def estimate_cutoffs_plot(output_file,
df_plt,
df_cell_estimate_cutoff,
df_fit=None,
scale_x_log10=False,
save_plot=True):
"""Plot UMI counts by sorted cell barcodes."""
if min(df_plt['umi_counts']) <= 0:
fix_log_scale = min(df_plt['umi_counts']) + 1
df_plt['umi_counts'] = df_plt['umi_counts'] + fix_log_scale
gplt = plt9.ggplot()
gplt = gplt + plt9.theme_bw()
if len(df_plt) <= 50000:
gplt = gplt + plt9.geom_point(mapping=plt9.aes(x='barcode',
y='umi_counts'),
data=df_plt,
alpha=0.05,
size=0.1)
else:
gplt = gplt + plt9.geom_line(mapping=plt9.aes(x='barcode',
y='umi_counts'),
data=df_plt,
alpha=0.25,
size=0.75,
color='black')
gplt = gplt + plt9.geom_vline(mapping=plt9.aes(xintercept='n_cells',
color='method'),
data=df_cell_estimate_cutoff,
alpha=0.75,
linetype='dashdot')
gplt = gplt + plt9.scale_color_brewer(palette='Dark2', type='qual')
if scale_x_log10:
gplt = gplt + plt9.scale_x_continuous(
trans='log10', labels=comma_labels, minor_breaks=0)
else:
gplt = gplt + plt9.scale_x_continuous(labels=comma_labels,
minor_breaks=0)
gplt = gplt + plt9.scale_y_continuous(
trans='log10', labels=comma_labels, minor_breaks=0)
gplt = gplt + plt9.labs(title='',
y='UMI counts',
x='Barcode index, sorted by UMI count',
color='Cutoff')
# Add the fit of the droplet utils model
if df_fit:
gplt = gplt + plt9.geom_line(mapping=plt9.aes(x='x', y='y'),
data=df_fit,
alpha=1,
color='yellow')
if save_plot:
gplt.save('{}.png'.format(output_file), dpi=300, width=5, height=4)
return gplt
def test_wrong_bases():
# x axis not transformed
p = (ggplot(df, aes('x', 'x')) +
annotation_logticks(sides='b', size=.75, base=10) + geom_point())
with pytest.warns(PlotnineWarning):
p.draw_test()
# x axis not transform, but ticks requested for a different base
p = (ggplot(df, aes('x', 'x')) +
annotation_logticks(sides='b', size=.75, base=10) +
scale_x_continuous(trans=log_trans(8)) + geom_point())
with pytest.warns(PlotnineWarning):
p.draw_test()
# x axis is discrete
df2 = df.assign(discrete=pd.Categorical([str(a) for a in df['x']]))
p = (ggplot(df2, aes('discrete', 'x')) +
annotation_logticks(sides='b', size=.75, base=None) + geom_point())
with pytest.warns(PlotnineWarning):
p.draw_test()
# y axis is discrete
df2 = df.assign(discrete=pd.Categorical([str(a) for a in df['x']]))
p = (ggplot(df2, aes('x', 'discrete')) +
annotation_logticks(sides='l', size=.75, base=None) + geom_point())
with pytest.warns(PlotnineWarning):
p.draw_test()
def test_annotation_stripes_continuous_transformed():
pdf = mtcars.assign(am=pd.Categorical(mtcars.am))
p = (ggplot(pdf) +
annotation_stripes(fills=["red", "green", "blue"], alpha=0.1) +
geom_jitter(aes("hp", "wt", color="am"), random_state=5) +
scale_x_continuous(trans='log2'))
assert p == "annotation_stripes_continuous_transformed"
def round_2_plot():
if not os.path.exists(round_2_df_path):
eprint(f'Downloading {round_2_df_url} to {round_2_df_path}')
urlretrieve(round_2_df_url, round_2_df_path)
verify_checksum(round_2_df_checksum, round_2_df_path)
df = pd.read_json(round_2_df_path)
p = (
ggplot(df) + aes(x='char_percent', y='correct', color='Dataset') +
facet_wrap('Guessing_Model', nrow=1) + stat_summary_bin(
fun_data=mean_no_se, bins=20, shape='.', linetype='None',
size=0.5) + scale_y_continuous(breaks=np.linspace(0, 1, 6)) +
scale_x_continuous(breaks=[0, .5, 1]) +
coord_cartesian(ylim=[0, 0.7]) +
ggtitle('Round 2 Attacks and Models') +
xlab('Percent of Question Revealed') + ylab('Accuracy') + theme(
#legend_position='top', legend_box_margin=0, legend_title=element_blank(),
strip_text_x=element_text(margin={
't': 6,
'b': 6,
'l': 1,
'r': 5
})) +
scale_color_manual(values=['#FF3333', '#66CC00', '#3333FF', '#FFFF33'],
name='Questions'))
p.save('2019_tacl_trick/auto_fig/round_2_json.pdf', width=7.0, height=1.7)
def plot_time_curve_with_threshold(self):
toplot = self.aggregated.melt(
id_vars='hour',
value_vars=['number_bacteria', 'number_actin'],
value_name='counts',
var_name='Object')
colors = self.create_color_list()
myfig = (
ggplot(toplot, aes("hour", "counts", color="Object")) +
geom_point() + geom_line() + labels.xlab("Time [hours]") +
labels.ylab("Average number of objects/nuclei") +
pn.scale_colour_manual(values=colors,
labels=list(self.sel_channel_time.value),
name="") + pn.labs(colour="") +
pn.scale_x_continuous(
breaks=np.sort(self.result.hour.unique()),
labels=list(np.sort(self.result.hour.unique()).astype(str))))
self.time_curve_fig = myfig
self.out_plot2.clear_output()
with self.out_plot2:
display(myfig)
def plot_time_curve_by_channel(self, b=None):
"""Callback to polot time curve of number of bacteria/nuclei for
each selected channel. Called by plot_time_curve_button."""
