def distance(target, source, insertcost, deletecost, replacecost):
n = len(target) + 1
m = len(source) + 1
# set up dist and initialize values
dist = [[0 for j in range(m)] for i in range(n)]
for i in range(1, n):
dist[i][0] = dist[i - 1][0] + insertcost
for j in range(1, m):
dist[0][j] = dist[0][j - 1] + deletecost
# align source and target strings
for j in range(1, m):
for i in range(1, n):
inscost = insertcost + dist[i - 1][j]
delcost = deletecost + dist[i][j - 1]
if (source[j - 1] == target[i - 1]): add = 0
else: add = replacecost
substcost = add + dist[i - 1][j - 1]
dist[i][j] = min(inscost, delcost, substcost)
# save a hinton plot of the distance matrix
normdist = [[(dist[i][j] / dist[n - 1][m - 1]) - 0.5 for j in range(m)]
for i in range(n)]
hinton.hinton(numpy.array(normdist))
plt.title("Distance matrix")
plt.savefig('distance_hinton.png', format='png')
# return min edit distance
return dist[n - 1][m - 1]
def clear():
# einzelne umgebungen:
global gesetzt
global noch_da
global pfad
pfad = ""
noch_da = 4
gesetzt = False
global rejection_sam_var
rejection_sam_var = [graph, "", "", 20]
global schwellwertffn
global hintonnumberffn
global alphaffn
global ownexample
ownexample = ""
schwellwertffn.set(2.5)
hintonnumberffn.set(20)
alphaffn.set(0.2)
# fuer alle:
anzeigefenster.delete("all")
printfenster.delete(1.0, END)
if not error:
zeichne.draw(graph, "", "")
umgebung = graph.__class__.__name__
if umgebung == "FFNetwork":
matrix = np.empty([15, 15]) # fuer hinton diagramm
#kantengewichte wieder zufaellig machen!
for node in graph.get_nodes():
for edge in node.get_edges():
weight = random.randint(0, 100) / 100
edge.set_weight(weight)
# matrix fuellen:
for zeile in range(15):
for spalte in range(15):
matrix[zeile][spalte] = 0
for node in graph.get_nodes():
name = node.name()
for edge in node.get_edges():
dest = edge.end().name()
weight = edge.weight
# ueberpruefung worein es geschireben wird:
if name[0] == "I":
oben = int(name[1])
elif name[0] == "H":
oben = int(name[1]) + 9
if dest[0] == "H":
runter = int(dest[1]) + 9
else:
runter = 14
matrix[oben][14 - runter] = weight
matrix[runter][14 - oben] = weight
hinton.hinton(matrix)
fig = plt.figure()
plt.subplot(311)
plt.imshow(mean_aligned)
plt.subplot(312)
plt.imshow(mean_unaligned)
plt.subplot(313)
plt.plot(np.sum(mean_aligned,axis=0),label = 'Aligned')
plt.plot(np.sum(mean_unaligned,axis=0), label = 'Unaligned')
plt.legend()
fig.savefig(folder_name + '/' + 'var_%ist_mean.png' % (model_num_states))
plt.close(fig)
# Plot neuron emissions
fig = plt.figure()
hinton(p_emissions)
fig.savefig(folder_name + '/' + 'var_emissions_%ist.png' % model_num_states)
plt.close(fig)
print('file%i, %s, model%i, taste%i' %(file,cond_dir,model_num_states,taste))
# _____ _ _ _ _____
# / ____| | | | | | | / ____|
# | | __ _| | ___ _ _| | __ _| |_ ___ | | ___ _ __ _ __
# | | / _` | |/ __| | | | |/ _` | __/ _ \ | | / _ \| '__| '__|
# | |___| (_| | | (__| |_| | | (_| | || __/ | |___| (_) | | | |
# \_____\__,_|_|\___|\__,_|_|\__,_|\__\___| \_____\___/|_| |_|
#
# Start -> tastes x neurons x trials x time
# For every time bin, correlate firing across all trials of a neuron with taste ranks -> neurons x time
# e.g. 15 data points for every palatability rank if 15 trials per taste
# Average ABSOLUTE correlations across neurons -> time
parameters=cg.parameters,
step_rule=Scale(learning_rate=0.1))
train_set = H5PYDataset('mushrooms.hdf5', which_sets=('train',))
train_stream = DataStream.default_stream(
train_set, iteration_scheme=SequentialScheme(
train_set.num_examples, batch_size=128))
test_set = H5PYDataset('mushrooms.hdf5', which_sets=('test',))
test_stream = DataStream.default_stream(
test_set, iteration_scheme=SequentialScheme(
test_set.num_examples, batch_size=128))
main = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
FinishAfter(after_n_epochs=10),
Printing(),
TrainingDataMonitoring([cost, error_rate], after_batch=True, prefix='train'),
DataStreamMonitoring([cost, error_rate], after_batch=True, data_stream=test_stream, prefix='test'),
Plot('Train',
channels=[['train_cost', 'test_cost'], ['train_error_rate', 'test_error_rate']])
])
main.run()
hinton(W1.get_value())
hinton(W2.get_value())
### MAP Outputs ###
alpha, beta, scaling, expected_latent_state, expected_latent_state_pair = model_MAP.E_step(
)
# Save figures in appropriate directories
for i in range(data.shape[0]):
fig = plt.figure()
raster(data=data[i, :],
expected_latent_state=expected_latent_state[:,
i, :])
fig.savefig(folder_name + '/' + '%i_map_%ist.png' %
(i, model_num_states))
plt.close(fig)
fig = plt.figure()
hinton(model_MAP.p_transitions.T)
plt.title('Log_lik = %f' % model_MAP.log_posterior[-1])
plt.suptitle('Model converged = ' + str(model_MAP.converged))
fig.savefig(folder_name + '/' +
'hinton_map_%ist.png' % model_num_states)
plt.close(fig)
# Save data in appropriate spots in HDF5 file
node_name = '/map_hmm/%s/taste_%i/states_%i' % (
cond_dir, taste, model_num_states)
try:
hf5.remove_node(node_name, recursive=True)
except:
pass
threshold=1e-4,
n_cpu=mp.cpu_count())
### MAP Outputs ###
alpha, beta, scaling, expected_latent_state, expected_latent_state_pair = model_MAP.E_step(
)
for i in range(data.shape[0]):
fig = plt.figure()
raster(data[i, :],
expected_latent_state=expected_latent_state[:, i, :])
plt.savefig(plot_dir + '/' + 'clust%i_%i_map_%ist.png' %
(cluster, i, model_num_states))
plt.close(fig)
plt.figure()
hinton(model_MAP.p_transitions)
plt.title('log_post = %f' % model_MAP.log_posterior[-1])
plt.suptitle('Model converged = ' + str(model_MAP.converged))
plt.savefig(plot_dir + '/' + 'transitions_map_clust%i_%ist.png' %
(cluster, model_num_states))
plt.close()
plt.figure()
hinton(model_MAP.p_emissions)
plt.title('log_post = %f' % model_MAP.log_posterior[-1])
plt.suptitle('Model converged = ' + str(model_MAP.converged))
plt.savefig(plot_dir + '/' + 'emissions_map_clust%i_%ist.png' %
(cluster, model_num_states))
plt.close()
plt.figure()
A[idx,x_theory.index(temp[1])] = -1/4
A[x_theory.index(temp[1]),idx] = -1/3 * 1/Unq[temp[1]] # for u_k, Partial fu
B[idx,0] = 1/4 * (Na[temp[0]]/maxNa + vq[temp[0]]/maxvq + avgsumva[temp[0]]/maxavgsumva)
else:
## Other PostTypeId
A[idx,idx] = 1
idx += 1
for user in ulist:
A[idx,idx] = 1
B[idx,0] = 0 # 1/3 * (repu[user.split(':')[0]]/maxrepu)
#B[idx,0] = 1/3 * (repu[user.split(':')[0]]/maxrepu)
idx += 1
if hinplot == 1:
hinton(A, 'Amat')
hinton(B, 'Bvec')
solution = np.linalg.solve(A,B)
out = dict(zip(x_theory,solution))
#pprint.pprint(out)
if hinplot == 1:
hinton(solution, 'Sol')
## Analysing results
ans = {}
ques = {}
uuser = {}
userQ = {}
anslist = []
queslist = []
batch_size=128))
test_set = H5PYDataset('mushrooms.hdf5', which_sets=('test', ))
test_stream = DataStream.default_stream(test_set,
iteration_scheme=SequentialScheme(
test_set.num_examples,
batch_size=128))
main = MainLoop(model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
FinishAfter(after_n_epochs=10),
Printing(),
TrainingDataMonitoring([cost, error_rate],
after_batch=True,
prefix='train'),
DataStreamMonitoring([cost, error_rate],
after_batch=True,
data_stream=test_stream,
prefix='test'),
Plot('Train',
channels=[['train_cost', 'test_cost'],
['train_error_rate', 'test_error_rate']])
])
main.run()
hinton(W1.get_value())
hinton(W2.get_value())
def PCA_missing_data(plot=True):
#Principal Component Analysis, with randomly missing data
q = 2 #latent dimension
d = 5 #observation dimension
N = 200
niters = 200
Nmissing = 100
true_W = np.random.randn(d,q)
true_Z = np.random.randn(N,q)
true_mean = np.random.randn(d,1)
true_prec = 20.
