def corr_models(self):
df = self._acc()
df = df[df.sel]
corr = lambda x,y: 1 - scipy.spatial.distance.sqeuclidean(x,y) / len(x)
cr = []
for m1 in df.model.unique():
d1 = df[df.model==m1]
for s1 in df.dataset.unique():
for m2 in df.model.unique():
d2 = df[df.model==m2]
for s2 in df.dataset.unique():
if m1==m2 and s1==s2:
r = np.nan
else:
r = corr(d1[d1.dataset==s1].model_accuracy.values,
d2[d2.dataset==s2].model_accuracy.values)
cr.append([m1,s1,m2,s2,r])
cr = pandas.DataFrame(cr, columns=['model1', 'dataset1',
'model2', 'dataset2', 'lr'])
cr = stats.factorize(cr)
crs = cr.set_index(['model1', 'dataset1', 'model2', 'dataset2'])
crs = crs.unstack(['model2', 'dataset2'])
sns.plt.figure()
sns.heatmap(crs)
print(crs)
print(cr.groupby(['dataset1', 'dataset2']).mean())
self.show(pref='mcorr')
if self.myexp.html is not None:
self.myexp.html.writetable(crs, caption='proportion of matching responses')
g = cr.groupby(['dataset1', 'dataset2']).mean()
self.myexp.html.writetable(g, caption='mean proportion of matching responses')
def corr_models(mods1_dis, mods2_dis):
df = []
for mods1_label, mods1_data in mods1_dis.items():
inds = np.triu_indices(mods1_data.shape[0], k=1)
for mods2_label, mods2_data in mods2_dis.items():
corr = np.corrcoef(mods1_data[inds], mods2_data[inds])[0,1]
df.append([mods1_label, mods2_label, corr])
df = pandas.DataFrame(df, columns=['perception', 'models', 'correlation'])
df = stats.factorize(df)
sns.factorplot('perception', 'correlation', 'models',
data=df, kind='bar')
return df
def corr_neural_model(neural_dis, mod_dis):
df = []
for mod_label, mod_data in mod_dis.items():
for k in xrange(neural_dis.values()[0].shape[-1]):
nd_avg = []
for neural_label, neural_data in neural_dis.items():
inds = np.triu_indices(mod_data.shape[0], k=1)
corr = np.corrcoef(neural_data[:,:,k][inds],
mod_data[inds])[0,1]
df.append([neural_label, mod_label, k, corr])
df = pandas.DataFrame(df, columns=['neural', 'model',
'subjid', 'correlation'])
df = stats.factorize(df)
sns.factorplot('neural', 'correlation', 'model', data=df, kind='bar')
return df
def avg_hop2008(self, dis, plot=True):
# dis = self.dissimilarity()
df = []
for layer, dis in dis.items():
df.extend(self._avg(dis, layer, 'px'))
df.extend(self._avg(dis, layer, 'shape'))
other = dis.copy()
for sh in np.unique(self.dims['px']):
ss = self.dims['px'] == sh
other.T[ss].T[ss] = np.nan
for sh in np.unique(self.dims['shape']):
ss = self.dims['shape'] == sh
other.T[ss].T[ss] = np.nan
inds = range(len(other))
n = 0
for si, s1 in enumerate(inds):
for s2 in inds[si + 1:]:
if not np.isnan(other[s1, s2]):
df.append([layer, 'other', n, s1, s2, other[s1, s2]])
n += 1
df = pandas.DataFrame(
df, columns=['layer', 'kind', 'n', 'i', 'j', 'dissimilarity'])
if plot:
df = stats.factorize(df, order={'kind': ['px', 'shape', 'other']})
# df = df[df.feature != 'other']
# agg = stats.aggregate(df, values='dissimilarity', rows='layer',
# cols='feature', yerr='n')
agg = df.pivot_table(index='n',
columns=['layer', 'kind'],
values='dissimilarity')
sns.factorplot('layer',
'dissimilarity',
'kind',
data=df,
kind='bar')