if self.aggregated is None:
self.data_aggregation()
if len(self.sel_channel_time.value) == 0:
print("Select at least one channel")
else:
subset = self.aggregated[self.aggregated.channel.isin(
self.sel_channel_time.value)].copy(deep=True)
subset.loc[:, "channel"] = subset.channel.astype(
pd.CategoricalDtype(self.sel_channel_time.value, ordered=True))
colors = self.create_color_list()
myfig = (
ggplot(subset, aes("hour", "normalized", color="channel")) +
geom_point() + geom_line() + labels.xlab("Time [hours]") +
labels.ylab("Average number of bacteria/nuclei") +
pn.scale_colour_manual(
values=colors,
labels=list(self.sel_channel_time.value),
name="") + pn.labs(colour="") + pn.scale_x_continuous(
breaks=np.sort(self.result.hour.unique()),
labels=list(
np.sort(self.result.hour.unique()).astype(str))))
self.time_curve_fig = myfig
self.out_plot2.clear_output()
with self.out_plot2:
display(myfig)
def round_1_plot():
df = pd.read_csv('2019_tacl_trick/data/round_1.csv')
model_dtype = CategoricalDtype(['DAN', 'RNN', 'IR'], ordered=True)
df['Model'] = df['Model'].astype(model_dtype)
# This following is a hack so that the legend widths are the same across plots
def rename(x):
if x == 'Round 1 - IR Adversarial':
return 'Round 1 - IR Adversarial '
else:
return x
df['Dataset'] = df['Dataset'].map(rename)
p = (ggplot(df) + aes(x='x', y='y', color='Dataset') +
facet_wrap('Model', nrow=1) + geom_point(size=1.0, shape='o') +
scale_y_continuous(breaks=np.linspace(0, 1, 6), limits=[0, 0.6]) +
scale_x_continuous(breaks=[0, .5, 1]) +
xlab('Percent of Question Revealed') + ylab('Accuracy') +
ggtitle('Round 1 Attacks and Models') +
theme(strip_text_x=element_text(margin={
't': 6,
'b': 6,
'l': 1,
'r': 5
})) + scale_color_manual(
values=['#FF3333', '#66CC00', '#3333FF', '#FFFF33'],
name='Questions'))
p.save('2019_tacl_trick/auto_fig/round_1_csv.pdf', width=7.0, height=1.7)
def plot_train_test(ags):
frontiers = data.train_test(ags)
frontiers, model = data.train_test_model(frontiers)
labs = frontiers.sort_values('train_flops').groupby(
'elo').first().reset_index()
desc = f'log₁₀(test) = {model.params[1]:.1f} · log₁₀(train) + {model.params[2]:.1g} · elo + {model.params[0]:.0f}'
return (
pn.ggplot(
frontiers,
pn.aes(x='train_flops', y='test_flops', color='elo',
group='elo')) + pn.geom_line(size=.5, show_legend=False) +
pn.geom_line(pn.aes(y='test_flops_hat'),
size=.25,
show_legend=False,
linetype='dashed')
# + pn.geom_point(size=.5, show_legend=False)
+ pn.geom_text(pn.aes(label='elo.astype(int)'),
labs,
show_legend=False,
size=6,
nudge_y=+.2) + pn.scale_color_cmap(limits=(-1500, 0)) +
pn.scale_x_continuous(trans='log10') +
pn.scale_y_continuous(trans='log10') + pn.annotate(
'text', 1.5e13, 5e9, label=desc, ha='left', size=6, family='serif')
+ pn.labs(x='Train-time compute (FLOPS-seconds)',
y='Test-time compute (FLOPS-seconds)') + plot.IEEE())
def scatter_plot(df,
xcol,
ycol,
domain,
xname=None,
yname=None,
log=False,
width=6,
height=6,
clamp=True,
tickCount=5):
assert len(domain) == 2
POINT_SIZE = 0.5
DASH_PATTERN = (0, (3, 1))
if xname == None:
xname = xcol
if yname == None:
yname = ycol
# formater for axes' labels
ax_formatter = mizani.custom_format('{:n}')
if clamp: # clamp overflowing values if required
df = df.copy(deep=True)
df.loc[df[xcol] > domain[1], xcol] = domain[1]
df.loc[df[ycol] > domain[1], ycol] = domain[1]
# generate scatter plot
scatter = p9.ggplot(df)
scatter += p9.aes(x=xcol, y=ycol)
scatter += p9.geom_point(size=POINT_SIZE, na_rm=True)
scatter += p9.labs(x=xname, y=yname)
if log: # log scale
scatter += p9.scale_x_log10(limits=domain, labels=ax_formatter)
scatter += p9.scale_y_log10(limits=domain, labels=ax_formatter)
else:
scatter += p9.scale_x_continuous(limits=domain, labels=ax_formatter)
scatter += p9.scale_y_continuous(limits=domain, labels=ax_formatter)
#scatter += p9.theme_xkcd()
scatter += p9.theme_bw()
scatter += p9.theme(
panel_grid_major=p9.element_line(color='#666666', alpha=0.5))
scatter += p9.theme(figure_size=(width, height))
# generate additional lines
scatter += p9.geom_abline(intercept=0, slope=1,
linetype=DASH_PATTERN) # diagonal
scatter += p9.geom_vline(xintercept=domain[1],
linetype=DASH_PATTERN) # vertical rule
scatter += p9.geom_hline(yintercept=domain[1],
linetype=DASH_PATTERN) # horizontal rule
res = scatter
return res
def limits(x, y=None, xbreaks=None, ybreaks=None):
if y is None:
y = x
x0, x1 = x
y0, y1 = y
if xbreaks is None:
xbreaks = np.linspace(x0, x1, x1 - x0 + 1)
if ybreaks is None:
ybreaks = np.linspace(y0, y1, y1 - y0 + 1)
# We want these plots to continue to the top and left.
return [ gg.coord_cartesian(xlim=x, ylim=y),
gg.scale_x_continuous(limits=(x0, None), breaks=xbreaks),
gg.scale_y_continuous(limits=(y0, None), breaks=ybreaks) ]
return [ gg.scale_x_continuous(limits=x, breaks=xbreaks),
gg.scale_y_continuous(limits=y, breaks=ybreaks) ]
def __plot(
self,
plot_data,
x,
y,
colour,
lbl_x,
lbl_y,
facet,
facet_scales,
facet_by,
smoothed,
points,
error_bars,
save,
):
cbbPalette = [
"#000000",
"#E69F00",
"#56B4E9",
"#009E73",
"#0072B2",
"#D55E00",
"#CC79A7",
]
plt = ggplot(data=plot_data, mapping=aes(x=x, y=y, colour=colour))
plt += xlab(lbl_x)
plt += ylab(lbl_y)
# + facet_grid("site~", scales="free")
# + geom_line()
if facet:
# TODO: use facet as save
nrow, ncol = self.get_facet_rows(plot_data, facet_by)
plt += facet_wrap(facet_by, nrow=nrow, ncol=ncol, scales=facet_scales)
if points:
plt += geom_point()
if error_bars:
# TODO use generic way to compute them
pass
# self.plt += geom_errorbar(aes(ymin="ACI_mean - ACI_std", ymax="ACI_mean + ACI_std"))
# TODO: use smooth as save
if smoothed:
plt += geom_smooth(
method="mavg",
se=False,
method_args={"window": 4, "center": True, "min_periods": 1},
)
else:
plt += geom_line()
plt += scale_colour_manual(values=cbbPalette, guide=False)
plt += scale_x_continuous(labels=label_x)
plt += theme(figure_size=(15, 18), dpi=150)
if save:
plt.save(**save)
return plt
def plot_histogram(df_plot,
variable_column,
output_file='plot_distribution',
facet_column='none',
x_log10=False):
"""Plot plot_distribution to png.
Parameters
----------
df_plot : pandas.DataFrame
DataFrame with <variable_column> as a column.
variable_column : string
String of variable_column column to plot.
output_file : string
Basename of output file.
facet_column : string
Column to facet the plot by.