Xdata_full = np.dot(true_Z,true_W.T) + true_mean.T
Xdata_observed = Xdata_full + np.random.randn(N,d)*np.sqrt(1./true_prec)
#erase some data
missing_index_i = np.argsort(np.random.randn(N))[:Nmissing]
missing_index_j = np.random.multinomial(1,np.ones(d)/d,Nmissing).nonzero()[1]
Xdata = Xdata_observed.copy()
Xdata[missing_index_i,missing_index_j] = np.nan
#set up the problem...
Ws = [nodes.Gaussian(d,np.zeros((d,1)),np.eye(d)*1e-3) for i in range(q)]
W = nodes.hstack(Ws)
Mu = nodes.Gaussian(d,np.zeros((d,1)),np.eye(d)*1e-3)
Beta = nodes.Gamma(d,1e-3,1e-3)
Zs = [nodes.Gaussian(q,np.zeros((q,1)),np.eye(q)) for i in range(N)]
Xs = [nodes.Gaussian(d,W*z+Mu,Beta) for z in Zs]
[xnode.observe(xval.reshape(d,1)) for xnode,xval in zip(Xs,Xdata)]
#make a network object
net = Network()
net.addnode(W)
net.fetch_network()# automagically fetches all of the other nodes...
#infer!
net.learn(100)
#plot
if plot:
import pylab
import hinton
#compare true and learned W
Qtrue,Rtrue = np.linalg.qr(true_W)
Qlearn,Rlearn = np.linalg.qr(W.pass_down_Ex())
pylab.figure();pylab.title('True W')
hinton.hinton(Qtrue)
pylab.figure();pylab.title('E[W]')
hinton.hinton(Qlearn)
if q==2:#plot the latent variables
pylab.figure();pylab.title('true Z')
pylab.scatter(true_Z[:,0],true_Z[:,1],50,true_Z[:,0])
pylab.figure();pylab.title('learned Z')
learned_Z = np.hstack([z.pass_down_Ex() for z in Zs]).T
pylab.scatter(learned_Z[:,0],learned_Z[:,1],50,true_Z[:,0])
#recovered X mean
X_rec = np.hstack([x.pass_down_Ex() for x in Xs]).T
#Recovered X Variance
#slight hack here - set q variance of observed nodes to zeros (it should be random...)
