def genFakeData(td, method="genX", decay=0):
for sub in subs:
fake_data[sub] = []
fake_irts[sub] = []
for setnum in range(numsets):
dataset = []
irtset = []
for listnum, trimval in enumerate(listlengths[sub]):
td.trim = trimval
## SEED START
# find animal key in USFanimals or substitute with closest animal
startAnimal = real_data[sub][listnum][0]
if startAnimal == 'polarbear': startAnimal = 'bear'
if startAnimal == 'beagle': startAnimal = 'dog'
if startAnimal == 'bulldog': startAnimal = 'dog'
if startAnimal == 'cheetah': startAnimal = 'tiger'
if startAnimal == 'python': startAnimal = 'snake'
if startAnimal == 'ferret': startAnimal = 'possum'
if startAnimal == 'hedgehog': startAnimal = 'chipmunk'
for key, val in USFanimals.iteritems():
if val == startAnimal:
td.startX = ('specific', key)
## SEED END
print str(sub), str(td.trim)
if method == "genX":
[data, irts, alter_graph] = rw.genX(USFnet, td)
irtset.append(irts)
elif method == "frequency":
data = [rw.nodeDegreeSearch(USFnet, td) for i in range(3)]
elif method == "cbdfs":
data = [rw.cbdfs(USFnet, td) for i in range(3)]
elif method == "sa":
data = [
rw.spreadingActivationSearch(USFnet, td, decay)
for i in range(3)
]
data = numToAnimal(data, USFanimals)[0]
dataset.append(data)
fake_data[sub].append(dataset)
fake_irts[sub].append(irtset)
return fake_data
usfg=nx.to_networkx_graph(usf)
numnodes=len(items)
usfsize=160
misses=[]
f=open('usf_sims_misses.csv','a', 0) # write/append to file with no buffering
for numlists in range(2,18):
for listlength in [15,30,50,70]:
misses=0
crs=0
for gnum in range(100): # how many samples for each numlists/listlength combo
print numlists, listlength, gnum
# generate toy lists
Xs=[rw.genX(usfg)[0:listlength] for i in range(numlists)]
itemsinXs=np.unique(rw.flatten_list(Xs)) # list of items that appear in toy data
notinx=[] # nodes not in trimmed X
for i in range(usfsize):
if i not in itemsinXs:
notinx.append(i)
miss=sum([len([j for j in usfg.neighbors(i) if j > i]) for i in notinx])
misses=misses+miss
cr=sum([(usfsize-i-1) for i in range(len(notinx))])
cr=cr-miss
crs=crs+cr
import rw
import numpy as np
import random
import math
import sys
numnodes=15
numlinks=4
probRewire=.3
numx=3
graph_seed=1
x_seed=graph_seed
randomseed=1
g,a=rw.genG(numnodes,numlinks,probRewire,seed=graph_seed)
Xs=[rw.genX(g, seed=x_seed+i) for i in range(numx)]
def PfromB(b):
P=np.empty([numnodes,numnodes],dtype=float) # set P from b (transition matix P[to,from])
for colnum, col in enumerate(b.T):
for rownum, row in enumerate(col):
P[rownum,colnum]=math.exp(row)/sum([math.exp(i) for i in col])
return P
def regml():
#random.seed(randomseed) # only for replicability
#np.random.seed(randomseed)
# free parameters
c=.75 # value used by Kwang, and Bottou, 2012
numedges = numnodes * (numlinks / 2) # number of edges in graph
numx = 3 # How many observed lists?
trim = 1 # ~ What proportion of graph does each list cover?
# PARAMETERS FOR RECONTRUCTING GRAPH
jeff = 0.9 # 1-IRT weight
beta = (1 / 1.1
) # for gamma distribution when generating IRTs from hidden nodes
#graph_seed=None
#x_seed=None
graph_seed = 65 # make sure same toy network is generated every time
x_seed = 65 # make sure same Xs are generated every time
# toy data
g, a = rw.genG(numnodes, numlinks, probRewire, seed=graph_seed)
[Xs,
steps] = zip(*[rw.genX(g, seed=x_seed + i, use_irts=1) for i in range(numx)])
Xs = list(Xs)
steps = list(steps)
irts = rw.stepsToIRT(steps, beta, seed=x_seed)
[Xs, irts, alter_graph] = rw.trimX(trim, Xs, irts, g)
# Find best graph!
#best_graph, bestval=rw.findBestGraph(Xs, irts, jeff, beta, numnodes)
best_graph, bestval = rw.findBestGraph(Xs, numnodes=numnodes)
# convert best graph to networkX graph, write to file
#g=nx.to_networkx_graph(best_graph)
#nx.write_dot(g,subj+".dot")
# generate data for `numsub` participants, each having `numlists` lists of `listlengths` items
seednum = 0 # seednum=150 (numsubs*numlists) means start at second sim, etc.