Returns
-------
NULL
"""
df_plot['x'] = df_plot[variable_column]
if x_log10:
if np.any(df_plot['x'].values < 0):
return 1
elif np.any(df_plot['x'].values == 0):
df_plot['x'] = np.log10(df_plot['x'].values + 1e-10)
variable_column = variable_column + ' (log10)'
else:
df_plot['x'] = np.log10(df_plot['x'].values)
variable_column = variable_column + ' (log10)'
gplt = plt9.ggplot(df_plot, plt9.aes(x='x'))
gplt = gplt + plt9.theme_bw()
gplt = gplt + plt9.geom_histogram(alpha=0.8)
gplt = gplt + plt9.scale_x_continuous(
# trans='log10',
# labels=comma_labels,
minor_breaks=0)
gplt = gplt + plt9.scale_y_continuous(
# trans='log10',
# labels=comma_labels,
minor_breaks=0)
gplt = gplt + plt9.labs(title='', x=variable_column)
gplt = gplt + plt9.theme(axis_text_x=plt9.element_text(angle=-45, hjust=0))
if facet_column != 'none':
gplt = gplt + plt9.facet_wrap('~ {}'.format(facet_column), ncol=5)
n_facets = df_plot[facet_column].nunique()
gplt.save('{}.png'.format(output_file),
dpi=300,
width=6 * (n_facets / 4),
height=4 * (n_facets / 4),
limitsize=False)
else:
gplt.save('{}.png'.format(output_file), dpi=300, width=4, height=4)
return 0
def scatter_plot2(df1, df2, xcol, ycol, domain, color1='black', color2='red', xname=None, yname=None, log=False, width=6, height=6, clamp=True, tickCount=5):
assert len(domain) == 2
POINT_SIZE = 1.5
DASH_PATTERN = (0, (6, 2))
if xname is None:
xname = xcol
if yname is None:
yname = ycol
# formatter for axes' labels
ax_formatter = mizani.custom_format('{:n}')
if clamp: # clamp overflowing values if required
df1 = df1.copy(deep=True)
df1.loc[df1[xcol] > domain[1], xcol] = domain[1]
df1.loc[df1[ycol] > domain[1], ycol] = domain[1]
df2 = df2.copy(deep=True)
df2.loc[df2[xcol] > domain[1], xcol] = domain[1]
df2.loc[df2[ycol] > domain[1], ycol] = domain[1]
# generate scatter plot
scatter = p9.ggplot(df1)
scatter += p9.aes(x=xcol, y=ycol)
scatter += p9.geom_point(size=POINT_SIZE, na_rm=True, color=color1, alpha=0.5)
scatter += p9.geom_point(size=POINT_SIZE, na_rm=True, data=df2, color=color2, alpha=0.5)
scatter += p9.labs(x=xname, y=yname)
# rug plots
scatter += p9.geom_rug(na_rm=True, sides="tr", color=color1, alpha=0.05)
scatter += p9.geom_rug(na_rm=True, sides="tr", data=df2, color=color2, alpha=0.05)
if log: # log scale
scatter += p9.scale_x_log10(limits=domain, labels=ax_formatter)
scatter += p9.scale_y_log10(limits=domain, labels=ax_formatter)
else:
scatter += p9.scale_x_continuous(limits=domain, labels=ax_formatter)
scatter += p9.scale_y_continuous(limits=domain, labels=ax_formatter)
# scatter += p9.theme_xkcd()
scatter += p9.theme_bw()
scatter += p9.theme(panel_grid_major=p9.element_line(color='#666666', alpha=0.5))
scatter += p9.theme(panel_grid_minor=p9.element_blank())
scatter += p9.theme(figure_size=(width, height))
scatter += p9.theme(text=p9.element_text(size=24, color="black"))
# generate additional lines
scatter += p9.geom_abline(intercept=0, slope=1, linetype=DASH_PATTERN) # diagonal
scatter += p9.geom_vline(xintercept=domain[1], linetype=DASH_PATTERN) # vertical rule
scatter += p9.geom_hline(yintercept=domain[1], linetype=DASH_PATTERN) # horizontal rule
res = scatter
return res
def test_annotation_logticks_base_8():
base = 8
df = pd.DataFrame({'x': base**np.arange(4)})
# The grid should align with the logticks
p = (ggplot(df, aes('x', 'x')) + annotation_logticks(sides='b', size=.75) +
geom_point() + scale_x_continuous(trans=log_trans(base=base)) +
theme(panel_grid_minor=element_line(color='green'),
panel_grid_major=element_line(color='red')))
assert p == 'annotation_logticks_base_8'
def plot_optimal_model_size(ags):
from statsmodels.formula import api as smf
results = {}
for b, g in ags.groupby('boardsize'):
ordered = g.sort_values('elo').copy()
ordered['params'] = g.width**2 * g.depth
left = np.log10(g.train_flops.min())
right = np.log10(g.train_flops.max())
for f in np.linspace(left, right, 11)[1:]:
subset = ordered[ordered.train_flops <= 10**f]
results[b, 10**f] = subset.params.iloc[-1]
df = pd.Series(results).reset_index()
df.columns = ['boardsize', 'approx_flops', 'params']
model = smf.ols('np.log10(params) ~ np.log10(approx_flops) + 1', df).fit()
left, right = np.log10(df.approx_flops.min()), np.log10(
df.approx_flops.max())
preds = pd.DataFrame({'approx_flops': 10**np.linspace(left, right, 21)})
preds['params'] = 10**model.predict(preds)
labs = df.sort_values('approx_flops').groupby(
'boardsize').last().reset_index()
labs['params'] = labs.apply(
lambda r: df[df.approx_flops <= r.approx_flops].params.max(), axis=1)
points = df.sort_values('approx_flops').groupby(
'boardsize').last().reset_index()
desc = f'log₁₀(params) = {model.params[1]:.2f} · log₁₀(compute) − {-model.params[0]:.1f}'
return (
pn.ggplot(df, pn.aes(x='approx_flops', y='params')) +
pn.geom_line(pn.aes(color='factor(boardsize)', group='boardsize'),
show_legend=False) +
pn.geom_line(data=preds, linetype='dashed', size=.25) +
pn.geom_point(pn.aes(color='factor(boardsize)', group='boardsize'),
data=points,
size=.5,
show_legend=False) +
pn.geom_text(pn.aes(
color='factor(boardsize)', group='boardsize', label='boardsize'),
data=labs,
nudge_y=+.5,
show_legend=False,
size=6) +
pn.annotate(
'text',
1e9, 2e7, label=desc, ha='left', size=6, family='serif') +
pn.scale_x_continuous(trans='log10') +
pn.scale_y_continuous(trans='log10') + pn.scale_color_hue(l=.4) +
pn.labs(x='Train-time compute (FLOPS-seconds)',
y='Optimal model size (params)') + plot.IEEE())
def plot_convergence(pile):
stops = range(100, int(len(pile) / 10), utils.bills_per_pound)
dist_stats = pd.DataFrame([get_sample_dist(pile, size) for size in stops])
return (
pn.ggplot(dist_stats) + pn.geom_line(pn.aes(x='size', y='mean')) +
pn.geom_line(
pn.aes(x='size', y='lower'), color='#FF5500', linetype='dotted') +
pn.geom_line(
pn.aes(x='size', y='upper'), color='#FF5500', linetype='dotted') +
pn.scale_x_continuous(breaks=stops) +
pn.theme(axis_text_x=pn.element_text(angle=270, hjust=1)))
def plot_fees(fees, title, y_axis, years, filename):
p = pn.ggplot(fees, pn.aes('year', y_axis, color = 'conference', shape = 'conference')) + \
pn.geom_point() + \
pn.geom_line() + \
pn.labs(title = title, x = 'Year', y = 'Fee (€)') + \
pn.ylim(0, 1000) + \
pn.theme_light() + \
pn.scale_x_continuous(breaks = years) + \
pn.scale_colour_discrete(name = 'Conference') + \
pn.scale_shape_discrete(name = 'Conference')
p.save(filename, width=6, height=3, dpi=300)
def plot_series(self, y: str,
series_state_df: pd.DataFrame) -> plotnine.ggplot:
