a question and a request
Opened this issue · 12 comments
Hello, take the liberty to interrupt, I downloaded the dataset isic2018 through the official website for the split is 2594 sheets Where should I get the extra 100 data from the 2694 in the paper?
If convenient, can you share your train.info .log?
2022-04-23 04:21:43 - val epoch: 299, loss: 0.3479
2022-04-23 04:21:43 - train: epoch 300, iter:0, loss: 0.1446, lr: 0.000999023230571994
2022-04-23 04:21:49 - train: epoch 300, iter:20, loss: 0.1743, lr: 0.000999023230571994
2022-04-23 04:21:55 - train: epoch 300, iter:40, loss: 0.1683, lr: 0.000999023230571994
2022-04-23 04:22:00 - train: epoch 300, iter:60, loss: 0.1629, lr: 0.000999023230571994
2022-04-23 04:22:06 - train: epoch 300, iter:80, loss: 0.1638, lr: 0.000999023230571994
2022-04-23 04:22:12 - train: epoch 300, iter:100, loss: 0.1653, lr: 0.000999023230571994
2022-04-23 04:22:17 - train: epoch 300, iter:120, loss: 0.1645, lr: 0.000999023230571994
2022-04-23 04:22:23 - train: epoch 300, iter:140, loss: 0.1616, lr: 0.000999023230571994
2022-04-23 04:22:29 - train: epoch 300, iter:160, loss: 0.1630, lr: 0.000999023230571994
2022-04-23 04:22:36 - train: epoch 300, iter:180, loss: 0.1610, lr: 0.000999023230571994
2022-04-23 04:22:42 - train: epoch 300, iter:200, loss: 0.1623, lr: 0.000999023230571994
2022-04-23 04:22:48 - train: epoch 300, iter:220, loss: 0.1662, lr: 0.000999023230571994
2022-04-23 04:23:55 - val epoch: 300, loss: 0.3976, miou: 0.7376317996433632, f1_or_dsc: 0.8490081728416308, accuracy: 0.9178844489673577, specificity: 0.9080565533258367, sensitivity: 0.94843156776658, confusion_matrix: [[36379921 3683576]
[ 664696 12224895]]
2022-04-23 04:25:04 - test of best model, loss: 0.2816,miou: 0.8071077684988276, f1_or_dsc: 0.8932591432212099, accuracy: 0.9481092018656211, specificity: 0.9661648357855531, sensitivity: 0.8919885821047386, confusion_matrix: [[38707942 1355555]
[ 1392223 11497368]]
Hi, thanks for following our work. The above result is the best one when training and testing our model on the ISIC2018 dataset. And the datasets of two datasets and the all weights will be update in the readme.md file later in the form of a link.
First of all, thanks for your reply. I have another question, the following is my log, but the result is very different with yours (my environment is ubuntu server and RTX3090). The code has not been changed, and the dataset uses isic2018_256.
2023-05-05 18:05:59 - train: epoch 300, iter:0, loss: 0.5077, lr: 0.000999023230571994
2023-05-05 18:06:00 - train: epoch 300, iter:20, loss: 0.6688, lr: 0.000999023230571994
2023-05-05 18:06:01 - train: epoch 300, iter:40, loss: 0.6885, lr: 0.000999023230571994
2023-05-05 18:06:02 - train: epoch 300, iter:60, loss: 0.6861, lr: 0.000999023230571994
2023-05-05 18:06:02 - train: epoch 300, iter:80, loss: 0.6950, lr: 0.000999023230571994
2023-05-05 18:06:03 - train: epoch 300, iter:100, loss: 0.6876, lr: 0.000999023230571994
2023-05-05 18:06:04 - train: epoch 300, iter:120, loss: 0.6910, lr: 0.000999023230571994
2023-05-05 18:06:05 - train: epoch 300, iter:140, loss: 0.6931, lr: 0.000999023230571994
2023-05-05 18:06:06 - train: epoch 300, iter:160, loss: 0.6945, lr: 0.000999023230571994
2023-05-05 18:06:07 - train: epoch 300, iter:180, loss: 0.6977, lr: 0.000999023230571994
2023-05-05 18:06:08 - train: epoch 300, iter:200, loss: 0.6980, lr: 0.000999023230571994
2023-05-05 18:06:09 - train: epoch 300, iter:220, loss: 0.6949, lr: 0.000999023230571994
2023-05-05 18:06:27 - val epoch: 300, loss: 1.2017, miou: 0.3388203635288079, f1_or_dsc: 0.5061476098791312, accuracy: 0.7951686633583206, specificity: 0.8953288610390496, sensitivity: 0.45815140322845055, confusion_matrix: [[35191396 4114157]
[ 6329580 5351875]]
2023-05-05 18:06:47 - test of best model, loss: 0.9647,miou: 0.38495204677204725, f1_or_dsc: 0.5559066794684588, accuracy: 0.8147856606922297, specificity: 0.9065609635361191, sensitivity: 0.5059817462807501, confusion_matrix: [[35632880 3672673]
[ 5770852 5910603]]
Hello, this result is indeed somewhat strange. The code I uploaded on GitHub is only a part of my complete project files, and I will check it when I have time later. You can try using the updated dataset link in the readme file for retraining and testing, and the results should not be biased so much. In addition, because the model parameters in this article are very small, it is certain that there is some bias in the training results.
Thank you for your answer.
The ISIC2018 of the official dataset is 2594, is it convenient to ask how the remaining 100 sheets came about?
2694 is composed of 2594 for training and 100 for validation from the offical site.
Thank you very much for your reply, I ran the code again this afternoon with your dataset, and it was not much different from the results in the paper table.
Thank you very much for your reply, I ran the code again this afternoon with your dataset, and it was not much different from the results in the paper table.
你好，看到你使用了这个代码，想请问一下你每个epoch会花费多少时间呢