with open('gradualpluscutoff7.csv', 'w', 0) as fh:
fh.write("method,simnum,listnum,hit,miss,fa,cr,cost,startseed\n")
for simnum in range(numsims):
data = [] # Xs using usf_item indices
datab = [
] # Xs using ss_item indices (nodes only generated by subject)
numnodes = []
items = [] # ss_items
startseed = seednum # for recording
for sub in range(numsubs):
Xs = rw.genX(usf_graph_nx, toydata, seed=seednum)[0]
data.append(Xs)
# renumber dictionary and item list
itemset = set(rw.flatten_list(Xs))
numnodes.append(len(itemset))
ss_items = {}
convertX = {}
for itemnum, item in enumerate(itemset):
ss_items[itemnum] = usf_items[item]
convertX[item] = itemnum
items.append(ss_items)
Xs = [[convertX[i] for i in x] for x in Xs]
numlinks=4 # initial number of edges per node (must be even) probRewire=.3 # probability of re-wiring an edge numedges=numnodes*(numlinks/2) # number of edges in graph numx=3 # How many observed lists? trim=1 # ~ What proportion of graph does each list cover? # PARAMETERS FOR RECONTRUCTING GRAPH jeff=0.9 # 1-IRT weight beta=(1/1.1) # for gamma distribution when generating IRTs from hidden nodes #graph_seed=None #x_seed=None graph_seed=65 # make sure same toy network is generated every time x_seed=65 # make sure same Xs are generated every time # toy data g,a=rw.genG(numnodes,numlinks,probRewire,seed=graph_seed) [Xs,steps]=zip(*[rw.genX(g, seed=x_seed+i,use_irts=1) for i in range(numx)]) Xs=list(Xs) steps=list(steps) irts=rw.stepsToIRT(steps, beta, seed=x_seed) [Xs,irts,alter_graph]=rw.trimX(trim,Xs,irts,g) # Find best graph! best_graph, bestval=rw.findBestGraph(Xs, irts, jeff, beta, numnodes) # convert best graph to networkX graph, write to file #g=nx.to_networkx_graph(best_graph) #nx.write_dot(g,subj+".dot")
import rw
import numpy as np
steps=[]
ehs=[]
meaneh=[]
meanstep=[]
fh1 = open('ehmean.csv','w')
fh2 = open('eh.csv','w')
for i in range(10000):
print i
g,a=rw.genG(20,4,.3)
X,step=rw.genX(g,use_irts=1)
eh=rw.expectedHidden([X],a,len(a))[0]
steps.append(step)
ehs.append(eh)
meaneh.append(np.mean(eh))
meanstep.append(np.mean(step))
for num, i in enumerate(meaneh):
fh1.write(str(i) + "," + str(meanstep[num-1]) + "\n")
ehs=rw.flatten_list(ehs)
steps=rw.flatten_list(steps)
for num, i in enumerate(ehs):
fh2.write(str(i) + "," + str(steps[num-1]) + "\n")
fh1.close()
toygraph = rw.Toygraphs({
'graphtype': "steyvers",
'numnodes': 50,
'numlinks': 6
})
fh = open('priming_test.csv', 'w')
seed = 15
for td in toydata:
print "numx: ", td.numx
# generate data with priming and fit best graph
g, a = rw.genG(toygraph, seed=seed)
[Xs, irts, alter_graph] = rw.genX(g, td, seed=seed)
bestgraph_priming, ll = rw.uinvite(Xs,
td,
toygraph.numnodes,
fitinfo=fitinfo,
seed=seed,
debug=True)
priming_cost = rw.cost(bestgraph_priming, a)
print priming_cost
td.priming = 0.0
# fit best graph assuming no priming
bestgraph_nopriming, ll = rw.uinvite(Xs,
td,
toygraph.numnodes,
fitinfo=fitinfo,
irts = rw.Irts({
'data': [],
'irttype': "gamma",
'beta': (1 / 1.1),
'irt_weight': 0.9,
'rcutoff': 20
})
fitinfo = rw.Fitinfo({
'tolerance': 1500,
'startGraph': "naiverw",
'prob_multi': 1.0,
'prob_overlap': 0.5
})
x_seed = 1
graph_seed = 1
td = toydata[0]
g, a = rw.genG(toygraphs, seed=graph_seed)
[Xs,
irts.data] = zip(*[rw.genX(g, td, seed=x_seed + i) for i in range(td.numx)])
Xs = list(Xs)
irts.data = list(irts.data)
[Xs, irts.data, alter_graph] = rw.trimX(td.trim, Xs,
irts.data) # trim data when necessary
rw.drawMat(a, cmap=plt.cm.BuGn)
newmat = rw.drawMatChange(Xs, a, td, (0, 1), keep=0) # e.g.
## generate fake_ data and irts yoked to real_ data
numsets = 100 # number of sets of fake data per SS
fake_data = {}
fake_irts = {}
for subj in subs:
graph, items = rw.read_csv(real_graphs,
cols=('node1', 'node2'),
header=1,
filters={
'subj': str(subj),
'uinvite': '1'
})
graph = nx.from_numpy_matrix(graph)
fake_data[subj] = []
fake_irts[subj] = []
for setnum in range(numsets):
dataset = []
irtset = []
for trimval in listlengths[subj]:
toydata.trim = trimval
print str(subj), str(toydata.trim)
[data, irts, alter_graph] = rw.genX(graph, toydata)
data = numToAnimal(data, items)[0]
dataset.append(data)
irtset.append(irts)
fake_data[subj].append(dataset)
fake_irts[subj].append(irtset)