As Jeremy talks about all the characters, I start to think about revisiting my Cryptography courses. ^^
Someone should write a tool that automatically converts Jeremyâs lecture transcripts into a Q&A format and auto-posts to Stack Overflow.
Basically a programmatic way to find gems in his lectures where his advice is in this format:
âIf you get this error, fix it with this!â
do a wget https://s3.amazonaws.com/text-datasets/nietzsche.txt
Coming back to this - am I right in assuming word based would be more efficient in normal vocabulary tasks while character based would be better for things like code?
Yeap just got it from https://raw.githubusercontent.com/mbernico/lstm_bot/master/lstm_bot/nietzsche.txt
Thanks!
You can combine them as well.
Was mostly watching live in class, which Jupyter notebook was all the character RNN (last notebook)? Couldnât find which one it was in Crestle? I did a git pull
and couldnât figure out which one it was?
lesson6-rnn
Doesnât show up, was I pulling from the wrong repo? Had this problem 2 weeks ago too. =(
Here it is directly⌠https://github.com/fastai/fastai/blob/master/courses/dl1/lesson6-rnn.ipynb
Not sure if helpful
Iâm not using Crestle, hopefully somebody else can help you with that.
Why do we have batch_first
option in rnn? See for ex: http://pytorch.org/docs/0.3.0/nn.html?highlight=lstm#torch.nn.LSTM
In what case one or the other is needed?
Oh ya I think I just had the repo locally on my PC at home and not on this Macbook.
I need to rewatch many parts of this lecture.
How would you do that? And would it improve anything as compared to character based model?
Oh crap yeah now I remember!
Use apply_cats()
I think I only had a single item in my validation set at that point. I meant to point it out, but forgot!