"""
"""
aes_kwargs = dict(x="round_number",
y=y,
color="factor(game_id)",
group="factor(game_id)")
return (plotnine.ggplot(series_state_df) + plotnine.aes(**aes_kwargs) +
plotnine.geom_point() + plotnine.geom_line() +
plotnine.theme_light() +
plotnine.scale_x_continuous(breaks=range(1, 20, 1)))
def create(self, file_path: str) -> None:
(ggplot(self._data, aes("loc")) +
geom_histogram(bins=100, fill="#1e4f79") +
facet_grid(facets="category ~ .", scales='free_y') +
scale_x_continuous(trans=asinh_trans(), labels=asinh_labels) +
scale_y_continuous(labels=comma_format())
#+ scale_y_continuous(labels=lambda l: ["%.2f%%" % (v * 100 / len(self._data)) for v in l])
+ ggtitle("Class Sizes") + xlab("Lines of Code") +
ylab("Number of Classes") +
theme_classic(base_size=32, base_family="Helvetica") +
theme(text=element_text(size=32), subplots_adjust={"hspace": 0.1
})).save(file_path,
width=8,
height=18)
def create(self, file_path: str) -> None:
(ggplot(self._data, aes("value")) +
geom_histogram(bins=100, fill="#1e4f79") +
facet_wrap(facets="variable", scales="free", ncol=3) +
scale_x_continuous(trans=asinh_trans(), labels=asinh_labels) +
scale_y_continuous(labels=comma_format()) +
ggtitle("Distributions of QMOOD Quality Attributes") +
xlab("Quality Attribute Value") + ylab("Number of Projects") +
theme_classic(base_size=32, base_family="Helvetica") +
theme(text=element_text(size=32),
subplots_adjust={
"wspace": 0.35,
"hspace": 0.35
})).save(file_path, width=24, height=12)
def plot_training_curves(ags):
df = ags[ags.test_nodes == 64].copy()
df['g'] = df.run + df.test_nodes.astype(str)
return (pn.ggplot(
df,
pn.aes(x='train_flops',
y='400/np.log(10)*elo',
group='g',
color='factor(boardsize)')) + pn.geom_line() +
pn.geom_point(size=.5) + pn.scale_x_continuous(trans='log10') +
pn.scale_color_discrete(name='Boardsize') +
pn.labs(x='Training FLOPS',
y='Elo v. perfect play',
title='All agents\' training curves') + plot.mpl_theme() +
plot.poster_sizes())
def test_area_aesthetics():
p = (ggplot(df, aes('x', 'ymax+2', group='factor(z)')) + geom_area() +
geom_area(aes('x+width', alpha='z')) +
geom_area(aes('x+2*width', linetype='factor(z)'),
color='black',
fill=None,
size=2) +
geom_area(aes('x+3*width', color='z'), fill=None, size=2) +
geom_area(aes('x+4*width', fill='factor(z)')) +
geom_area(aes('x+5*width', size='z'), color='black', fill=None) +
scale_x_continuous(
breaks=[i * 2 * np.pi
for i in range(7)],
labels=['0'] + [r'${}\pi$'.format(2 * i) for i in range(1, 7)]))
assert p + _theme == 'area_aesthetics'
def plot_sample_efficiency(ags):
df = ags.query('boardsize == 9 & test_nodes == 64').copy()
df['params'] = df.train_flops / df.samples
return (
pn.ggplot(
df,
pn.aes(x='samples',
y='400/np.log(10)*elo',
color='params',
group='run')) + pn.geom_line() + pn.geom_point() +
pn.scale_x_continuous(trans='log10') +
pn.scale_color_continuous(trans='log10', name='Params') + pn.labs(
title=
'Bigger networks might not be comute efficient, but they are sample efficient',
y='Elo v. perfect play',
x='Train FLOPS') + plot.mpl_theme() + plot.poster_sizes() +
plot.no_colorbar_ticks())
def plot_params(ags):
df = ags.query('boardsize == 9 & test_nodes == 64').copy()
df['params'] = df.train_flops / df.samples
return (
pn.ggplot(
df,
pn.aes(x='train_flops',
y='400/np.log(10)*elo',
color='params',
group='run')) + pn.geom_line() + pn.geom_point() +
pn.scale_x_continuous(trans='log10') +
pn.scale_color_continuous(trans='log10', name='Params') + pn.labs(
title=
'Smaller networks are more compute efficient for lower performances, but plateau earlier',
y='Elo v. perfect play',
x='Train FLOPS') + plot.mpl_theme() + plot.poster_sizes() +
plot.no_colorbar_ticks())
def test_area_aesthetics():
p = (ggplot(df, aes('x', 'ymax+2', group='factor(z)')) +
geom_area() +
geom_area(aes('x+width', alpha='z')) +
geom_area(aes('x+2*width', linetype='factor(z)'),
color='black', fill=None, size=2) +
geom_area(aes('x+3*width', color='z'),
fill=None, size=2) +
geom_area(aes('x+4*width', fill='factor(z)')) +
geom_area(aes('x+5*width', size='z'),
color='black', fill=None) +
scale_x_continuous(
breaks=[i*2*np.pi for i in range(7)],
labels=['0'] + [r'${}\pi$'.format(2*i) for i in range(1, 7)])
)
assert p + _theme == 'area_aesthetics'
def plot_resid_var_trends(ags):
resid_var = data.residual_vars(ags)
return (
pn.ggplot(
resid_var,
pn.aes(x='ratio',
y='rv',
color='factor(predicted)',
group='predicted')) + pn.geom_line(size=2) +
pn.geom_text(pn.aes(label='seen'), nudge_y=-.1, size=14) +
pn.geom_point(size=4) + pn.scale_x_continuous(trans='log10') +
pn.scale_y_continuous(trans='log10') +
pn.scale_color_discrete(name='Predicted frontier') + pn.labs(
x='(cost of observed frontier)/(cost of predicted frontier)',
y='residual variance in performance',
title=
'Frontiers of small problems are good, cheap proxies for frontiers of expensive problems'
) + plot.mpl_theme() + plot.poster_sizes())
def test_inplace_add():
p = _p = ggplot(df)
p += aes('x', 'y')
assert p is _p
p += geom_point()
assert p is _p
p += stat_identity()
assert p is _p
p += scale_x_continuous()
assert p is _p
with pytest.warns(PlotnineWarning):
# Warning for; replacing existing scale added above
p += xlim(0, 10)
assert p is _p
p += lims(y=(0, 10))
assert p is _p
p += labs(x='x')
assert p is _p
p += coord_trans()
assert p is _p
p += facet_null()
assert p is _p
p += annotate('point', 5, 5, color='red', size=5)
assert p is _p
p += guides()
assert p is _p
p += theme_gray()
assert p is _p
th = _th = theme_gray()
th += theme(aspect_ratio=1)
assert th is _th
def test_wrong_bases():
# x axis not transformed
p = (ggplot(df, aes('x', 'x'))
+ annotation_logticks(sides='b', size=.75, base=10)
+ geom_point()
)
with pytest.warns(PlotnineWarning):
p.draw_test()
# x axis not transform, but ticks requested for a different base
p = (ggplot(df, aes('x', 'x'))
+ annotation_logticks(sides='b', size=.75, base=10)
+ scale_x_continuous(trans=log_trans(8))
+ geom_point()
)
with pytest.warns(PlotnineWarning):
p.draw_test()
def plot_flops_frontier(ags):
df = data.modelled_elos(ags)
return (
pn.ggplot(
df,
pn.aes(
x='train_flops', color='factor(boardsize)', group='boardsize'))
+ pn.geom_line(pn.aes(y='400/np.log(10)*elo'), size=2) + pn.geom_line(
pn.aes(y='400/np.log(10)*elohat'), size=1, linetype='dashed') +
pn.labs(
x='Training FLOPS',
y='Elo v. perfect play',
title=
'Performance is a sigmoid of compute, linearly scaled by board size'
) + pn.scale_x_continuous(trans='log10') +
