mwojc at p.lodz.pl
Tue Aug 9 02:56:31 CEST 2011
ffnet version 0.7 has been released and is available for download at:
This release contains couple of important changes:
- neural network can be trained now using the power of multi-processor
systems (see example mptrain.py)
- attributes necessary for calculation of network derivatives
are now generated only on demand;
- data normalization limits are not changed when re-trainig with new
data set; net.renormalize=True have to be set first;
- compatibility with newest versions of numpy, scipy and networkx
- support for *export to java* and *drawing network with drawffnet*
Basic API is left almost untouched. Exactly the same trainig scripts
as for older versions should work without problems.
What is ffnet?
ffnet is a fast and easy-to-use feed-forward neural
network training solution for python.
Unique features present in ffnet
1. Any network connectivity without cycles is allowed.
2. Training can be performed with use of several optimization
schemes including: standard backpropagation with momentum, rprop,
conjugate gradient, bfgs, tnc (with multiprocessing)
and genetic alorithm based optimization.
3. There is access to exact partial derivatives of network outputs
vs. its inputs.
4. Automatic normalization of data.
Basic assumptions and limitations
1. Network has feed-forward architecture.
2. Input units have identity activation function,
all other units have sigmoid activation function.
3. Provided data are automatically normalized, both input and output,
with a linear mapping to the range (0.15, 0.85).
Each input and output is treated separately (i.e. linear map is
unique for each input and output).
4. Function minimized during training is a sum of squared errors
of each output for each training pattern.
Excellent computational performance is achieved implementing core
functions in fortran 77 and wrapping them with f2py. ffnet outstands
in performance pure python training packages and is competitive to
'compiled language' software. Incorporation of multiprocessing capabilities
(tnc algorithm so far) makes ffnet ideal for large scale (really!)
problems. Moreover, a trained network can be exported to fortran sources,
compiled and called from many programming languages.
Basic usage of the package is outlined below.
See description of ffnet module and its functions
(and especially ffnet class) for detailed explanations.
>>> from ffnet import ffnet, mlgraph, savenet, loadnet, exportnet
>>> conec = mlgraph( (2,2,1) )
>>> net = ffnet(conec)
>>> input = [ [0.,0.], [0.,1.], [1.,0.], [1.,1.] ]
>>> target = [ [1.], [0.], [0.], [1.] ]
>>> net.train_tnc(input, target, maxfun = 1000)
>>> net.test(input, target, iprint = 2)
>>> savenet(net, "xor.net")
>>> exportnet(net, "xor.f")
>>> net = loadnet("xor.net")
>>> answer = net( [ 0., 0. ] )
>>> partial_derivatives = net.derivative( [ 0., 0. ] )
Usage examples with full description can be found in
examples directory of the source distribution.
More information about the Python-announce-list