Suggestion to make the model Converge for a Sentence Embedding

[dataislife]

Hi,

Instead of using Bow or Boc, I was thinking it could be possible to map that for sentence embeddings. Any suggestions to get the algorithm to Converge?
At the moment, I am struggling to get it converging.
After 93 epochs, with python main.py --model binary_bow --config 00, we have the following results which are quite low:

epoch: 1 | loss: 2208.521 | -logp(x|z): 2203.606 | kldiv: 4.915 | time: 0 min., 1 sec.
F1. 0.129 | acc. 0.807 | prec.: 0.154 | rec. 0.111
........
epoch: 50 | loss: 1909.999 | -logp(x|z): 1896.300 | kldiv: 13.700 | time: 1 min., 54 sec.
F1. 0.067 | acc. 0.782 | prec.: 0.075 | rec. 0.061
.........
epoch: 93 | loss: 1886.401 | -logp(x|z): 1873.151 | kldiv: 13.250 | time: 3 min., 19 sec.
F1. 0.042 | acc. 0.785 | prec.: 0.049 | rec. 0.036

Any feedback is welcome :)