for x in Xs:
if x.observed:
x.qcov *=0
var_rec = np.vstack([np.diag(x.qcov) for x in Xs]) + 1./np.diag(Beta.pass_down_Ex())
#plot each recovered signal in a separate figure
for i in range(d):
pylab.figure();pylab.title('recovered_signal '+str(i))
pylab.plot(Xdata_full[:,i],'g',marker='.',label='True') # 'true' values of missing data (without noise)
pylab.plot(X_rec[:,i],'b',label='Recovered') # recovered mising data values
pylab.plot(Xdata[:,i],'k',marker='o',linewidth=2,label='Observed') # with noise, and holes where we took out values
pylab.legend()
volume_x = np.hstack((np.arange(len(Xs)),np.arange(len(Xs))[::-1]))
volume_y = np.hstack((X_rec[:,i]+2*np.sqrt(var_rec[:,i]), X_rec[:,i][::-1]-2*np.sqrt(var_rec[:,i])[::-1]))
pylab.fill(volume_x,volume_y,'b',alpha=0.3)
print '\nBeta'
print true_prec,Beta.pass_down_Ex()[0,0]
print '\nMu'
print np.hstack((true_mean,Mu.pass_down_Ex()))
pylab.show()
Xs.reverse()
[a.update() for a in As]
[c.update() for c in Cs]
Q.update()
R.update()
print niters-i
#plot
if plot:
import pylab
import hinton
#plot hintons of learned (and true) matrices.
pylab.figure()
pylab.subplot(1,2,1)
pylab.title('True A')
hinton.hinton(true_A)
pylab.subplot(1,2,2)
pylab.title('E[A]')
hinton.hinton(A.pass_down_Ex())
pylab.figure()
pylab.subplot(1,2,1)
pylab.title('True C')
hinton.hinton(true_C)
pylab.subplot(1,2,2)
pylab.title('E[C]')
hinton.hinton(C.pass_down_Ex())
pylab.figure()
pylab.subplot(2,2,1)
pylab.title('True Q')
def restaurantffn_net_topology():
global graph
global error
global ffnname
number_nodes = [0, 0]
graph = FFNetwork("", anzeigefenster, printfenster, ki) # fuer clear_gui
folder = os.path.abspath(os.path.dirname("AIMA.py"))
folder = folder.split("\\aima")[0]
path = folder + "\\feedforward"
#path = os.path.abspath(os.path.dirname(__file__))
# datei mit selben namen.ffn wie bsp werden geoeffnet
ffnname = ffnname.split("/")
ffnname = ffnname[-1]
ffnname = ffnname.split(".")
ffnname = ffnname[0]
name = str(path) + "/" + ffnname + ".ffn"
print(name)
try:
datei = open(name).read()
matrix = np.empty([15, 15]) # fuer hinton diagramm
szenario = []
for string in datei.split("\n"):
szenario.append(string)
name = szenario[0]
numberofnodes = int(szenario[1])
numberofedges = int(szenario[2 + numberofnodes])
graph = FFNetwork(name, anzeigefenster, printfenster, ki)
# knoten und kanten in graphen aufnehmen:
for element in range(2, 2 + numberofnodes, 1):
szenario[element] = szenario[element].split()
nameofnode = szenario[element][0]
nodetype = szenario[element][1]
graph.add_node(nameofnode, nodetype)
if nameofnode[0] == "I":
number_nodes[0] += 1
if nameofnode[0] == "H":
number_nodes[1] += 1
for element in range(3 + numberofnodes,
3 + numberofnodes + numberofedges, 1):
szenario[element] = szenario[element].split()
start = szenario[element][0]
end = szenario[element][1]
start = graph.get_node(start)
end = graph.get_node(end)
weight = random.randint(0, 100) / 100
graph.add_edge(start.name(), end.name(), weight)
# matrix fuellen:
for zeile in range(15):
for spalte in range(15):
matrix[zeile][spalte] = 0
for node in graph.get_nodes():
name = node.name()
for edge in node.get_edges():
dest = edge.end().name()
weight = edge.weight
# ueberpruefung worein es geschireben wird:
if name[0] == "I":
oben = int(name[1])
elif name[0] == "H":
oben = int(name[1]) + 9
if dest[0] == "H":
runter = int(dest[1]) + 9
else:
runter = 14
