def behav_olddata(self):
self.model_name = 'behav'
sil = pandas.read_csv('snodgrass/data/sil_human_acc.csv',
header=None) / 100.
df = [
('color', .903, .169),
('gray', .892, .172),
# ('line drawing', .882, .171),
('silhouette', sil.mean().values[0], sil.std(ddof=1).values[0])
]
df = pandas.DataFrame(df, columns=['dataset', 'accuracy', 'stdev'])
n = 260
ci = df.stdev * 1.96 / np.sqrt(n)
df['ci_low'] = df.accuracy - ci
df['ci_high'] = df.accuracy + ci
sns.factorplot('dataset',
'accuracy',
data=df,
kind='bar',
color=self.colors['shape'])
hue = 'kind' if 'kind' in df else None
plot.plot_ci(df, what=['Rectangle'], hue=hue)
self.show(pref='acc')
self.html.writetable(df)
def pred_corr(self):
pref = 'pred_corr'
print()
print('{:=^50}'.format(' ' + pref + ' '))
value = 'model_accuracy'
method = 'euclidean'
df = self.get_data_all(pref, kind='compare', value=value, method=method)
behav = self.myexp.behav()
behav = behav.pivot_table(index=['kind', 'subjid'],
columns='no', values='acc')
df = base._set_ci(df, groupby=['models', 'dataset'])
g = sns.factorplot(x='dataset', y='consistency', hue='models',
data=df, kind='bar',
color=self.myexp.colors['shape'])
hue = 'kind' if 'kind' in df else None
plot.plot_ci(df, hue=hue)
sns.plt.ylim([0,1])
for sno, subset in enumerate(ORDER):
self.myexp.set_subset(subset)
rel = self.reliability(behav.loc[subset])
# import pdb; pdb.set_trace()
rel = ((1+rel[0])/2., (1+rel[1])/2.)
sns.plt.axhspan(rel[0], rel[1], xmin=(2*sno+1)/6.-1/7., xmax=(2*sno+1)/6.+1/7., facecolor='0.9', edgecolor='0.9', zorder=0)
self.show(pref=pref, suffix='all_' + value + '_' + method)
def plot_all(self,
df,
values,
dim,
pref='',
ceiling=None,
color=None,
ylim=[-.1, 1.1]):
df = _set_ci(df)
print(df)
gray = sns.color_palette('Set2', 8)[-1]
light = (.3, .3, .3) #colors[-2]
# if color is None:
# color = self.colors[0]
# light = sns.light_palette(color, n_colors=3)[1]
palette = []
for model in df.models.unique():
depth = df[df.models == model].depth.iloc[0]
if depth == 'shallow':
palette.append(gray)
elif depth == 'hmax':
palette.append(light)
elif depth == 'deep':
palette.append(color)
# dims = df.kind.unique()
# col = None if len(dims) == 1 else 'kind'
g = sns.factorplot(x='models',
y=values,
data=df,
kind='bar',
palette=palette)
sns.plt.ylim(ylim)
if 'ci_low' in df.columns:
hue = 'kind' if 'kind' in df else None
plot.plot_ci(df, hue=hue)
if 'accuracy' in df.columns and self.myexp.dims[dim] is not None:
plot_chance(self.myexp.dims[dim])
# for ax, dim in zip(g.axes.flat, dims):
sns.plt.title(dim)
if ceiling is not None:
sns.plt.axhspan(ceiling[0],
ceiling[1],
facecolor='0.9',
edgecolor='0.9',
zorder=0)
self.set_vertical_labels(g)
#pref = kind if pref == '' else pref + '_' + kind
self.show(pref=pref, suffix='all_' + dim)
def behav_olddata(self):
self.model_name = 'behav'
sil = pandas.read_csv('snodgrass/data/sil_human_acc.csv', header=None) / 100.
