A question about the training details

About one month ago, we trained the baseline with your released code (single GPU) and achieve almost the same performance as following:

• xxxx｜ PQ ｜mPQ+ ｜multi_r2
• ours ｜ 0.6138 ｜0.4937 ｜0.8369
• yours ｜ 0.6149 ｜ 0.4998 ｜0.8585

Considering the controllability for the code, we reconstruct the code with PyTorch-Lightning and use two GPUs and other parameters keep the same as your released baseline. But the result is disappointing, the mPQ down about 3 percent to 46.70 .

Now, we have checked every possible cause of the problem and speculated that it was due to the multi GPUs training.

So we were wondering how many GPUs you used in training. This is very important for us to improve our performance

Looking forward to your reply, thank you very much！

Hi!

Good news that you pretty matched the performance of our baseline. We trained the model with 2 GPUs to get the baseline result. Remember, this result was only on just a small validation set. The real performance measure will be on the final test set at the end, which will be much larger and more reliable. Good luck with finalising your algorithm - not long until submission on the preliminary test set opens!

Simon