search

lovecambi / qebrain

machine translation and quality estimation
BSD 2-Clause "Simplified" License
34 stars 18 forks source link

Label length in prediction #18

Open Venky-JING opened 4 years ago

Venky-JING commented 4 years ago

Hello! I'm sorry, I have a question about the label length when predict the test data.

The predicted tag length seems to be the longest sentence length in each batch, but what I want is the predicted tag of the actual length of each sentence.

This is the comparison between the real label length and the predicted label length by qebrain: real:32 openkiwi:32 qebrain:32 real:27 openkiwi:27 qebrain:32 real:18 openkiwi:18 qebrain:32 real:22 openkiwi:22 qebrain:32 real:14 openkiwi:14 qebrain:32 real:14 openkiwi:14 qebrain:32 real:22 openkiwi:22 qebrain:32 real:22 openkiwi:22 qebrain:32 real:9 openkiwi:9 qebrain:32 real:11 openkiwi:11 qebrain:32 real:20 openkiwi:20 qebrain:32 real:25 openkiwi:25 qebrain:32 real:22 openkiwi:22 qebrain:32 real:13 openkiwi:13 qebrain:32 real:20 openkiwi:20 qebrain:32 real:22 openkiwi:22 qebrain:32 real:19 openkiwi:19 qebrain:58 real:16 openkiwi:16 qebrain:58 real:12 openkiwi:12 qebrain:58 real:12 openkiwi:12 qebrain:58 real:25 openkiwi:25 qebrain:58 real:32 openkiwi:32 qebrain:58 real:17 openkiwi:17 qebrain:58 real:30 openkiwi:30 qebrain:58 real:25 openkiwi:25 qebrain:58 real:26 openkiwi:26 qebrain:58 real:58 openkiwi:58 qebrain:58 real:18 openkiwi:18 qebrain:58 real:24 openkiwi:24 qebrain:58 real:19 openkiwi:19 qebrain:58 real:29 openkiwi:29 qebrain:58 real:13 openkiwi:13 qebrain:58 real:8 openkiwi:8 qebrain:43 real:15 openkiwi:15 qebrain:43 real:19 openkiwi:19 qebrain:43 real:18 openkiwi:18 qebrain:43 real:20 openkiwi:20 qebrain:43 real:43 openkiwi:43 qebrain:43 real:17 openkiwi:17 qebrain:43 real:18 openkiwi:18 qebrain:43 real:15 openkiwi:15 qebrain:43 real:14 openkiwi:14 qebrain:43 real:19 openkiwi:19 qebrain:43 real:28 openkiwi:28 qebrain:43 real:23 openkiwi:23 qebrain:43 real:16 openkiwi:16 qebrain:43 real:12 openkiwi:12 qebrain:43 real:14 openkiwi:14 qebrain:43

How can I deal with this problem? Thank you very much

lovecambi commented 4 years ago

Because the training is batched, it is required to pad short sentences such that the current batch of data have the same length. You need to do some post-processing to remove padded positions.