df = [('color', .903, .169),
('gray', .892, .172),
# ('line drawing', .882, .171),
('silhouette', sil.mean().values[0], sil.std(ddof=1).values[0])]
df = pandas.DataFrame(df, columns=['dataset', 'accuracy', 'stdev'])
n = 260
ci = df.stdev * 1.96 / np.sqrt(n)
df['ci_low'] = df.accuracy - ci
df['ci_high'] = df.accuracy + ci
sns.factorplot('dataset', 'accuracy', data=df,
kind='bar', color=self.colors['shape'])
hue = 'kind' if 'kind' in df else None
plot.plot_ci(df, what=['Rectangle'], hue=hue)
self.show(pref='acc')
self.html.writetable(df)
def func_wrapper(self, df, values, ylim=[0, 1]):
g = plot_func(self, df, values, ylim=ylim)
if 'ci_low' in df.columns:
hue = 'kind' if 'kind' in df else None
plot.plot_ci(df, hue=hue)
if 'accuracy' in df.columns:
for value in self.dims.values():
plot_chance(value)
if self.model_name in ['googlenet', 'vgg-19']:
labels = g.axes.flat[0].get_xticklabels()
for label in labels:
if len(label.get_text()) > 3:
label.set_ha('right')
label.set_rotation(30)
# if self.model_name == 'googlenet':
# sns.plt.subplots_adjust(bottom=.25)
self.show(values)
def func_wrapper(self, df, values, ylim=[0, 1]):
g = plot_func(self, df, values, ylim=ylim)
if "ci_low" in df.columns:
hue = "kind" if "kind" in df else None
plot.plot_ci(df, hue=hue)
if "accuracy" in df.columns:
for value in self.dims.values():
plot_chance(value)
if self.model_name in ["googlenet", "vgg-19"]:
labels = g.axes.flat[0].get_xticklabels()
for label in labels:
if len(label.get_text()) > 3:
label.set_ha("right")
label.set_rotation(30)
# if self.model_name == 'googlenet':
# sns.plt.subplots_adjust(bottom=.25)
self.show(values)
def pred_corr(self):
pref = 'pred_corr'
print()
print('{:=^50}'.format(' ' + pref + ' '))
value = 'model_accuracy'
method = 'euclidean'
df = self.get_data_all(pref,
kind='compare',
value=value,
method=method)
behav = self.myexp.behav()
behav = behav.pivot_table(index=['kind', 'subjid'],
columns='no',
values='acc')
df = base._set_ci(df, groupby=['models', 'dataset'])
g = sns.factorplot(x='dataset',
y='consistency',
hue='models',
data=df,
kind='bar',
color=self.myexp.colors['shape'])
hue = 'kind' if 'kind' in df else None
plot.plot_ci(df, hue=hue)
sns.plt.ylim([0, 1])
for sno, subset in enumerate(ORDER):
self.myexp.set_subset(subset)
rel = self.reliability(behav.loc[subset])
# import pdb; pdb.set_trace()
rel = ((1 + rel[0]) / 2., (1 + rel[1]) / 2.)
sns.plt.axhspan(rel[0],
rel[1],
xmin=(2 * sno + 1) / 6. - 1 / 7.,
xmax=(2 * sno + 1) / 6. + 1 / 7.,
facecolor='0.9',
edgecolor='0.9',
zorder=0)
self.show(pref=pref, suffix='all_' + value + '_' + method)
def plot_all(self, df, values, dim, pref="", ceiling=None, color=None, ylim=[-0.1, 1.1]):
df = _set_ci(df)
print(df)
gray = sns.color_palette("Set2", 8)[-1]
light = (0.3, 0.3, 0.3) # colors[-2]
# if color is None:
# color = self.colors[0]
# light = sns.light_palette(color, n_colors=3)[1]
palette = []
for model in df.models.unique():
depth = df[df.models == model].depth.iloc[0]
if depth == "shallow":
palette.append(gray)
elif depth == "hmax":
palette.append(light)
elif depth == "deep":
palette.append(color)
# dims = df.kind.unique()
# col = None if len(dims) == 1 else 'kind'
g = sns.factorplot(x="models", y=values, data=df, kind="bar", palette=palette)
sns.plt.ylim(ylim)
if "ci_low" in df.columns:
hue = "kind" if "kind" in df else None
plot.plot_ci(df, hue=hue)
if "accuracy" in df.columns and self.myexp.dims[dim] is not None:
plot_chance(self.myexp.dims[dim])
# for ax, dim in zip(g.axes.flat, dims):
sns.plt.title(dim)
if ceiling is not None:
sns.plt.axhspan(ceiling[0], ceiling[1], facecolor="0.9", edgecolor="0.9", zorder=0)
self.set_vertical_labels(g)
# pref = kind if pref == '' else pref + '_' + kind
self.show(pref=pref, suffix="all_" + dim)
