How to train with Coverage?

[XuemingQiu]

when i have finished the 500k iters training with PG ,which makes the Couvergae is False. The Rouge Score is higher than you provided!
`

ROUGE-1:
rouge_1_f_score: 0.3682 with confidence interval (0.3660, 0.3704)
rouge_1_recall: 0.5194 with confidence interval (0.5166, 0.5225)
rouge_1_precision: 0.2983 with confidence interval (0.2961, 0.3004)

ROUGE-2:
rouge_2_f_score: 0.1588 with confidence interval (0.1568, 0.1608)
rouge_2_recall: 0.2248 with confidence interval (0.2220, 0.2277)
rouge_2_precision: 0.1286 with confidence interval (0.1268, 0.1304)

ROUGE-l:
rouge_l_f_score: 0.3390 with confidence interval (0.3369, 0.3412)
rouge_l_recall: 0.4786 with confidence interval (0.4758, 0.4814)
rouge_l_precision: 0.2746 with confidence interval (0.2724, 0.2766)
But the I continuing training from PG last time model with Coverage = True , and train with 25k , why rouge score is so low? ----------------------------------------
ROUGE-1:
rouge_1_f_score: 0.3670 with confidence interval (0.3650, 0.3689)
rouge_1_recall: 0.5149 with confidence interval (0.5123, 0.5174)
rouge_1_precision: 0.2983 with confidence interval (0.2962, 0.3004)

ROUGE-2:
rouge_2_f_score: 0.1510 with confidence interval (0.1490, 0.1528)
rouge_2_recall: 0.2120 with confidence interval (0.2093, 0.2146)
rouge_2_precision: 0.1229 with confidence interval (0.1213, 0.1246)

ROUGE-l:
rouge_l_f_score: 0.3370 with confidence interval (0.3350, 0.3388)
rouge_l_recall: 0.4729 with confidence interval (0.4704, 0.4754)
rouge_l_precision: 0.2738 with confidence interval (0.2718, 0.2758)
****************************************`

May i some thing wrong to train with coverage?
I check the output , the covergae model output with no repeate, the pg has? but why rouge score is so low?

[Wenjun-Peng]

Do you figure out what cause it? I trained the model for 500k iters with coverage=Fasle, Then truned off coverage for 200k iters, and got results:
{'rouge1_fmeasure': tensor(0.3744),
'rouge1_precision': tensor(0.3811),
'rouge1_recall': tensor(0.3919),

'rouge2_fmeasure': tensor(0.1642),
'rouge2_precision': tensor(0.1683),
'rouge2_recall': tensor(0.1711),

'rougeL_fmeasure': tensor(0.2603),
'rougeL_precision': tensor(0.2646),
'rougeL_recall': tensor(0.2729),

'rougeLsum_fmeasure': tensor(0.2603),
'rougeLsum_precision': tensor(0.2646),
'rougeLsum_recall': tensor(0.2729)}

by the way, I use torchmetrics for measure results instead of pyrouge, but I don't think it will cause inconsistant results