def plot_loss(self, X, Y, loss, show=True):
# Graphic display
title = str(self.step) + ': Epochs:' + str(self.epochs) + \
" batch:" + str(self.batch_size) + \
" alpha:" + str(self.learning_rate)
train_X1 = X[:, 0]
fig = plt.figure()
ax = fig.gca(projection='3d')
ax.text2D(0.05, 0.95, title, transform=ax.transAxes)
# trisurf plot
# ax.plot_trisurf(np.ravel(train_X1),
# np.ravel(train_X2), np.ravel(Y), color='b')
ax.plot_trisurf(np.ravel(train_X1), np.ravel(Y), np.ravel(loss))
ax.set_xlabel('Weight')
ax.set_ylabel('bias')
ax.set_zlabel('Loss')
ax.legend()
# save image to buffer
buf = io.BytesIO()
if show:
# If want to view image in Tensorboard, do not show plot => Strange
# bug!!!
image_path = self.logs_path + "/images"
if not os.path.exists(image_path):
os.mkdir(image_path)
filename = image_path + "/" + \
"loss-" + \
time.strftime("%Y%m%d-%H%M%S", time.localtime()) + ".png"
plt.savefig(filename, format='png')
plt.show()
else:
plt.savefig(buf, format='png')
buf.seek(0)
plt.close()
# plt.clf()
return buf
# cPickle.dump([difficulties, generosities, ave_switches], handle) #read with open(filename, 'rb') as handle: data = cPickle.load(handle) difficulties=data[0] generosities=data[1] ave_switches=data[2] fig = plt.figure() ax = Axes3D(fig) Axes3D.scatter(ax, difficulties, generosities, ave_switches, cmap=cm.jet) Axes3D.plot_trisurf(ax, difficulties, generosities, ave_switches, cmap=cm.jet) plt.show() print ave_switches # griddata and contour. xi = np.linspace(min(difficulties),max(difficulties),15) yi = np.linspace(min(generosities),max(generosities),15) xi = np.linspace(0,1,15) yi = np.linspace(0,1,15) print len(difficulties), len(generosities), len(ave_switches)
import glob
def read_and_overview(filename):
"""Read NetCDF using read_generic_netcdf and print upper level dictionary keys
"""
test = read_generic_netcdf(filename)
print("\nPrint keys for file %s" % os.path.basename(filename))
for key in test.keys():
print("\t%s" % key)
pfad = ('/automount/ags/velibor/gpmdata/nicht3dComposit/*.nc')
pfad_3d = sorted(glob.glob(pfad))[1]
comp3d = get_wradlib_data_file(pfad_3d)
read_and_overview(comp3d)
from netCDF4 import Dataset
import numpy as np
comp3d = Dataset(pfad_3d, mode='r')
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
Axes3D.plot_trisurf(x,y,z)
def block_analysis(model=False):
# BLOCK BY BLOCK
pre_metacog_switches={0:[], 1:[], 2:[]}
post_metacog_switches={0:[], 1:[], 2:[]}
difficulties={0:[], 1:[], 2:[]}
generosities={0:[], 1:[], 2:[]}
pre_met_swi_rews={0:[], 1:[], 2:[]}
post_met_swi_rews={0:[], 1:[], 2:[]}
for condition in set(data.conditions):
for payrate in set(data.payrates):
pre_metacog_switches[condition].append(gameAnalyzer.metacog_switches(data.bandit_data[condition][payrate], n_trials=subject_trials, post_metacog=False))
post_metacog_switches[condition].append(gameAnalyzer.metacog_switches(data.bandit_data[condition][payrate], n_trials=subject_trials, post_metacog=True))
difficulties[condition].append(1-abs(payrate[1]-payrate[0]))
generosities[condition].append((payrate[1]+payrate[0])/2)
pre_met_swi_rews[condition].append(gameAnalyzer.metacog_rewards(data.bandit_data[condition][payrate], n_trials=subject_trials, post_metacog=False))
post_met_swi_rews[condition].append(gameAnalyzer.metacog_rewards(data.bandit_data[condition][payrate], n_trials=subject_trials, post_metacog=True))
print pre_metacog_switches
print post_metacog_switches
print np.mean(pre_metacog_switches[0])
print np.mean(pre_metacog_switches[1])
print np.mean(pre_metacog_switches[2])
print np.mean(post_metacog_switches[0])
print np.mean(post_metacog_switches[1])
print np.mean(post_metacog_switches[2])
s0=sum(post_metacog_switches[0])
s1=sum(post_metacog_switches[1])
print s0, s1
print pre_met_swi_rews
print post_met_swi_rews
print np.mean(post_met_swi_rews[0]), np.mean(post_met_swi_rews[1])
normdif0=np.array(post_metacog_switches[0])-np.array(post_met_swi_rews[0])
normdif1=np.array(post_metacog_switches[1])-np.array(post_met_swi_rews[1])
print normdif0
print normdif1
print np.mean(normdif0), np.mean(normdif1)
collapsing=np.array(post_metacog_switches[0])-np.array(post_metacog_switches[1])
normcollapsing=normdif0-normdif1
# p=0.057
print 'pvalue from ttest: {0}'.format(stats.ttest_rel(post_metacog_switches[0],post_metacog_switches[1])[1])
print 'same for rewards (pvalue from ttest): {0}'.format(stats.ttest_rel(post_met_swi_rews[0], post_met_swi_rews[1])[1])
print '...and subtracted: {0}'.format(stats.ttest_rel(normdif0, normdif1)[1])
fig = plt.figure()
ax = Axes3D(fig)
Axes3D.scatter(ax, difficulties[0], generosities[0], collapsing, cmap=cm.jet)
Axes3D.plot_trisurf(ax, difficulties[0], generosities[0], collapsing, cmap=cm.jet)
plt.show()
