Keras模型（RNN）预测Floydhub上的奇怪输出

Question

I am very confused about something. As far as I can tell, something is very wrong with my model only when I save it through Floydhub. I'm not even sure if the issue is with Floydhub- but I don't understand what could be happening so I'm blaming it on them for now.

I am trying to run a RNN using Keras that generates text on a word level.

I'm using this dummy dataset (the data has a few issues, but for the purpose of this error report it should work). If you look at the dataset, you'll see it has a fair amount of <newline> words.

When training on my laptop and saving the model, I get predictions like this: <newline> <newline> jerry you the looks <newline> <newline> <newline> jerry you know <newline> <newline> bye the <newline> <newline> <newline> jerry flip a the the to <newline> <newline> jerry elaine you <newline> <newline> jerry elaine it <newline> <newline> jerry <newline> <newline> <newline> elaine have <newline> <newline> you jerry <newline> <newline> jerry her in <newline> <newline> <newline> i just <newline> <newline> <newline> <newline> <newline> he cometh back <newline> <newline> the you <newline> <newline> <newline> <newline> <newline> no of <newline> <newline> elaine <newline> <newline> <newline> <newline> elaine the me <newline> <newline> <newline>

However, when I train via Floydhub (using exactly the same code- only changing paths) and saving the model, I get stuff like this: strengths dont turns hu hosting sittin avoided yayou sittin them tie sittin hu tie turns turns he biography them hereand its battery car afternoon tie into into tie sittin thanks alone turns turns brilliant minute quones shhhhh folks its car turns turns brilliant minute location decided turns turns brilliant biography them sometimes sitting thanks thanks thanks closes turns turns jer grape thursday jerrys jerrys national biography comin turns turns brilliant grape hu drawn minute paper hu probably hu mashed again turns turns jer grape office larry jerrys shop coin lie hescrazylook turns turns jer grape hu decided surprised ive meatloaf

Not a <newline> in sight, just a random selection of words in the vocabulary- no clear pattern that I can see whatsover. Surely, something is wrong with the weights being saved on Floydhub.

If you go to my repository and check the output between training and prediction it should be easy to see the difference.

On the readme you'll see some instructions on how to train or run predictions. Again- everything works if I train and save on my home PC, so I honestly can't figure out what's going on with Floydhub. I'm not even sure if the problem is with Floydhub- so if anyone has any ideas about what could be causing this issue, or how I could more effectively debug the issue , please let me know.

My code can be seen here (rnn.py)

I'm completely stumped :confused:

Thanks

Answer 1

Turns out set() doesn't return a stable order. To fix the issue I used sort() on my words set before the mapping and tokenising started.

The reason it was working on my laptop and not on Floydhub is because set() maintains a stable order when inside the same environment, although of course on Floydhub each time you run it, you run it in a different environment.