Did anyone tried to reimplement seq2seq translation code using the material presenting in v3 of the course? Seems, all I can find is about fastai v0.7 or something…
See the NLP course, there is a seq2seq notebook github.com/fastai/course-nlp
It uses fastai 1.0
Yes, https://github.com/fastai/course-nlp/blob/master/7-seq2seq-translation.ipynb is what I’ve checked before posting. But it doesn’t use modularized preprocessing that was used in https://github.com/fastai/course-v3/blob/master/nbs/dl2/12_text.ipynb
The later I feel to be much more clean.
Still cannot reproduce.
Can someone take a look at https://github.com/akhavr/fastai-lectures-v3-seq2seq/blob/master/seq2seq.ipynb ? Might be I’m missing something stupid, since the model doesn’t train at all.
Got a deep belief that something is wrong with optimization routines in v3 course.
Gave up, so far, trying to reproduce seq2seq on fastai-dev v2
I’m working on major updates to this including transformer implementations and huggingface integration.
Is there a “work in progress” snapshot somewhere? Because, right now it uses fastai v1, and there’s a seq2seq fastai-v1 notebook here https://github.com/fastai/course-nlp/blob/master/7-seq2seq-translation.ipynb
Not at the moment.