98%|█████████▊| 45/46 [00:37<00:00, 1.21it/s]tensor([[[[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
...,
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.]]]], device='cuda:0')
100%|██████████| 46/46 [00:38<00:00, 1.20it/s]
tensor([[[[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
...,
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.],
[0., 0., 0., ..., 0., 0., 0.]]]], device='cuda:0')
Traceback (most recent call last):
File "P:\CJJ\EGE-UNet-main\train.py", line 169, in
main(config)
File "P:\CJJ\EGE-UNet-main\train.py", line 154, in main
loss = test_one_epoch(
File "P:\CJJ\EGE-UNet-main\engine.py", line 147, in test_one_epoch
TN, FP, FN, TP = confusion[0,0], confusion[0,1], confusion[1,0], confusion[1,1]
IndexError: index 1 is out of bounds for axis 1 with size 1

print(confusion) is [[3014656]]
It is caused by only one category?

Hello,
I've been working with your released available dataset (isic2017/2018) and I noticed an issue with the mask images. It appears that the mask images provided are not binary as expected for segmentation tasks. This may affect the performance of models being trained for segmentation purposes and the evaluation of those models. You can refer the code below,

import numpy as np
from PIL import Image

path="./data_isic1718/isic2017/train/masks/ISIC_0000000_segmentation.png"

print(np.unique(np.array(Image.open(path))))

output like this:
[  0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17
  18  19  20  22  23  24  25  26  27  29  30  31  32  33  34  35  36  37
  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  55  57  59
  60  61  62  64  65  66  67  68  69  70  71  72  73  74  75  76  77  78
  79  80  81  82  83  84  85  86  87  88  89  90  91  93  94  95  96  97
  98 100 101 102 103 104 105 106 107 108 109 110 112 113 115 116 117 118
 119 120 121 122 123 124 126 128 129 130 131 134 135 136 137 138 139 141
 142 143 144 145 146 147 148 149 150 151 152 153 154 155 157 158 159 160
 161 162 165 166 167 171 172 173 174 175 176 177 178 179 181 182 183 184
 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203
 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255]
 

Hello, we use the BUSI dataset similar to single-object segmentation where the gradient cannot descent.

I use official code and isic2017 dataset to train, but unable to reproduce paper results, val miou is about 0.34.

Could author provide some experiment results of speed test in some GPU, likes 1080TI, 3080. Thanks

论文里面写的是0.053，但是实际上是0.0458，我觉得应该不是我复现的问题吧，因为我是直接用的你开源的模型代码

Why does the testing phase after training end in errors?

#----------Testing----------#
0%| | 0/650 [00:00<?, ?it/s]
......
ValueError: height and width must be > 0

When running your code today, I found that the code I wrote could not predict the desired segmentation graph.

非常感谢您提供的简洁清晰的代码，我在调试您的代码的时候发现如果将gt_ds设置为False，程序将会报错。请问该如何修改？

thx for your work!
Can this pth weight be converted to onnx format and reasoned about?

I look at the code and it seems more realistic to just output torch.sigmoid(out0) when testing with 'gt_ds':=False, but I followed this and found that self.GAB1-5 must be entered in the @JCruan519

针对模型直接用torch.onnx.export 转换为onnx，这时没有提示错误，当采用onnxsim 进行优化时，提示gmhpa/Resize_2_output_0 这个错误，指向了Grouped_multi_axis_Hadamard_Product_Attention 中的F.interpolate( params_zy），提示为这里没有定义；针对原始的onnx转换为mnn模型也没有提示错误，但是执行mnn时也提示了如onnxsim 同样的错误，指向了params_zy 的 resize 操作结果没有定义，不知道这个是什么原因。

Hello, my test results on Unet are different from the original paper. My result on Unet is 86.73.
Thank you!

I would like to ask, how is the dataset processed?
Is it true that the official data's training set and validation set are merged and then divided into an experimental training set and test set with a 7:3 ratio?

Thanks for your contribution. I just have a question about the input size. I want to train with a larger input size (512 x 512),, but will it impact performance a lot? or it still adaptive .

为什么我训练一个epoch要一两个小时，batch_size为512，训练集大概一万张，测试集2000张，而且我改了gpu_id也不可以用多卡训练？请问是我那个参数设置的问题吗？

Thanks for your good reasearch and I trained the network in sky datasets.
compared with u2netp, the model size has been reduced by 10 times, and the inference speed has increased by more than 14 times although the model has a slight decrease in accuracy.

Hi. May I know is there any reason to wrap all "self.gt_conv*" layers with "nn.Sequential" instead of directly assign them as "nn.Conv2d" layer?

https://github.com/JCruan519/EGE-UNet/blob/main/models/egeunet.py#L195

Can the best saved model be made public?

想问一下作者大大，所有的对比实验都是自己复现的代码还是开源代码？论文中U-Net训练ISIC2017数据集DSC能够达到86.99。而我使用UNet-pytorch只能达到83.976。这个结果差距会不会太大了一点？


另外EGE-UNet在ISIC2017数据集的复现结果也略微偏低，请问这是我显卡的问题嘛？