
File "E:\深度学习\github\unet多分类\unet-nested-multiple-classification-master\LovaszSoftmax\pytorch\lovasz_losses.py", line 167, in lovasz_softmax
loss = lovasz_softmax_flat(*flatten_probas(probas, labels, ignore), classes=classes)
File "E:\深度学习\github\unet多分类\unet-nested-multiple-classification-master\LovaszSoftmax\pytorch\lovasz_losses.py", line 190, in lovasz_softmax_flat
raise ValueError('Sigmoid output possible only with 1 class')
ValueError: Sigmoid output possible only with 1 class

[MEYE66]

不能把category number的类别定义为mask的分割种类呢，即不需要再把mask的颜色种类重新按0,1,2,3这样排序，保留原本的颜色？

[zonasw]

不能把category number的类别定义为mask的分割种类呢，即不需要再把mask的颜色种类重新按0,1,2,3这样排序，保留原本的颜色？

可以，但你需要对加载数据的代码做一些修改。

[wangfusheng-cqu]

这代码只能运用于图像尺寸一样的数据集吗？？？