Can not reproduce results

[chinmay5]

Hi,

I managed to run the code but, I am struggling to reproduce the results. Here are the logs for the U-Net model based on the default configuration.

total Threshold: 0.300 =====> Evaluation IOU: 0.7384; F1_score: 0.8487; Acc: 0.9050

Thanks.

[WHU-YH-jx]

Thank you for your interest in our work. Considering the points you've raised, I think there are a few areas that may need attention:

1. Data Preprocessing: In our paper, we applied data augmentation to the dataset prior to training. Taking the ER dataset as an example, we performed operations such as rotation and flipping, resulting in a total of 942 training images.
2. Training Strategy: During model training, we utilized a warm-up training strategy, which helps the model learn and converge more effectively.
   I hope this information is helpful to you.

[chinmay5]

Hi,

Thank you for the response.

Unfortunately, I am still unable to reproduce the results.

The ER dataset has 125 samples. I used the data_augmentation.py script to create the augmentations and got 125 * 8 = 1000 training samples.

Do I need to apply the same augmentations on the test samples and create their copies as well? I do not think that was mentioned in the paper, but I may be wrong.

I use the train_unet.py script for my training. I use a step-size of 1 and a window size of 5. The script uses 1000 warmup steps. Does this value need changing?

Thanks.