Inference on Word CNN about tcn HOT 4 CLOSED

You mean the case where you want to (for instance) generate the next word/character? You probably don't want to use seq_len=1, but instead an amount that is reasonable for a good history size. For instance, you can use seq_len=100, so that the previous 100 tokens will contribute to the next prediction (which is output[:,:,-1]).

Let's say you have input x with dimension (N, L) where L=100 now. Then evaluating model(x) will give you an output z of (N, L) as well, where z_i is TCN's prediction of the next token after x_i. TCN's signature is very similar to that of the RNN/LSTM in PyTorch.

Hope this helps :-)

from tcn.

@jerrybai1995 yes this is the case. Thanks! Regarding the words split for next token, supposed to set seq_len=100 it means the next 100 chars i.e. it could break on a subword?

from tcn.

Setting seq_len=100 means use (up to) the previous 100 tokens (word or char) for prediction. If you are doing character level language modeling, then yes, we could be breaking on a subword.

from tcn.

@jerrybai1995 thanks ok so this means I should use the word_cnn rather than the char_cnn to avoid subwords. Thank you.

from tcn.