Deep learning in pythonista .?

dgelessus

Most Python machine learning libraries are written in native code (C, C++, Assembly, Fortran, etc.), so you cannot install them in Pythonista (because of iOS restrictions). npdl is a bit different, it uses only pure Python and numpy, so it should be usable in Pythonista.

If your code hangs on the model.fit call, maybe it just takes a while to run? Depending on what your device is, your code will run much slower than on a normal computer, especially for code that requires a lot of computation. Try letting it run for a few minutes and see if anything happens. (If your device gets warm, you can tell that it's still doing something. :))

Note: I know nothing about machine learning, so I can't help very much, sorry.

Kush

Yeah I tried it on my iPhone 5s and let it run for 3-4 minutes .. i just wanted to know if it was actually running and working on more powerful devices ir not .?

zrzka

Post some script (+ what to install), so, I can run it on iPad Pro and will let you know for how long it runs.

Kush

Well i just pip installed npdl
Then git cloned this repo:
https://github.com/oujago/NumpyDL
And ran the lstm-character.py in examples folder
But it hung on train model
I figured there were too many characters in tiny shakespeare.txt
So i reduced them but it still didnt went past the train model..
However its working fine in my pc .. still took a lot of time though

zrzka

Here're results from the iPad Pro 2nd gen 12.9". I had to change max iterations to 10, it took much longer with 100.

Started: 2017-09-18 10:50:13.803954
data has 1115389 charactres,  65nique.
Building model ...
Train model ...
iter 1, train-[loss 4.1734, acc 0.0362]; 
iter 2, train-[loss 4.1724, acc 0.0362]; 
iter 3, train-[loss 4.1714, acc 0.0375]; 
iter 4, train-[loss 4.1704, acc 0.0375]; 
iter 5, train-[loss 4.1695, acc 0.0400]; 
iter 6, train-[loss 4.1685, acc 0.0413]; 
iter 7, train-[loss 4.1675, acc 0.0437]; 
iter 8, train-[loss 4.1665, acc 0.0437]; 
iter 9, train-[loss 4.1655, acc 0.0462]; 
iter 10, train-[loss 4.1646, acc 0.0475]; 
Started: 2017-09-18 10:50:13.803954, Ended: 2017-09-18 11:21:58.137529
Duration: 1904.333575

Approximately 32 minutes. You should run these tasks on a computer as @dgelessus already mentioned. These pure Python implementations will be always slow.

Kush

Oh wow thanks .. decreasing the max iter definitely helped and i also found out that decreasing the n_out to 30(from 300) ie reducing the no of nodes made it a lot faster

zrzka

Just run it on my desktop (MacBook Pro (15-inch, Late 2016), 2,9 GHz Intel Core i7, 16 GB 2133 MHz LPDDR3) for fun ...

/Users/robertvojta/anaconda/bin/python3 /Users/robertvojta/Work/purposefly/NumpyDL/examples/lstm-character-lm.py
Started: 2017-09-18 22:14:10.355194
data has 1115389 charactres,  65nique.
Building model ...
Train model ...
iter 1, train-[loss 4.1730, acc 0.0262]; 
iter 2, train-[loss 4.1716, acc 0.0262]; 
iter 3, train-[loss 4.1702, acc 0.0262]; 
iter 4, train-[loss 4.1688, acc 0.0262]; 
iter 5, train-[loss 4.1674, acc 0.0250]; 
iter 6, train-[loss 4.1660, acc 0.0262]; 
iter 7, train-[loss 4.1646, acc 0.0262]; 
iter 8, train-[loss 4.1632, acc 0.0275]; 
iter 9, train-[loss 4.1619, acc 0.0288]; 
iter 10, train-[loss 4.1605, acc 0.0300]; 
Started: 2017-09-18 22:14:10.355194, Ended: 2017-09-18 22:15:04.876571
Duration: 54.521377

... and the result is less than one minute, 32x faster :) Okay, it's not just about desktop vs iPad, but also pure Python vs optimized NumPy libs, ...

Kush

Haha yeah pure python libraries are indeed slower

ellie_ff1493

iOS 11 has a deep learning api in it

chibill

For pretrained models.