kegsay / automata Goto Github PK

View Code? Open in Web Editor NEW

3.0 3.0 2.0 58 KB

A pure Go artificial neural network library

License: Apache License 2.0

Go 100.00%

automata's People

Contributors

Stargazers

Watchers

automata's Issues

Swap to slices over maps where possible

Running a cpu profile on the tests shows:

(pprof) top5
Showing nodes accounting for 8020ms, 65.63% of 12220ms total
Dropped 38 nodes (cum <= 61.10ms)
Showing top 5 nodes out of 39
      flat  flat%   sum%        cum   cum%
    2790ms 22.83% 22.83%     2950ms 24.14%  runtime.mapiternext /usr/local/go/src/runtime/hashmap.go
    2000ms 16.37% 39.20%     2360ms 19.31%  runtime.mapaccess1_fast64 /usr/local/go/src/runtime/hashmap_fast.go
    1320ms 10.80% 50.00%     1770ms 14.48%  runtime.mapassign_fast64 /usr/local/go/src/runtime/hashmap_fast.go
    1100ms  9.00% 59.00%     7540ms 61.70%  github.com/kegsay/automata.(*Neuron).Activate /home/kegsay/go/src/github.com/kegsay/automata/neuron.go
     810ms  6.63% 65.63%      810ms  6.63%  runtime.aeshash64 /usr/local/go/src/runtime/asm_amd64.s

It's telling that map accesses are the top 3, all that hashing... Array accesses would be better.

One way of doing this would be to use the ID types as offsets into a slice. This will require a bit of a rework.

Implement LSTM

This is the whole point of developing this library as it is the creme de la creme of the paper.

Implement .dot support

It'd be nice to port a Network to a .dot file which can then be visualised using dot.

Implement cross entropy coster

The paper does this, so we should too.

Implement cross validation support in the Trainer

It feels a bit cheaty to be giving the network the input it will then need to recall. Leave-one-out cross validation or similar would be a nice improvement.

Recommend Projects