Where does <unk> come from? about nmt-chatbot HOT 4 CLOSED

<unk> is one of the three special tokens. It basically means what you think - 'unknown'. The purpose for that token is to replace every entity in training that doesn't exist in vocab. Same for inference (when you are "asking" AI).
So every entity that is not included in vocab will be replaced with <unk> and AI will respond with sentences that are including <unk>s. To lower number of that tokens with responses you have to include more entities in vocab files (make "bigger" vocab files).
Also, if you are training your model, you will see a lot of <unk>s. Later you should see them less and less the further in training process.

from nmt-chatbot.

Standard settings are tuned for 4GB of VRAM, so just try to raise 'vocab_size' as much as possible. If you would like to raise it further, you have some possibilities to free some VRAM memory (in exchange for overall model abilities), you can:

• change 'num_units' to 256
• uncomment and change 'batch_size' to 64
• comment out 'encoder_type'

But it's hard to tell how model will change - you will have to train couple of them using diffrent settings and compare results.

from nmt-chatbot.

Thank you!
With the standard settings, how high can I go with vocab size using 8GB VRAM?

from nmt-chatbot.

FYI: vocab size = 100k is not working, 75k is working

from nmt-chatbot.