zlinao/variational-transformer

Variational Transformers for Diverse Response Generation

variational-transformer's Issues

print()s not printing

Hey,

For some reason I can't seem to run print() functions through the code. I've isolated the problem to this being the case when save_path is included as an argument. For example:

python3 main.py \
--model trs --emb_dim 300 --hidden_dim 300 \
--hop 4 --heads 4 --cuda --batch_size 128 \
--lr 0.001 --pretrain_emb --kl_ceiling 0.48 \
--aux_ceiling 1 --full_kl_step 20000 \
--save_path save/trs_new_bow_batch/ > save/trs_new_bow_batch/out.txt \

Won't print() anything in the code, whereas:

python3 main.py \
--model trs --emb_dim 300 --hidden_dim 300 \
--hop 4 --heads 4 --cuda --batch_size 128 \
--lr 0.001 --pretrain_emb --kl_ceiling 0.48 \
--aux_ceiling 1 --full_kl_step 20000 \
#--save_path save/trs_new_bow_batch/ > save/trs_new_bow_batch/out.txt \

Will run print() correctly.

Has anyone else run into this issue? Is there any stdout/buffer flushing that could be playing this up? I'm on Ubuntu running PyTorch 1.7.1

Is PRETRAIN necessary?

Hi,
I notice that SVT is trained after loading the parameters of a pretrained model (including encoder & decoder).
I am curious about if pretraining is necessary. Had you tried train SVT from scratch?
What's the difference between these two training schemes?

---

如果您能看懂中文，希望可以加微信高效沟通，我的微信号 He_2262，感激不尽！

maybe i find a bug

when adding the data in data/persona/*_self_original.txt，the context is not fit to the responce.

There are many utterences on each line. This code connects the first few dialogues on line n with the first few dialogues on line n + 1, but the former is not the context of the latter, ignoring so many information(or context) on line n.

How to test with a specific epoch

Thank you for releasing your code!
In utils/config.py, there is a parameter "--test", but it cannot be used in main.py to calculate the metrics in Line 36&37.
How to test the model with a specific given epoch?
Thx