pn.scale_color_discrete(name='Boardsize') +
pn.coord_cartesian(None,
(None, 0)) + plot.mpl_theme() + plot.poster_sizes())
def yoy_growth():
"""
This creates figures showing the number of questions versus year in dataset
"""
with open('data/external/datasets/qanta.mapped.2018.04.18.json') as f:
year_pages = defaultdict(set)
year_questions = Counter()
for q in json.load(f)['questions']:
if q['page'] is not None:
year_pages[q['year']].add(q['page'])
year_questions[q['year']] += 1
start_year = min(year_pages)
# 2017 is the earlier year we have a full year's worth of data, including partial 2018 isn't accurate
end_year = min(2017, max(year_pages))
upto_year_pages = defaultdict(set)
upto_year_questions = Counter()
for upto_y in range(start_year, end_year + 1):
for curr_y in range(start_year, upto_y + 1):
upto_year_questions[upto_y] += year_questions[curr_y]
for page in year_pages[curr_y]:
upto_year_pages[upto_y].add(page)
year_page_counts = {}
for y, pages in upto_year_pages.items():
year_page_counts[y] = len(pages)
year_page_counts
year_rows = []
for y, page_count in year_page_counts.items():
year_rows.append({'year': y, 'value': page_count, 'Quantity': 'Distinct Answers'})
year_rows.append({'year': y, 'Quantity': 'Total Questions', 'value': upto_year_questions[y]})
year_df = pd.DataFrame(year_rows)
count_cat = CategoricalDtype(categories=['Total Questions', 'Distinct Answers'], ordered=True)
year_df['Quantity'] = year_df['Quantity'].astype(count_cat)
eprint(year_df[year_df.Quantity == 'Total Questions'])
p = (
ggplot(year_df)
+ aes(x='year', y='value', color='Quantity')
+ geom_line() + geom_point()
+ xlab('Year')
+ ylab('Count up to Year (inclusive)')
+ theme_fs()
+ scale_x_continuous(breaks=list(range(start_year, end_year + 1, 2)))
)
p.save(path.join(output_path, 'question_answer_counts.pdf'))
def plot_test(ags):
df = ags.query('boardsize == 9').groupby('run').apply(
lambda df: df[df.idx == df.idx.max()]).copy()
df['test_flops'] = df.test_nodes * (df.train_flops / df.samples)
subset = df.query('test_nodes == 64').sort_values('test_flops')
selection = [
subset.loc[ELO * subset.elo > e].iloc[0].run
for e in np.linspace(-2000, -500, 4)
]
df = df[df.run.isin(selection)].copy()
df['params'] = df.width**2 * df.depth
df['arch'] = df.apply(lambda r: '{depth}×{width}'.format(**r), axis=1)
labels = df.sort_values('test_flops').reset_index(
drop=True).groupby('run').first().reset_index()
return (pn.ggplot(
df, pn.aes(x='test_flops', y='ELO*elo', color='params', group='run')) +
pn.geom_point(size=.25, show_legend=False) +
pn.geom_line(size=.5, show_legend=False) +
pn.geom_text(pn.aes(label='test_nodes'),
nudge_y=-50,
show_legend=False,
size=4,
va='top') + pn.geom_text(pn.aes(label='test_nodes'),
nudge_y=-50,
show_legend=False,
size=4,
va='top') +
pn.geom_text(pn.aes(label='arch'),
data=labels,
show_legend=False,
size=6,
nudge_x=-.1,
ha='right') + pn.scale_x_continuous(trans='log10') +
pn.scale_color_cmap('plasma',
trans='log10',
limits=(df.params.min(), 10 * df.params.max()))
+ pn.coord_cartesian(
(3.5, None)) + pn.labs(x='Test-time compute (FLOPS-seconds)',
y='Elo v. perfect play') + plot.IEEE())
def create_state_plot(
self,
x: str,
y: str,
data: pd.DataFrame,
aes_kwargs: Tuple[str, str] = None) -> plotnine.ggplot:
"""
"""
aes_kwargs = ({} if aes_kwargs is None else aes_kwargs)
x_breaks = data[x]
max_y = max(data[y])
y_interval = int(max_y / 20) + 1
y_breaks = range(0, max_y + 1, y_interval)
return (plotnine.ggplot(data) + plotnine.aes(x=x, y=y, **aes_kwargs) +
plotnine.geom_line() + plotnine.geom_point() +
plotnine.theme_light() +
plotnine.scale_x_continuous(breaks=x_breaks) +
plotnine.scale_y_continuous(breaks=y_breaks))
def quick_color_check(target_matrix, source_matrix, num_chips):
""" Quickly plot target matrix values against source matrix values to determine
over saturated color chips or other issues.
Inputs:
source_matrix = a 22x4 matrix containing the average red value, average green value, and
average blue value for each color chip of the source image
target_matrix = a 22x4 matrix containing the average red value, average green value, and
average blue value for each color chip of the target image
num_chips = number of color card chips included in the matrices (integer)
:param source_matrix: numpy.ndarray
:param target_matrix: numpy.ndarray
:param num_chips: int
"""
# Imports
from plotnine import ggplot, geom_point, geom_smooth, theme_seaborn, facet_grid, geom_label, scale_x_continuous, \
scale_y_continuous, scale_color_manual, aes
import pandas as pd
# Extract and organize matrix info
tr = target_matrix[:num_chips, 1:2]
tg = target_matrix[:num_chips, 2:3]
tb = target_matrix[:num_chips, 3:4]
sr = source_matrix[:num_chips, 1:2]
sg = source_matrix[:num_chips, 2:3]
sb = source_matrix[:num_chips, 3:4]
# Create columns of color labels
red = []
blue = []
green = []
for i in range(num_chips):
red.append('red')
blue.append('blue')
green.append('green')
# Make a column of chip numbers
chip = np.arange(0, num_chips).reshape((num_chips, 1))
chips = np.row_stack((chip, chip, chip))
# Combine info
color_data_r = np.column_stack((sr, tr, red))
color_data_g = np.column_stack((sg, tg, green))
color_data_b = np.column_stack((sb, tb, blue))
all_color_data = np.row_stack((color_data_b, color_data_g, color_data_r))
# Create a dataframe with headers
dataset = pd.DataFrame({'source': all_color_data[:, 0], 'target': all_color_data[:, 1],
'color': all_color_data[:, 2]})
# Add chip numbers to the dataframe
dataset['chip'] = chips
dataset = dataset.astype({'color': str, 'chip': str, 'target': float, 'source': float})
# Make the plot
p1 = ggplot(dataset, aes(x='target', y='source', color='color', label='chip')) + \
geom_point(show_legend=False, size=2) + \
geom_smooth(method='lm', size=.5, show_legend=False) + \
theme_seaborn() + facet_grid('.~color') + \
geom_label(angle=15, size=7, nudge_y=-.25, nudge_x=.5, show_legend=False) + \
scale_x_continuous(limits=(-5, 270)) + scale_y_continuous(limits=(-5, 275)) + \
scale_color_manual(values=['blue', 'green', 'red'])
# Reset debug
if params.debug is not None:
if params.debug == 'print':
p1.save(os.path.join(params.debug_outdir, 'color_quick_check.png'))
elif params.debug == 'plot':
print(p1)
def analyze_nir_intensity(gray_img, mask, bins=256, histplot=False):
"""This function calculates the intensity of each pixel associated with the plant and writes the values out to
a file. It can also print out a histogram plot of pixel intensity and a pseudocolor image of the plant.