matrix[oben][14 - runter] = weight
matrix[runter][14 - oben] = weight
hinton.hinton(matrix)
# unspruengliche Matrix mit noch zufaelligen Werten anzeigen:
plt.show()
# testen ob genug input/output
output = 0
input = 0
for node in graph.get_nodes():
if node.get_type() == "output":
output += 1
if node.get_type() == "input":
input += 1
if (output == 0 or input == 0):
printfenster.insert(END, "Not enough Input/Outputnodes")
zeichne.draw(graph, "", "")
return number_nodes
except:
error = True
printfenster.insert(END, "Error\n")
#a missing data problem
Nmissing = int(N * missing_pc / 100)
observed2 = observed.copy()
missingi = np.argsort(np.random.rand(N))[:Nmissing]
missingj = np.random.randint(0, d - q,
Nmissing) #last q columns will be complete
observed2[missingi, missingj] = np.NaN
b = PCA_EM_missing(observed2, q)
b.learn(niters)
from hinton import hinton
import pylab
colours = np.arange(N) # to colour the dots with
hinton(linalg.qr(trueW.T)[1].T)
pylab.title('true transformation')
pylab.figure()
hinton(linalg.qr(a.W.T)[1].T)
pylab.title('reconstructed transformation')
pylab.figure()
hinton(linalg.qr(b.W.T)[1].T)
pylab.title('reconstructed transformation (missing data)')
pylab.figure()
pylab.subplot(3, 1, 1)
pylab.plot(latents)
pylab.title('true latents')
pylab.subplot(3, 1, 2)
pylab.plot(a.m_Z)
pylab.title('reconstructed latents')
pylab.subplot(3, 1, 3)
for t in range(50,150):
g[t].weights[4:7] = np.ones(3)
f,Y = model.simulate(g)
f_est,P,M = model.estimate_fields(g,Y)
f_est = [f0] + f_est
#model.estimate_kernels(f,g,Y)
if True:
vmin = -2
vmax = 5
plt.figure()
hinton.hinton(model.LDS.A)
plt.figure()
plt.subplot(1,3,1)
Z = []
for fi in f:
Z.append([fi(s) for s in np.linspace(2,8,100)])
Z_true = np.array(Z)
plt.imshow(Z_true,vmin=vmin,vmax=vmax)
plt.xlabel('space')
plt.ylabel('time')
plt.title('true field')
plt.subplot(1,3,2)
Z = []
for fi in f_est:
Z.append([fi(s) for s in np.linspace(2,8,100)])
Z_est = np.array(Z)
idx] = -1 / 3 * 1 / Unq[temp[1]] # for u_k, Partial fu
B[idx, 0] = 1 / 4 * (Na[temp[0]] / maxNa + vq[temp[0]] / maxvq +
avgsumva[temp[0]] / maxavgsumva)
else:
## Other PostTypeId
A[idx, idx] = 1
idx += 1
for user in ulist:
A[idx, idx] = 1
B[idx, 0] = 0 # 1/3 * (repu[user.split(':')[0]]/maxrepu)
#B[idx,0] = 1/3 * (repu[user.split(':')[0]]/maxrepu)
idx += 1
if hinplot == 1:
hinton(A, 'Amat')
hinton(B, 'Bvec')
solution = np.linalg.solve(A, B)
out = dict(zip(x_theory, solution))
#pprint.pprint(out)
if hinplot == 1:
hinton(solution, 'Sol')
## Analysing results
ans = {}
ques = {}
uuser = {}
userQ = {}
anslist = []
queslist = []
#a missing data problem
Nmissing = int(N*missing_pc/100)
observed2 = observed.copy()
missingi = np.argsort(np.random.rand(N))[:Nmissing]
missingj = np.random.randint(0,d-q,Nmissing)#last q columns will be complete
observed2[missingi,missingj] = np.NaN
b = PCA_EM_missing(observed2,q)
b.learn(niters)
from hinton import hinton
import pylab
colours = np.arange(N)# to colour the dots with
hinton(linalg.qr(trueW.T)[1].T)
pylab.title('true transformation')
pylab.figure()
hinton(linalg.qr(a.W.T)[1].T)
pylab.title('reconstructed transformation')
pylab.figure()
hinton(linalg.qr(b.W.T)[1].T)
pylab.title('reconstructed transformation (missing data)')
pylab.figure()
pylab.subplot(3,1,1)
pylab.plot(latents)
pylab.title('true latents')
pylab.subplot(3,1,2)
pylab.plot(a.m_Z)
pylab.title('reconstructed latents')
pylab.subplot(3,1,3)