# griddata and contour.
xi = np.linspace(min(difficulties[0]),max(difficulties[0]),15)
yi = np.linspace(min(generosities[0]),max(generosities[0]),15)
xi = np.linspace(0,1,15)
yi = np.linspace(0,1,15)
print len(difficulties[0]), len(generosities[0]), len(collapsing)
zi = griddata(difficulties[0],generosities[0],collapsing,xi,yi,interp='nn')#linear')
#plt.contour(xi,yi,zi,15,linewidths=0.5,colors='k')
plt.contourf(xi,yi,zi,cmap=cm.jet)#,
#norm=plt.normalize(vmax=abs(zi).max(), vmin=-abs(zi).max()))
plt.scatter(difficulties[0], generosities[0], marker='s', c=collapsing, s=200, cmap=cm.jet)
plt.colorbar() # draw colorba
plt.xlim([0,1])
plt.ylim([0,1])
plt.show()
# plt.scatter(difficulties[0], generosities[0], marker='s', c=collapsing, s=200, cmap=cm.coolwarm)
# plt.colorbar()
# plt.xlim([0,1])
# plt.ylim([0,1])
# plt.show()
if model:
#MCMC inference for rate difference
modeldata=np.array([post_metacog_switches[0],post_metacog_switches[1]])
#modeldata=np.array([post_met_swi_rews[0],post_met_swi_rews[1]]) #this is dubious, but good, nonsignificant
#modeldata=np.array([normdif0,normdif1]) #this does not work, negative values.. should adapt model for negative reward triggering switch.
model=pymc.MCMC(rateDifferenceModel.make_model(modeldata, subject_trials))
model.sample(iter=10100, burn=100, thin=10, progress_bar=False)
deltas=model.trace('delta')[:,:]
mdeltas=np.mean(deltas,0)
print 'supermean deltas: {0}'.format(np.mean(mdeltas))
alldeltas=deltas.reshape(deltas.shape[0]*deltas.shape[1])
print 'testing: {0}, {1}'.format(sum(alldeltas<0), len(alldeltas))
print 'aproximate pvalue from bayesian inference: {0}'.format(float(sum(alldeltas<0))/len(alldeltas))
# p=0.058
print 'pvalue for deltas != 0 from ttest (taking mean): {0}'.format(stats.ttest_1samp(mdeltas, 0))
print 'pvalue for deltas != 0 from ttest (all together): {0}'.format(stats.ttest_1samp(alldeltas, 0))
print 'delta shape: {0}'.format(deltas.shape)
pvals=np.asfarray(np.sum(deltas<0,0))/deltas.shape[0]
print 'try something better? 16 p values: {0}'.format(pvals)
print 'same, bonferroni corrected: {0}'.format(pvals*len(pvals))
print np.where(pvals==pvals.min())
npgens0=np.array(generosities[0])
npgens1=np.array(generosities[1])
npdifs0=np.array(difficulties[0])
npdifs1=np.array(difficulties[1])
print 'most significant effect @G={0}, D={1}'.format(npgens0[pvals==pvals.min()],npdifs0[pvals==pvals.min()])
print 'just a check @G={0}, D={1}'.format(npgens1[pvals==pvals.min()],npdifs1[pvals==pvals.min()])
# plt.hist(np.mean(deltas,0))
# plt.show()
plt.hist(alldeltas)
plt.show()
# utils.plot_and_correlate(difficulties[0],mdeltas)
# utils.plot_and_correlate(difficulties[0],pre_metacog_switches[0])
# utils.plot_and_correlate(difficulties[0],post_metacog_switches[0])
# utils.plot_and_correlate(difficulties[0],post_metacog_switches[1])
# utils.plot_and_correlate(generosities[0],mdeltas)
# utils.plot_and_correlate(generosities[0],pre_metacog_switches[0])
# utils.plot_and_correlate(generosities[0],post_metacog_switches[0])
# utils.plot_and_correlate(generosities[0],post_metacog_switches[1])
# utils.plot_and_correlate(generosities[0], difficulties[0])
fig = plt.figure()
ax = Axes3D(fig)
Axes3D.scatter(ax, difficulties[0], generosities[0], mdeltas, cmap=cm.jet)
Axes3D.plot_trisurf(ax, difficulties[0], generosities[0], mdeltas, cmap=cm.jet)
plt.show()