Inputs:
gray_img = 8- or 16-bit grayscale image data
mask = Binary mask made from selected contours
bins = number of classes to divide spectrum into
histplot = if True plots histogram of intensity values
Returns:
analysis_images = NIR histogram image
:param gray_img: numpy array
:param mask: numpy array
:param bins: int
:param histplot: bool
:return analysis_images: list
"""
params.device += 1
# apply plant shaped mask to image
mask1 = binary_threshold(mask, 0, 255, 'light')
mask1 = (mask1 / 255)
masked = np.multiply(gray_img, mask1)
# calculate histogram
if gray_img.dtype == 'uint16':
maxval = 65536
else:
maxval = 256
# Make a pseudo-RGB image
rgbimg = cv2.cvtColor(gray_img, cv2.COLOR_GRAY2BGR)
hist_nir, hist_bins = np.histogram(masked, bins, (1, maxval))
hist_bins1 = hist_bins[:-1]
hist_bins2 = [float(round(l, 2)) for l in hist_bins1]
hist_nir1 = [float(l) for l in hist_nir]
# make hist percentage for plotting
pixels = cv2.countNonZero(mask1)
hist_percent = (hist_nir / float(pixels)) * 100
# No longer returning a pseudocolored image
# make mask to select the background
# mask_inv = cv2.bitwise_not(mask)
# img_back = cv2.bitwise_and(rgbimg, rgbimg, mask=mask_inv)
# img_back1 = cv2.applyColorMap(img_back, colormap=1)
# mask the background and color the plant with color scheme 'jet'
# cplant = cv2.applyColorMap(rgbimg, colormap=2)
# masked1 = apply_mask(cplant, mask, 'black')
masked1 = cv2.bitwise_and(rgbimg, rgbimg, mask=mask)
# cplant_back = cv2.add(masked1, img_back1)
if params.debug is not None:
if params.debug == "print":
print_image(masked1, os.path.join(params.debug_outdir, str(params.device) + "_masked_nir_plant.jpg"))
if params.debug == "plot":
plot_image(masked1)
analysis_images = []
if histplot is True:
hist_x = hist_percent
bin_labels = np.arange(0, bins)
dataset = pd.DataFrame({'Grayscale pixel intensity': bin_labels,
'Proportion of pixels (%)': hist_x})
fig_hist = (ggplot(data=dataset,
mapping=aes(x='Grayscale pixel intensity',
y='Proportion of pixels (%)'))
+ geom_line(color='red')
+ scale_x_continuous(breaks=list(range(0, bins, 25))))
analysis_images.append(fig_hist)
if params.debug == "print":
fig_hist.save(os.path.join(params.debug_outdir, str(params.device) + '_nir_hist.png'))
elif params.debug == "plot":
print(fig_hist)
outputs.add_observation(variable='nir_frequencies', trait='near-infrared frequencies',
method='plantcv.plantcv.analyze_nir_intensity', scale='frequency', datatype=list,
value=hist_nir1, label=hist_bins2)
# Store images
outputs.images.append(analysis_images)
return analysis_images
def analyze_color(rgb_img, mask, hist_plot_type=None):
"""Analyze the color properties of an image object
Inputs:
rgb_img = RGB image data
mask = Binary mask made from selected contours
hist_plot_type = 'None', 'all', 'rgb','lab' or 'hsv'
Returns:
analysis_image = histogram output
:param rgb_img: numpy.ndarray
:param mask: numpy.ndarray
:param hist_plot_type: str
:return analysis_images: list
"""
params.device += 1
if len(np.shape(rgb_img)) < 3:
fatal_error("rgb_img must be an RGB image")
# Mask the input image
masked = cv2.bitwise_and(rgb_img, rgb_img, mask=mask)
# Extract the blue, green, and red channels
b, g, r = cv2.split(masked)
# Convert the BGR image to LAB
lab = cv2.cvtColor(masked, cv2.COLOR_BGR2LAB)
# Extract the lightness, green-magenta, and blue-yellow channels
l, m, y = cv2.split(lab)
# Convert the BGR image to HSV
hsv = cv2.cvtColor(masked, cv2.COLOR_BGR2HSV)
# Extract the hue, saturation, and value channels
h, s, v = cv2.split(hsv)
# Color channel dictionary
channels = {"b": b, "g": g, "r": r, "l": l, "m": m, "y": y, "h": h, "s": s, "v": v}
# Histogram plot types
hist_types = {"ALL": ("b", "g", "r", "l", "m", "y", "h", "s", "v"),
"RGB": ("b", "g", "r"),
"LAB": ("l", "m", "y"),
"HSV": ("h", "s", "v")}
if hist_plot_type is not None and hist_plot_type.upper() not in hist_types:
fatal_error("The histogram plot type was " + str(hist_plot_type) +
', but can only be one of the following: None, "all", "rgb", "lab", or "hsv"!')
# Store histograms, plotting colors, and plotting labels
histograms = {
"b": {"label": "blue", "graph_color": "blue",
"hist": [float(l[0]) for l in cv2.calcHist([channels["b"]], [0], mask, [256], [0, 255])]},
"g": {"label": "green", "graph_color": "forestgreen",
"hist": [float(l[0]) for l in cv2.calcHist([channels["g"]], [0], mask, [256], [0, 255])]},
"r": {"label": "red", "graph_color": "red",
"hist": [float(l[0]) for l in cv2.calcHist([channels["r"]], [0], mask, [256], [0, 255])]},
"l": {"label": "lightness", "graph_color": "dimgray",
"hist": [float(l[0]) for l in cv2.calcHist([channels["l"]], [0], mask, [256], [0, 255])]},
"m": {"label": "green-magenta", "graph_color": "magenta",
"hist": [float(l[0]) for l in cv2.calcHist([channels["m"]], [0], mask, [256], [0, 255])]},
"y": {"label": "blue-yellow", "graph_color": "yellow",
"hist": [float(l[0]) for l in cv2.calcHist([channels["y"]], [0], mask, [256], [0, 255])]},
"h": {"label": "hue", "graph_color": "blueviolet",
"hist": [float(l[0]) for l in cv2.calcHist([channels["h"]], [0], mask, [256], [0, 255])]},
"s": {"label": "saturation", "graph_color": "cyan",
"hist": [float(l[0]) for l in cv2.calcHist([channels["s"]], [0], mask, [256], [0, 255])]},
"v": {"label": "value", "graph_color": "orange",
"hist": [float(l[0]) for l in cv2.calcHist([channels["v"]], [0], mask, [256], [0, 255])]}
}
# Create list of bin labels for 8-bit data
binval = np.arange(0, 256)
bin_values = [l for l in binval]
analysis_images = []
# Create a dataframe of bin labels and histogram data
dataset = pd.DataFrame({'bins': binval, 'blue': histograms["b"]["hist"],
'green': histograms["g"]["hist"], 'red': histograms["r"]["hist"],
'lightness': histograms["l"]["hist"], 'green-magenta': histograms["m"]["hist"],
'blue-yellow': histograms["y"]["hist"], 'hue': histograms["h"]["hist"],
'saturation': histograms["s"]["hist"], 'value': histograms["v"]["hist"]})
# Make the histogram figure using plotnine
if hist_plot_type is not None:
if hist_plot_type.upper() == 'RGB':
df_rgb = pd.melt(dataset, id_vars=['bins'], value_vars=['blue', 'green', 'red'],
var_name='Color Channel', value_name='Pixels')
hist_fig = (ggplot(df_rgb, aes(x='bins', y='Pixels', color='Color Channel'))
+ geom_line()
+ scale_x_continuous(breaks=list(range(0, 256, 25)))
+ scale_color_manual(['blue', 'green', 'red'])
)
analysis_images.append(hist_fig)
elif hist_plot_type.upper() == 'LAB':
df_lab = pd.melt(dataset, id_vars=['bins'],
