crystal-fann/Lobby

Support and Discussion for crystal-fann shard

Hugo Abonizio

@hugoabonizio

hi @bararchy!

I'm having an issue with crystal-fann, can you help me with that?
I'm trying to train a Network::Standard on iris dataset, but my model always evaluate to the same class, it's not learning the patterns
am I doing something wrong?

https://gist.github.com/anonymous/6acbb9787167f72fb2ea2b5ff17a30a9

running it on neural_net.cr sample gives 96% accuracy

but with crystal-fann it's about 33% (always the same class)

Bar Hofesh

@bararchy

Hi @hugoabonizio I would strongly suggest not using single iiteration learning method, and instead use batch with the Data structure example , I would also suggest playing around with different activation functions , FANNs default to liner , which is in 90% of the cases not what you want ;)

Also, if you prefer the lib to do most things for you automatically, try the Cascade network, for me on different data sets it gave best results (keep in mind the net will build its own hidden network, so only input and output are relevant.)

Please update me on how it go, I'll also try your example

Hugo Abonizio

@hugoabonizio

@bararchy thanks for answering!

I was using batch training, like that:

ann.train_batch(data, {:max_runs => 500000, :desired_mse => 0.01_f64, :log_each => 1000})

but changed to single iteration to make it more similar to the neural_net sample I was testing

I think the default activation function is sigmoid (http://leenissen.dk/fann/html/files2/advancedusage-txt.html), but I'm testing others anyway

I was trying the same parameters as found in https://genbattle.bitbucket.io/blog/2016/03/19/Simple-Artificial-Neural-Networks-with-FANN-and-C/

Bar Hofesh

@bararchy

@hugoabonizio Does using Cascade and batch gives you a better result ?

Hugo Abonizio

@hugoabonizio

I'll try Cascade later and let you know the results!

Bar Hofesh

@bararchy

:thumbsup: thanks

Oh, one more thing, which source did you use to get the FANN lib from ?

Hugo Abonizio

@hugoabonizio

hi @bararchy

I followed this installation http://leenissen.dk/fann/wp/help/installing-fann/ and cloned from github repo

on lib_fann.cr file I was replacing @[Link("doublefann")] with @[Link("fann")] to experiment

with doublefann it's giving 96% accuracy now!

(but sometimes it's 30%, I think that's a problem with the random weights initialization)

I think it's fine now, thanks!

Bar Hofesh

@bararchy

Oh I see XD
well, as we binded against douablefann we changed Float64\Float32 ratio to acompany this :) so using the non double version indeed introduces issues

thanks for the update :) @hugoabonizio

Hugo Abonizio

@hugoabonizio

I didn't thought about that :laughing:

thanks for your help!

Martyn Jago

@mjago

Hi @bararchy The example on github readme seems to fail for me as in I get a result > 0.1 ~ any ideas?

Max epochs     8000. Desired error: 0.0010000000.
Epochs            1. Current error: 0.5189931989. Bit fail 2.
Epochs         1000. Current error: 0.2500000000. Bit fail 4.
Epochs         2000. Current error: 0.2500000000. Bit fail 4.
Epochs         3000. Current error: 0.2500000596. Bit fail 4.
Epochs         4000. Current error: 0.2500000000. Bit fail 4.
Epochs         5000. Current error: 0.2500000000. Bit fail 4.
Epochs         6000. Current error: 0.2500000596. Bit fail 4.
Epochs         7000. Current error: 0.2500000000. Bit fail 4.
Epochs         8000. Current error: 0.2500000596. Bit fail 4.
[0.5000107397550064]
false

Bar Hofesh

@bararchy

@mjago hmmm it seems you get "stuck" at an error rate and not getting out of it.

This could be because of few issues, Have you compiled and used doublefann? did you used the latest version of crystal-fann ?

specs are passing for me on HEAD

does running crystal spec works for you?

Bar Hofesh

@bararchy

Also, a quick question, which of the examples didn't work?

Martyn Jago

@mjago

Specs work

Finished in 71.3 milliseconds
14 examples, 0 failures, 0 errors, 0 pending

I compiled github/libfann head

Bar Hofesh

@bararchy

hm... running specs on master shows

Fann::Network
  initializes
works
Fann::Network
  initializes standard network
  initializes cascade network
  free memory
  trains on single data
  Shows MSE
  trains and evaluate single data
  train on batchMax epochs     8000. Desired error: 0.0010000000.
Epochs            1. Current error: 0.2745950818. Bit fail 2.
Epochs           91. Current error: 0.0009867755. Bit fail 0.
  train on batch
  train on cascadeMax neurons 500. Desired error: 0.001000
Neurons       0. Current error: 0.250000. Total error:  1.0000. Epochs    51. Bit fail   4
Neurons       1. Current error: 0.000746. Total error:  0.0030. Epochs   117. Bit fail   0. candidate steepness 1.00. function FANN_GAUSSIAN_SYMMETRIC
Train outputs    Current error: 0.000000. Epochs    123
  train on cascade

Finished in 3.2 milliseconds
10 examples, 0 failures, 0 errors, 0 pending

So only 10 examples

how did you get 14?

are you using latest version of Crystal-fann?

Martyn Jago

@mjago

:smile:

Ok so I was using a shard to get the error, and cloned your repo to run the specs

Bar Hofesh

@bararchy

Ok, so I'll just make sure again, are you using latest crystal-fann ?

maybe the shard isn't using latest crystal-fann version?

Martyn Jago

@mjago

NeuraLegion/crystal-fann

shard is using that with branch: master

I get this .......Max epochs 8000. Desired error: 0.0010000000.
Epochs 1. Current error: 0.3286948800. Bit fail 2.
Epochs 63. Current error: 0.0009598084. Bit fail 0.
.Max neurons 500. Desired error: 0.001000
Neurons 0. Current error: 0.250000. Total error: 1.0000. Epochs 51. Bit fail 4
Neurons 1. Current error: 0.000618. Total error: 0.0025. Epochs 116. Bit fail 0. candidate steepness 0.25. function FANN_GAUSSIAN_SYMMETRIC
Train outputs Current error: 0.000000. Epochs 120
......

Finished in 79.43 milliseconds
14 examples, 0 failures, 0 errors, 0 pending
Sun Dec 17: crystal-fann/ >
```

I hate gitter sometimes

Bar Hofesh

@bararchy

@mjago You're right, it was my bad, I wans't on latest

can you tell me which of the README examples wont work for you?

basiclly all of them are in the specs, so it's wierd they work there but not in the code

Martyn Jago

@mjago

# Work on array of test data (batch)

Where communities thrive

People

Repo info

Activity