self.show(pref='avg')
return df
def avg_hop2008(self, dis, plot=True):
# dis = self.dissimilarity()
df = []
for layer, dis in dis.items():
df.extend(self._avg(dis, layer, 'px'))
df.extend(self._avg(dis, layer, 'shape'))
other = dis.copy()
for sh in np.unique(self.dims['px']):
ss = self.dims['px'] == sh
other.T[ss].T[ss] = np.nan
for sh in np.unique(self.dims['shape']):
ss = self.dims['shape'] == sh
other.T[ss].T[ss] = np.nan
inds = range(len(other))
n = 0
for si, s1 in enumerate(inds):
for s2 in inds[si+1:]:
if not np.isnan(other[s1,s2]):
df.append([layer, 'other', n, s1, s2, other[s1,s2]])
n += 1
df = pandas.DataFrame(df, columns=['layer', 'kind', 'n', 'i', 'j', 'dissimilarity'])
if plot:
df = stats.factorize(df, order={'kind': ['px', 'shape', 'other']})
# df = df[df.feature != 'other']
# agg = stats.aggregate(df, values='dissimilarity', rows='layer',
# cols='feature', yerr='n')
agg = df.pivot_table(index='n', columns=['layer', 'kind'],
values='dissimilarity')
sns.factorplot('layer', 'dissimilarity', 'kind', data=df,
kind='bar')
self.show(pref='avg')
return df
def corr_neural_model_avg(neural_dis, mod_dis):
df = []
avg = [['pixelwise',
('BA17', 'BA18', 'TOS', 'postPPA',
'LOTCobject', 'LOTCface')],
['shape',
('LOTCbody', 'LOTChand',
'VOTCobject', 'VOTCbody/face')],
['animate/inanimate',
('LOTCobject', 'LOTCface', 'LOTCbody', 'LOTChand',
'VOTCobject', 'VOTCbody/face')],
['nat/artifact', ('IPS', 'SPL',
'IPL', 'DPFC')]
]
for a, rois in avg:
for mod_label, mod_data in mod_dis.items():
for k in xrange(neural_dis.values()[0].shape[-1]):
nd_avg = []
for neural_label, neural_data in neural_dis.items():
if neural_label in rois:
nd_avg.append(neural_data[:,:,k])
nd_avg = np.average(nd_avg, axis=0)
inds = np.triu_indices(nd_avg.shape[0], k=1)
corr = np.corrcoef(nd_avg[inds],
mod_data[inds])[0,1]
df.append([a, mod_label, k, corr])
df = pandas.DataFrame(df, columns=['neural', 'model',
'subjid', 'correlation'])
df = stats.factorize(df)
sns.factorplot('neural', 'correlation', 'model', data=df, kind='bar')
return df
def corr_models(self):
df = self._acc()
df = df[df.sel]
corr = lambda x, y: 1 - scipy.spatial.distance.sqeuclidean(x, y) / len(
x)
cr = []
for m1 in df.model.unique():
d1 = df[df.model == m1]
for s1 in df.dataset.unique():
for m2 in df.model.unique():
d2 = df[df.model == m2]
for s2 in df.dataset.unique():
if m1 == m2 and s1 == s2:
r = np.nan
else:
r = corr(
d1[d1.dataset == s1].model_accuracy.values,
d2[d2.dataset == s2].model_accuracy.values)
cr.append([m1, s1, m2, s2, r])
cr = pandas.DataFrame(
cr, columns=['model1', 'dataset1', 'model2', 'dataset2', 'lr'])
cr = stats.factorize(cr)
crs = cr.set_index(['model1', 'dataset1', 'model2', 'dataset2'])
crs = crs.unstack(['model2', 'dataset2'])
sns.plt.figure()
sns.heatmap(crs)
print(crs)
print(cr.groupby(['dataset1', 'dataset2']).mean())
self.show(pref='mcorr')
if self.myexp.html is not None:
self.myexp.html.writetable(
crs, caption='proportion of matching responses')
g = cr.groupby(['dataset1', 'dataset2']).mean()
self.myexp.html.writetable(
g, caption='mean proportion of matching responses')
