训练为啥前几个 准确率啥的都是-1呢？是程序的设定吗？

[tyn513]

3%|▎ | 20/725 [00:33<17:52, 1.52s/it]training batch: 20, loss: 29.96524, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.852 6%|▌ | 40/725 [01:04<17:31, 1.53s/it]training batch: 40, loss: 24.01720, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.889 8%|▊ | 60/725 [01:36<17:31, 1.58s/it]training batch: 60, loss: 21.02634, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.821 11%|█ | 80/725 [02:09<17:02, 1.59s/it]training batch: 80, loss: 17.43504, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.859 14%|█▍ | 100/725 [02:42<17:20, 1.66s/it]training batch: 100, loss: 12.46207, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.908 17%|█▋ | 120/725 [03:16<17:41, 1.75s/it]training batch: 120, loss: 9.91527, precision: 1.000 recall: 0.029 f1: 0.057 accuracy: 0.923 19%|█▉ | 140/725 [03:49<15:39, 1.61s/it]training batch: 140, loss: 11.09897, precision: 0.636 recall: 0.146 f1: 0.237 accuracy: 0.915

[lmy86263]

3%|▎ | 20/725 [00:33<17:52, 1.52s/it]training batch: 20, loss: 29.96524, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.852 6%|▌ | 40/725 [01:04<17:31, 1.53s/it]training batch: 40, loss: 24.01720, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.889 8%|▊ | 60/725 [01:36<17:31, 1.58s/it]training batch: 60, loss: 21.02634, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.821 11%|█ | 80/725 [02:09<17:02, 1.59s/it]training batch: 80, loss: 17.43504, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.859 14%|█▍ | 100/725 [02:42<17:20, 1.66s/it]training batch: 100, loss: 12.46207, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.908 17%|█▋ | 120/725 [03:16<17:41, 1.75s/it]training batch: 120, loss: 9.91527, precision: 1.000 recall: 0.029 f1: 0.057 accuracy: 0.923 19%|█▉ | 140/725 [03:49<15:39, 1.61s/it]training batch: 140, loss: 11.09897, precision: 0.636 recall: 0.146 f1: 0.237 accuracy: 0.915

你是哪个模型？遇到了同样的问题。

[lmy86263]

3%|▎ | 20/725 [00:33<17:52, 1.52s/it]training batch: 20, loss: 29.96524, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.852 6%|▌ | 40/725 [01:04<17:31, 1.53s/it]training batch: 40, loss: 24.01720, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.889 8%|▊ | 60/725 [01:36<17:31, 1.58s/it]training batch: 60, loss: 21.02634, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.821 11%|█ | 80/725 [02:09<17:02, 1.59s/it]training batch: 80, loss: 17.43504, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.859 14%|█▍ | 100/725 [02:42<17:20, 1.66s/it]training batch: 100, loss: 12.46207, precision: -1.000 recall: 0.000 f1: -1.000 accuracy: 0.908 17%|█▋ | 120/725 [03:16<17:41, 1.75s/it]training batch: 120, loss: 9.91527, precision: 1.000 recall: 0.029 f1: 0.057 accuracy: 0.923 19%|█▉ | 140/725 [03:49<15:39, 1.61s/it]training batch: 140, loss: 11.09897, precision: 0.636 recall: 0.146 f1: 0.237 accuracy: 0.915

你是哪个模型？遇到了同样的问题。

看了源代码，发现作者使用的是exact match，如果没有match上，则使用初始化值，也就是-1，因此在前几轮中可能由于训练不充分，会出现-1，夸张的是，我用finetune bert+crf的前几个epoch都是-1，所以结论是：代码没有问题（或者，作者可以添加一些友好的提示，不成熟的建议。）

[Kyro-Beluga]

我也是这个问题。但是我都训练了20个epoch了，还是-1。除了batch_size改小了为24，其他都没变。感觉很无奈。

[stanleylsx]

把learning_rate改小，设定为5e-5

[stanleylsx]

已经修改为指标初始为0.0