Question about Training Results Discrepancy

[Halleyawoo]

Dear Author,

I followed your setup exactly and preprocessed the LA dataset as instructed. Then, I ran the following command:

bash train.sh -c 0 -e diffusion -t la_0.05 -i 'experiment1' -l 0.01 -w 10 -n 300 -d true

However, the final results I obtained are:

Final Dice of each class: [78.8]  
Final Jaccard of each class: [70.4]  
Final HD95 of each class: [12.5]  
Final ASD of each class: [2.3]  

Final Avg Dice: 78.77±0.0  
Final Avg Jaccard: 70.41±0.0  
Final Avg HD95: 12.54±0.0  
Final Avg ASD: 2.27±0.0  

There is a significant gap between my results and those reported in the paper. According to the training script, the early stopping occurred at epoch 168 when the Dice score was already quite high. However, during validation, the results still differ considerably from the ones in the paper.

Could you please help me understand what might be causing this discrepancy? Did I make any mistakes in the setup?

For reference, I used an A100 GPU and got the following training log:

Evaluation epoch 168, dice: 0.9429011344909668, [0.94290113]
Best eval dice is 0.9532027840614319 in epoch 118
Early stop.

The checkpoint was loaded from:

./logs/Exp_SSL_LA_0.05/diffusionexperiment1/fold1/
load checkpoint from ./logs/Exp_SSL_LA_0.05/diffusionexperiment1/fold1/ckpts/best_model.pth

I would really appreciate any guidance you could provide!

Best regards

[hhjjaaa]

There are many unstable factors in this code, and I found such problems in the abdominal data set of this paper. The result reported in the paper was 60, but I made the result of 62 and also made the result of 60, which is even far lower than 60. There are roughly such indicators as 40.

[hhjjaaa]

I found that batch size also has a certain effect on the result in the code, what is your opinion