Excalibur

@haibarasfs Currently the best performance I can get is F1 64 EM 50 (with long time training ......) I haven't figured out the difference. Ideally, it should reach around 77...

@haibarasfs You mean memory explode? If you use 1080ti card, then the batch_size maybe need to be smaller...

@andy840314 What do you think that may cause the 5.0 point gap between your implementation and that of Tensorflow QANet? I am also trying to reach the performance.

@andy840314 @hengruo I tested andy's code, it can achieve F1 74.128853 EM 62.707499 in 14 epochs, and F1 70.157651 EM 58.185461 in 4 epochs. I compared andy's version with hengruo's...

@hackiey That's great! I will test it

After fixed some data processing issue, my current performance is: Epoch 1: train accuracy: 0.8593 dev accuracy: 0.7004, train_loss: 13.87, dev_loss: 20.79 Epoch 2: train accuracy: 0.9206 dev accuracy: 0.7157,...

具体的运行时间多长记不清了，很多年前跑的实验。但是我印象中不会超过两三天。建议试试只怕一小部分数据看看时间，等比例估计一下。处理好的数据手头上也没有，时间太久远了 difonjohaiv ***@***.***> 于2023年3月19日周日 03:23写道： > 数据已经跑了36小时了，还是没有得到输出文件。请问有朋友跑完feature_extractor.py并获得输出文件吗 > > — > Reply to this email directly, view it on GitHub > , > or unsubscribe > > . > You...

打错字了，只跑一小部分数据预估时间。 Steve Zhang ***@***.***> 于2023年3月19日周日 11:56写道： > > 具体的运行时间多长记不清了，很多年前跑的实验。但是我印象中不会超过两三天。建议试试只怕一小部分数据看看时间，等比例估计一下。处理好的数据手头上也没有，时间太久远了 > > difonjohaiv ***@***.***> 于2023年3月19日周日 03:23写道： > >> 数据已经跑了36小时了，还是没有得到输出文件。请问有朋友跑完feature_extractor.py并获得输出文件吗 >> >> — >> Reply to this email directly, view it on...

@calliwen I will update the README asap.