# griddata and contour.
xi = np.linspace(min(difficulties[0]),max(difficulties[0]),15)
yi = np.linspace(min(generosities[0]),max(generosities[0]),15)
xi = np.linspace(0,1,15)
yi = np.linspace(0,1,15)
zi = griddata(difficulties[0],generosities[0],mdeltas,xi,yi,interp='nn')#linear')
#plt.contour(xi,yi,zi,15,linewidths=0.5,colors='k')
plt.contourf(xi,yi,zi,cmap=cm.jet)#,
#norm=plt.normalize(vmax=abs(zi).max(), vmin=-abs(zi).max()))
plt.scatter(difficulties[0], generosities[0], marker='s', c=mdeltas, s=200, cmap=cm.jet)
plt.colorbar() # draw colorba
plt.xlim([0,1])
plt.ylim([0,1])
plt.show()
def plot(self, X, Y, show=True):
# Graphic display
predicted = self.sess.run(self.predict_model(X))
title = str(self.step) + ': Epochs:' + str(self.epochs) + \
" batch:" + str(self.batch_size) + \
" alpha:" + str(self.learning_rate)
if self.n_features == 1: # plot 2D image
plt.title(title)
# plt.plot(self.train_X, self.train_Y, c='b', label='Original data')
# plt.plot(self.train_X, predicted, c='r', label='Fitted')
plt.scatter(X, Y, c='b', label='Original data')
plt.scatter(X, predicted, c='r', label='Fitted')
plt.legend()
else: # plot 3D for X1, X2
if (self.n_features == 2):
train_X1, train_X2 = np.hsplit(X, self.n_features)
else:
train_X1, train_X2, _ = np.hsplit(X, self.n_features)
fig = plt.figure()
ax = fig.gca(projection='3d')
ax.text2D(0.05, 0.95, title, transform=ax.transAxes)
# line plot
# x = np.squeeze(np.asarray(train_X1))
# y = np.squeeze(np.asarray(train_X2))
# z = np.squeeze(np.asarray(Y))
# z1 = np.squeeze(np.asarray(predicted))
# ax.plot(x, y, zs=z, c='b', label='Original data')
# ax.plot(x, y, zs=z1, c='r', label='Fitted')
# statter plot
ax.scatter(train_X1,
train_X2,
Y,
c='b',
marker='s',
label='Original data')
# ax.scatter(train_X1, train_X2, predicted,
# c='r', marker='v', label='Fitted')
# trisurf plot
# ax.plot_trisurf(np.ravel(train_X1),
# np.ravel(train_X2), np.ravel(Y), color='b')
ax.plot_trisurf(np.ravel(train_X1),
np.ravel(train_X2),
np.ravel(predicted),
color='r')
ax.set_xlabel('X1')
ax.set_ylabel('X2')
ax.set_zlabel('Y')
ax.legend()
# save image to buffer
buf = io.BytesIO()
if show:
# If want to view image in Tensorboard, do not show plot => Strange
# bug!!!
image_path = self.logs_path + "/images"
if not os.path.exists(image_path):
os.mkdir(image_path)
filename = image_path + "/" + \
time.strftime("%Y%m%d-%H%M%S", time.localtime()) + ".png"
plt.savefig(filename, format='png')
plt.show()
else:
plt.savefig(buf, format='png')
buf.seek(0)
plt.close()
# plt.clf()
return buf