itdxer / neupy Goto Github PK

NeuPy is a Tensorflow based python library for prototyping and building neural networks

License: MIT License

Python 50.68% JavaScript 0.05% HTML 0.10% Shell 0.01% Jupyter Notebook 49.17%

artificial-neural-networks boltzmann-machine deep-learning deep-neural-networks deeplearning hopfield lvq neupy neural-network pnn rbm sofm som tensorflow

neupy's People

Contributors

Stargazers

Watchers

Forkers

wavelets webbo89 yanlinaung iblech zhangscth prabhjotsl wubr2000 ml-ai-nlp-ir sonia2599 mbhatt1 neocher zhdbeng heiyixueren edwardbetts disc5 stczhc albertwy vamsijkrishna lvsun kod3r insertcod3 killedision curtiszimmerman lihaossu drbeco paulhendricks mayblue9 isilicon fernandoferreira-me neo4reo bigr-lab poseidon1214 benjamesbabala chituma110 jinyu0310 kradygithub einsteinjava judit-b hlpassion gokul0625 aiadventures anna-jach webdiscover theerit xnsam fox-glue alibaheri ndn-love anorouzi trainkin alonsopg vishalbelsare alephramos geoblog meelement hbcbh1999 shashank-bhatt-07 aaasss0636 dandanw924 susmithanagaboina nlile hulalazz renlongzheng alexwickstrom nazmulislam025 aishwarya1711 muhammedashraf9244 miyainnyc 4dlcreate shatha2014 ahmadjordan multiplecrashes msdevana dpatchigolla fgdbtkd qinghuajiang smcclure879 wgalvao ruanguoqing buptqq raideryi alvinhaidar temp3rr0r sy-lucky degerli emmettfc chenjunkai2018 amoswish gautam-balachandran none53 wjianxz pichyyyyy rediscovera bickymz dundunbug leo-lulei aliozay zapaishchykova xiaonanguaye trochez

neupy's Issues

Identify shared documentation as separate class

Run memory tests for the algorithms

Format input data in train and prediction procedures.

Possible solution

def train(self, input_data):
    input_data = format_data(input_data)

Test Conjugate Gradient update functions

Integrate ``format_data`` function in ``predict_raw`` method

Problem with quasi-Newton update functions

Add different weight initialisation methods

Modifications for scikit-learn pipeline compatibility

Implement transform and fit_transfom in BaseSkeleton class

Crashes my RAM

I am using oja's rule on dataset of size 400x156300. It seems to crash my RAM. I am not sure what is causing this. Please help.
I have 12 GB of RAM.
Tried using memmap still crash!!

convert memmap and reduce precision

[num_sample,num_feat]=train_data.shape
filename = path.join(mkdtemp(), 'train_data.dat')
memmap_train = np.memmap(filename, dtype='float32', mode='w+', shape=(num_sample,num_feat))
memmap_train[:] = train_data[:]
del train_data,test_data

apply oja's rule

ojanet = algorithms.Oja(minimized_data_size=1250,step=1e-10,verbose=True,show_epoch=1)
ojanet.train(memmap_train, epsilon=1e-3,epochs=10000)
red_train_data = ojanet.predict(memmap_train)
ojanet.plot_errors(logx=False)
pdb.set_trace()

Add test to Weight decay and Weight elimination algorithms that test compatibility with step update algorithms

Integrate sparse matrix module from the SciPy library in the CMAC network.

Review list of property type classes and remove or modify useless

Related to module neupy.core.properties

Change `error` to choice property in the BaseNetwork class

Also there should be possibility to initialize custom error function

Identify required properties in already implemented algorithms

Get rid of `functions` module

Rid of `Feedforward` class

Add more tests for the wolfe search algorithm.

Implement continuous BAM algorithm

Control `show_epoch` option if train epoch runs too fast.

Dropout layer raises exception for Wolfe search optimization

theano.gof.fg.MissingInputError: ("An input of the graph, used to compute Elemwise{sub,no_inplace}(<TensorType(float64, scalar)>, <TensorType(float64, scalar)>), was not provided and not given a value.Use the Theano flag exception_verbosity='high',for more information on this error.", <TensorType(float64, scalar)>)

Possibility to use Oja algorithm for a huge matrices

Add displacement functions for CMAC network

Now only works diagonal displacement.

ReLU derivative function

Hinton diagram shows negative legend even if all values are positive.

Hide legend if there is no negative values in the input matrix

Test all properties

Check all documentation and change it to a new style before release/v0.2.0

Issues with pandas data frames

Add tests to verify that all the algorithms work with pandas data frames and fix the ones that don't work

Remove `use_raw_predict_at_error` option from Feedforward class

Copy input data by default in algorithms that don't transform it

There also must be available option that disable this possibility

Representation for the memory based networks

Define __repr__ method for the memory based networks

Load digit data set for the discrete networks test from the library

Possible use case

from neupy.datasets import load_discrete_digits
digits = load_discrete_digits()

`show_epoch` property shows just 10 epochs

By default for a case when number of epochs is greater or equal to 10

Logging issue in iPython notebook

Problem with verbose mode for ensemble algorithms

Implement functionality that help identify where user terminal support styles and switch it's behaviour in case when it's doesn't support them.

dhnet = algorithms.DiscreteHopfieldNetwork(mode='sync')
dhnet.train(data)
dhnet.energy(otherdata)