In [1]:

    

%pylab inline


    

    




Populating the interactive namespace from numpy and matplotlib

In [2]:

    

import sys
sys.version
sys.path += ["/usr/local/lib/python2.7/dist-packages"]


    

In [3]:

    

import clstm
import numpy


    

In [4]:

    

inputs = clstm.Sequence()
outputs = clstm.Sequence()
targets = clstm.Sequence()


    

In [5]:

    

a = array(randn(100,11,7),'f')
clstm.sequence_of_array(inputs,a)
print inputs.size(), inputs.depth(), inputs.batchsize()
#clstm.array_of_sequence(b,inputs)
#print b.shape


    

    




100 11 7

In [6]:

    

b = numpy.empty_like(a)
clstm.array_of_sequence(b,inputs)


    

In [7]:

    

b = numpy.empty((1,1,1),'f')
clstm.array_of_sequence(b,inputs)
b.shape


    

    
Out[7]:





(100, 11, 7)

In [8]:

    

net = clstm.make_net_init("bidi","ninput=11:nhidden=20:noutput=13")
net


    

    
Out[8]:





<clstm.INetwork; proxy of <Swig Object of type 'std::shared_ptr< INetwork > *' at 0x7fbbba8e8c30> >

In [9]:

    

print clstm.network_info(net)


    

    




.Stacked: 0.000100 0.900000 11 13
.Stacked.Parallel: 0.000100 0.900000 11 40
.Stacked.Parallel.NPLSTM_SigmoidTanhTanh: 0.000100 0.900000 11 20
.Stacked.Parallel.Reversed: 0.000100 0.900000 11 20
.Stacked.Parallel.Reversed.NPLSTM_SigmoidTanhTanh: 0.000100 0.900000 11 20
.Stacked.SoftmaxLayer: 0.000100 0.900000 40 13

In [10]:

    

temp = clstm.make_layer("LSTM")
temp


    

In [11]:

    

clstm.sequence_of_array(net.inputs, a)
print net.inputs.size(), net.inputs.depth(), net.inputs.batchsize()


    

    




100 11 7

In [12]:

    

print net.outputs.size(), net.outputs.depth(), net.outputs.batchsize()
net.forward()
print net.outputs.size(), net.outputs.depth(), net.outputs.batchsize()


    

    




0 -1 -1
100 13 7

In [13]:

    

deltas = zeros(0,'f')
clstm.array_of_sequence(deltas,net.outputs)
deltas *= 0
clstm.sequence_of_array(net.d_outputs,deltas)


    

In [14]:

    

print net.d_inputs.size(), net.d_inputs.depth(), net.d_inputs.batchsize()
net.backward()
print net.d_inputs.size(), net.d_inputs.depth(), net.d_inputs.batchsize()


    

    




0 -1 -1
100 11 7

In [ ]: