lipiji / SongNet

Code for ACL 2020 paper "Rigid Formats Controlled Text Generation":https://www.aclweb.org/anthology/2020.acl-main.68/
MIT License
230 stars 40 forks source link

Bad results.. #25

Closed pavelxx1 closed 3 years ago

pavelxx1 commented 3 years ago

Hi I trained on small dataset

/content/SongNet
9
2300 667 2599
7
2
9
vocab
done
vocab.size = 1215
batch_acm 99, loss 5.277, acc 0.102, nll 6.265, ppl 86.928, x_acm 1584, lr 0.000002
batch_acm 199, loss 3.566, acc 0.262, nll 4.273, ppl 20.652, x_acm 3179, lr 0.000005
batch_acm 299, loss 2.664, acc 0.323, nll 3.223, ppl 9.619, x_acm 4774, lr 0.000008
batch_acm 399, loss 2.123, acc 0.396, nll 2.580, ppl 6.136, x_acm 6374, lr 0.000010
batch_acm 499, loss 1.807, acc 0.452, nll 2.195, ppl 4.706, x_acm 7969, lr 0.000013
validating...
epoch-3-acm-499 nll= 1.846888825275015 ppl= 3.7909837519747205 count= 667.0
batch_acm 599, loss 1.580, acc 0.502, nll 1.921, ppl 3.879, x_acm 9564, lr 0.000015
batch_acm 699, loss 1.370, acc 0.561, nll 1.673, ppl 3.265, x_acm 11164, lr 0.000018
batch_acm 799, loss 1.151, acc 0.631, nll 1.418, ppl 2.733, x_acm 12759, lr 0.000020
batch_acm 899, loss 0.934, acc 0.701, nll 1.167, ppl 2.290, x_acm 14354, lr 0.000023
batch_acm 999, loss 0.695, acc 0.787, nll 0.890, ppl 1.882, x_acm 15954, lr 0.000025
validating...
epoch-6-acm-999 nll= 0.3973974670427314 ppl= 1.3300093074609851 count= 667.0
batch_acm 1099, loss 0.494, acc 0.856, nll 0.652, ppl 1.589, x_acm 17549, lr 0.000028
batch_acm 1199, loss 0.332, acc 0.913, nll 0.460, ppl 1.384, x_acm 19144, lr 0.000030
batch_acm 1299, loss 0.223, acc 0.945, nll 0.330, ppl 1.260, x_acm 20739, lr 0.000033
batch_acm 1399, loss 0.157, acc 0.966, nll 0.252, ppl 1.192, x_acm 22339, lr 0.000035
batch_acm 1499, loss 0.117, acc 0.975, nll 0.208, ppl 1.156, x_acm 23934, lr 0.000038
validating...
epoch-10-acm-1499 nll= 0.08431253172289664 ppl= 1.060383505013393 count= 667.0
training time: 453sec.

and test result is unreadable text after execute polish.sh with my ckpt epoch10_batch_1499 and my vocab.txt

ps: my editedpolish_tpl.txt

['Gufd<s1>327711<s2>_____,____ менять.______ _____ сейчас. _________ любимый. ______ _____ много.']
0.7558178901672363

result:

Gufd<s1>327711<s2>_____,____ менять.______ _____ сейчас. _________ любимый. ______ _____ много.
<bos>По-ти, мув менять.шозыха сйшаб с</s>
lipiji commented 3 years ago

Sorry that I cannot understand the language. The reasons may be: 1) The dataset is too small, please use a larger one; 2) Somewhere bugs.