value_vars=['lightness', 'green-magenta', 'blue-yellow'],
var_name='Color Channel', value_name='Pixels')
hist_fig = (ggplot(df_lab, aes(x='bins', y='Pixels', color='Color Channel'))
+ geom_line()
+ scale_x_continuous(breaks=list(range(0, 256, 25)))
+ scale_color_manual(['yellow', 'magenta', 'dimgray'])
)
analysis_images.append(hist_fig)
elif hist_plot_type.upper() == 'HSV':
df_hsv = pd.melt(dataset, id_vars=['bins'],
value_vars=['hue', 'saturation', 'value'],
var_name='Color Channel', value_name='Pixels')
hist_fig = (ggplot(df_hsv, aes(x='bins', y='Pixels', color='Color Channel'))
+ geom_line()
+ scale_x_continuous(breaks=list(range(0, 256, 25)))
+ scale_color_manual(['blueviolet', 'cyan', 'orange'])
)
analysis_images.append(hist_fig)
elif hist_plot_type.upper() == 'ALL':
s = pd.Series(['blue', 'green', 'red', 'lightness', 'green-magenta',
'blue-yellow', 'hue', 'saturation', 'value'], dtype="category")
color_channels = ['blue', 'yellow', 'green', 'magenta', 'blueviolet',
'dimgray', 'red', 'cyan', 'orange']
df_all = pd.melt(dataset, id_vars=['bins'], value_vars=s, var_name='Color Channel',
value_name='Pixels')
hist_fig = (ggplot(df_all, aes(x='bins', y='Pixels', color='Color Channel'))
+ geom_line()
+ scale_x_continuous(breaks=list(range(0, 256, 25)))
+ scale_color_manual(color_channels)
)
analysis_images.append(hist_fig)
# Hue values of zero are red but are also the value for pixels where hue is undefined
# The hue value of a pixel will be undefined when the color values are saturated
# Therefore, hue values of zero are excluded from the calculations below
# Calculate the median hue value
# The median is rescaled from the encoded 0-179 range to the 0-359 degree range
hue_median = np.median(h[np.where(h > 0)]) * 2
# Calculate the circular mean and standard deviation of the encoded hue values
# The mean and standard-deviation are rescaled from the encoded 0-179 range to the 0-359 degree range
hue_circular_mean = stats.circmean(h[np.where(h > 0)], high=179, low=0) * 2
hue_circular_std = stats.circstd(h[np.where(h > 0)], high=179, low=0) * 2
# Store into lists instead for pipeline and print_results
# stats_dict = {'mean': circular_mean, 'std' : circular_std, 'median': median}
# Plot or print the histogram
if hist_plot_type is not None:
if params.debug == 'print':
hist_fig.save(os.path.join(params.debug_outdir, str(params.device) + '_analyze_color_hist.png'))
elif params.debug == 'plot':
print(hist_fig)
# Store into global measurements
# RGB signal values are in an unsigned 8-bit scale of 0-255
rgb_values = [i for i in range(0, 256)]
# Hue values are in a 0-359 degree scale, every 2 degrees at the midpoint of the interval
hue_values = [i * 2 + 1 for i in range(0, 180)]
# Percentage values on a 0-100 scale (lightness, saturation, and value)
percent_values = [round((i / 255) * 100, 2) for i in range(0, 256)]
# Diverging values on a -128 to 127 scale (green-magenta and blue-yellow)
diverging_values = [i for i in range(-128, 128)]
# outputs.measurements['color_data'] = {
# 'histograms': {
# 'blue': {'signal_values': rgb_values, 'frequency': histograms["b"]["hist"]},
# 'green': {'signal_values': rgb_values, 'frequency': histograms["g"]["hist"]},
# 'red': {'signal_values': rgb_values, 'frequency': histograms["r"]["hist"]},
# 'lightness': {'signal_values': percent_values, 'frequency': histograms["l"]["hist"]},
# 'green-magenta': {'signal_values': diverging_values, 'frequency': histograms["m"]["hist"]},
# 'blue-yellow': {'signal_values': diverging_values, 'frequency': histograms["y"]["hist"]},
# 'hue': {'signal_values': hue_values, 'frequency': histograms["h"]["hist"]},
# 'saturation': {'signal_values': percent_values, 'frequency': histograms["s"]["hist"]},
# 'value': {'signal_values': percent_values, 'frequency': histograms["v"]["hist"]}
# },
# 'color_features': {
# 'hue_circular_mean': hue_circular_mean,
# 'hue_circular_std': hue_circular_std,
# 'hue_median': hue_median
# }
# }
outputs.add_observation(variable='blue_frequencies', trait='blue frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["b"]["hist"], label=rgb_values)
outputs.add_observation(variable='green_frequencies', trait='green frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["g"]["hist"], label=rgb_values)
outputs.add_observation(variable='red_frequencies', trait='red frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["r"]["hist"], label=rgb_values)
outputs.add_observation(variable='lightness_frequencies', trait='lightness frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["l"]["hist"], label=percent_values)
outputs.add_observation(variable='green-magenta_frequencies', trait='green-magenta frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["m"]["hist"], label=diverging_values)
outputs.add_observation(variable='blue-yellow_frequencies', trait='blue-yellow frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["y"]["hist"], label=diverging_values)
outputs.add_observation(variable='hue_frequencies', trait='hue frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["h"]["hist"], label=hue_values)
outputs.add_observation(variable='saturation_frequencies', trait='saturation frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["s"]["hist"], label=percent_values)
outputs.add_observation(variable='value_frequencies', trait='value frequencies',
method='plantcv.plantcv.analyze_color', scale='frequency', datatype=list,
value=histograms["v"]["hist"], label=percent_values)
outputs.add_observation(variable='hue_circular_mean', trait='hue circular mean',
method='plantcv.plantcv.analyze_color', scale='degrees', datatype=float,
value=hue_circular_mean, label='degrees')
outputs.add_observation(variable='hue_circular_std', trait='hue circular standard deviation',
method='plantcv.plantcv.analyze_color', scale='degrees', datatype=float,
value=hue_median, label='degrees')
outputs.add_observation(variable='hue_median', trait='hue median',
method='plantcv.plantcv.analyze_color', scale='degrees', datatype=float,
value=hue_median, label='degrees')
# Store images
outputs.images.append(analysis_images)
return analysis_images
def plot_char_percent_vs_accuracy_smooth(self, expo=False, no_models=False, columns=False):
if self.y_max is not None:
limits = [0, float(self.y_max)]
eprint(f'Setting limits to: {limits}')
else:
limits = [0, 1]
if expo:
if os.path.exists('data/external/all_human_gameplay.json') and not self.no_humans:
with open('data/external/all_human_gameplay.json') as f:
all_gameplay = json.load(f)
frames = []
for event, name in [('parents', 'Intermediate'), ('maryland', 'Expert'), ('live', 'National')]:
if self.merge_humans:
name = 'Human'
gameplay = all_gameplay[event]
if event != 'live':
control_correct_positions = gameplay['control_correct_positions']
control_wrong_positions = gameplay['control_wrong_positions']
control_positions = control_correct_positions + control_wrong_positions
control_positions = np.array(control_positions)
control_result = np.array(len(control_correct_positions) * [1] + len(control_wrong_positions) * [0])
argsort_control = np.argsort(control_positions)
control_x = control_positions[argsort_control]
control_sorted_result = control_result[argsort_control]
control_y = control_sorted_result.cumsum() / control_sorted_result.shape[0]
control_df = pd.DataFrame({'correct': control_y, 'char_percent': control_x})
control_df['Dataset'] = 'Regular Test'
control_df['Guessing_Model'] = f' {name}'
frames.append(control_df)
adv_correct_positions = gameplay['adv_correct_positions']
adv_wrong_positions = gameplay['adv_wrong_positions']
adv_positions = adv_correct_positions + adv_wrong_positions
adv_positions = np.array(adv_positions)
adv_result = np.array(len(adv_correct_positions) * [1] + len(adv_wrong_positions) * [0])
argsort_adv = np.argsort(adv_positions)
adv_x = adv_positions[argsort_adv]
adv_sorted_result = adv_result[argsort_adv]
adv_y = adv_sorted_result.cumsum() / adv_sorted_result.shape[0]
adv_df = pd.DataFrame({'correct': adv_y, 'char_percent': adv_x})
adv_df['Dataset'] = 'IR Adversarial'
adv_df['Guessing_Model'] = f' {name}'
frames.append(adv_df)
if len(gameplay['advneural_correct_positions']) > 0:
adv_correct_positions = gameplay['advneural_correct_positions']
adv_wrong_positions = gameplay['advneural_wrong_positions']
adv_positions = adv_correct_positions + adv_wrong_positions
adv_positions = np.array(adv_positions)
adv_result = np.array(len(adv_correct_positions) * [1] + len(adv_wrong_positions) * [0])
argsort_adv = np.argsort(adv_positions)
adv_x = adv_positions[argsort_adv]
adv_sorted_result = adv_result[argsort_adv]
adv_y = adv_sorted_result.cumsum() / adv_sorted_result.shape[0]
adv_df = pd.DataFrame({'correct': adv_y, 'char_percent': adv_x})
adv_df['Dataset'] = 'RNN Adversarial'
adv_df['Guessing_Model'] = f' {name}'
frames.append(adv_df)
human_df = pd.concat(frames)
human_vals = sort_humans(list(human_df['Guessing_Model'].unique()))
human_dtype = CategoricalDtype(human_vals, ordered=True)
human_df['Guessing_Model'] = human_df['Guessing_Model'].astype(human_dtype)
dataset_dtype = CategoricalDtype(['Regular Test', 'IR Adversarial', 'RNN Adversarial'], ordered=True)
human_df['Dataset'] = human_df['Dataset'].astype(dataset_dtype)
if no_models:
p = ggplot(human_df) + geom_point(shape='.')
else:
df = self.char_plot_df
if 1 not in self.rounds:
df = df[df['Dataset'] != 'Round 1 - IR Adversarial']
if 2 not in self.rounds:
df = df[df['Dataset'] != 'Round 2 - IR Adversarial']
df = df[df['Dataset'] != 'Round 2 - RNN Adversarial']
p = ggplot(df)
if self.save_df is not None:
eprint(f'Saving df to: {self.save_df}')
df.to_json(self.save_df)
if os.path.exists('data/external/all_human_gameplay.json') and not self.no_humans:
eprint('Loading human data')
p = p + geom_line(data=human_df)
if columns:
facet_conf = facet_wrap('Guessing_Model', ncol=1)
else:
facet_conf = facet_wrap('Guessing_Model', nrow=1)
if not no_models:
if self.mvg_avg_char:
chart = stat_smooth(method='mavg', se=False, method_args={'window': 400})
else:
chart = stat_summary_bin(fun_data=mean_no_se, bins=20, shape='.', linetype='None', size=0.5)
else:
chart = None
p = (
p + facet_conf
+ aes(x='char_percent', y='correct', color='Dataset')
)
if chart is not None:
p += chart
p = (
p
+ scale_y_continuous(breaks=np.linspace(0, 1, 6))
+ scale_x_continuous(breaks=[0, .5, 1])
+ coord_cartesian(ylim=limits)
+ xlab('Percent of Question Revealed')
+ ylab('Accuracy')
+ theme(
#legend_position='top', legend_box_margin=0, legend_title=element_blank(),
strip_text_x=element_text(margin={'t': 6, 'b': 6, 'l': 1, 'r': 5})
)
+ scale_color_manual(values=['#FF3333', '#66CC00', '#3333FF', '#FFFF33'], name='Questions')
)
if self.title != '':
p += ggtitle(self.title)
return p
else:
if self.save_df is not None:
eprint(f'Saving df to: {self.save_df}')
df.to_json(self.save_df)
return (
ggplot(self.char_plot_df)
+ aes(x='char_percent', y='correct', color='Guessing_Model')
+ stat_smooth(method='mavg', se=False, method_args={'window': 500})
+ scale_y_continuous(breaks=np.linspace(0, 1, 6))
+ coord_cartesian(ylim=limits)
)
def syntactic_diversity_plots():
with open('data/external/syntactic_diversity_table.json') as f:
rows = json.load(f)
parse_df = pd.DataFrame(rows)
parse_df['parse_ratio'] = parse_df['unique_parses'] / parse_df['parses']
melt_df = pd.melt(
parse_df,
id_vars=['dataset', 'depth', 'overlap', 'parses'],
value_vars=['parse_ratio', 'unique_parses'],
var_name='metric',
value_name='y'
)
def label_facet(name):
if name == 'parse_ratio':
return 'Average Unique Parses per Instance'
elif name == 'unique_parses':
return 'Count of Unique Parses'
def label_y(ys):
formatted_ys = []
for y in ys:
y = str(y)
if y.endswith('000.0'):
formatted_ys.append(y[:-5] + 'K')
else:
formatted_ys.append(y)
return formatted_ys
p = (
ggplot(melt_df)
+ aes(x='depth', y='y', color='dataset')
+ facet_wrap('metric', scales='free_y', nrow=2, labeller=label_facet)
+ geom_line() + geom_point()
+ xlab('Parse Truncation Depth') + ylab('')
+ scale_color_discrete(name='Dataset')
+ scale_y_continuous(labels=label_y)
+ scale_x_continuous(
breaks=list(range(1, 11)),
minor_breaks=list(range(1, 11)),
limits=[1, 10])
+ theme_fs()
)
p.save(path.join(output_path, 'syn_div_plot.pdf'))
p = (
ggplot(parse_df)
+ aes(x='depth', y='unique_parses', color='dataset')
+ geom_line() + geom_point()
+ xlab('Parse Truncation Depth')
+ ylab('Count of Unique Parses')
+ scale_color_discrete(name='Dataset')
+ scale_x_continuous(
breaks=list(range(1, 11)),
minor_breaks=list(range(1, 11)),
limits=[1, 10])
+ theme_fs()
)
p.save(path.join(output_path, 'n_unique_parses.pdf'))
p = (
ggplot(parse_df)
+ aes(x='depth', y='parse_ratio', color='dataset')
+ geom_line() + geom_point()
+ xlab('Parse Truncation Depth')
+ ylab('Average Unique Parses per Instance')
+ scale_color_discrete(name='Dataset')
+ scale_x_continuous(breaks=list(range(1, 11)), minor_breaks=list(range(1, 11)), limits=[1, 10])
+ scale_y_continuous(limits=[0, 1])
+ theme_fs()
)
p.save(path.join(output_path, 'parse_ratio.pdf'))