def __run_rf_single(df_x, s_y, ntree, mtry, node_size, mode, seed, set_label,
plot):
"""Runs random forest once."""
my_y = s_y
if mode == 'classification':
my_y = vectors.FactorVector(s_y)
r('set.seed(' + str(seed) + ')')
random_forest = importr('randomForest')
data_rf = random_forest.randomForest(x=df_x,
y=my_y,
ntree=ntree,
mtry=mtry,
nodesize=node_size,
importance=rinterface.TRUE,
keep_forest=rinterface.TRUE,
proximity=rinterface.TRUE)
# TODO do we use these plots?
# previous implementation would produce only one set of plots for
# all iterations (the rest were overwritten due to filename
# collisions), which makes me suspect we don't use them
# if plot:
# plot.standard(data_rf, s_y, set_label)
# data_rf is a list; we need the member named importance
# if we convert it directly to a pd.DataFrame, we lose the col and row
# names for some reason, so convert first to a rpy2 Matrix, then
# copy the col and row names from there
m_vim = data_rf.rx2('importance')
df_vim = pandas2ri.ri2py_dataframe(data_rf.rx2('importance'))
df_vim.columns = m_vim.colnames
# restore original names, which may have been mangled by R
df_vim['row.names'] = df_x.columns.values
# df_vim has one column "row.names" then...
if mode == 'classification':
# one column for each class, then "MeanDecreaseAccuracy"
# and "MeanDecreaseGini"
df_vim = df_vim[['row.names', 'MeanDecreaseAccuracy']]
else:
# "%IncMSE" and "IncNodePurity"
df_vim = df_vim[['row.names', 'IncNodePurity']]
df_vim.columns = ['Feature.Name', 'vim']
return df_vim
import pytest
from rpy2.robjects import vectors
from rpy2.robjects.packages import importr
from rpy2.ipython import html
base = importr('base')
@pytest.mark.parametrize(
'o,func', [(vectors.IntVector([1, 2, 3]), html.html_vector_horizontal),
(vectors.FloatVector([1, 2, 3]), html.html_vector_horizontal),
(vectors.StrVector(['a', 'b'
'c']), html.html_vector_horizontal),
(vectors.FactorVector(['a', 'b'
'c']), html.html_vector_horizontal),
(vectors.ListVector({
'a': 1,
'b': 2
}), html.html_rlist),
(vectors.DataFrame({
'a': 1,
'b': 'z'
}), html.html_rdataframe),
('x <- c(1, 2, 3)', html.html_sourcecode),
(base.c, html.html_ridentifiedobject)])
def test_html_func(o, func):
res = func(o)
assert isinstance(res, str)
def process_refuse_output(refuse_dir, feature_metadata, group_metadata, df_x,
s_y, mode, data_label, ntree, mtry, iss, outer_reps,
inner_reps):
run_info = data_label + "\nNtree: " + str(ntree) + "\nMtry: " + \
str(mtry) + "\nISS: " + str(iss) + "\nOuter: " + str(outer_reps)
data_results = None
parameter_search_stats = None
parameter_search_storage = None
refuse_table = None
vim_reps = None
for filename in os.listdir(refuse_dir):
filepath = os.path.join(refuse_dir, filename)
if filename == 'Data_Results.csv':
data_results = pd.DataFrame.from_csv(filepath)
elif filename == 'Parameter_search_stats.csv':
parameter_search_stats = pd.DataFrame.from_csv(filepath)
elif filename == 'Parameter_search_storage.csv':
parameter_search_storage = pd.DataFrame.from_csv(filepath)
elif filename == 'REFUSE_structured.csv':
refuse_structured = pd.DataFrame.from_csv(filepath)
elif filename == 'REFUSE_Table.csv':
refuse_table = pd.DataFrame.from_csv(filepath)
elif filename == 'vim_reps.csv':
vim_reps = pd.DataFrame.from_csv(filepath)
else:
print 'WARNING: unexpected file ' + filename
# make R load the vis script
script_dir = os.path.dirname(os.path.realpath(__file__))
ro.r.source(os.path.join(script_dir, 'vis.r'))
# set the working directory for R
ro.r.setwd(os.getcwd())
# draw plots
if parameter_search_stats is not None:
ro.r['plot.parameter.search.stats'](parameter_search_stats)
else:
print "WARNING: Missing 'Parameter_search_stats.csv'"
if parameter_search_storage is not None:
ro.r['plot.parameter.search.growth'](parameter_search_storage)
else:
print "WARNING: Missing 'Parameter_search_storage.csv'"
if data_results is not None:
data_results = fst_utils.add_feature_metadata(data_results,
feature_metadata, 'name')
ro.r['plot.compare.methods'](data_results)
plot_importance = ro.r['plot.importance']
plot_importance(vim=ro.FloatVector(data_results['vim']),
labels=ro.StrVector(data_results['name']),
group=ro.StrVector(data_results['segment3']),
main='Auto Segment 3')
plot_importance(vim=ro.FloatVector(data_results['vim']),
labels=ro.StrVector(data_results['name']),
group=ro.StrVector(data_results['segment5']),
main='Auto Segment 5')
else:
print "WARNING: Missing 'Data_Results.csv'"
if refuse_table is not None:
refuse_table = fst_utils.add_feature_metadata(refuse_table,
feature_metadata)
if mode == 'classification':
threshold = 0.6
keepers = refuse_table[refuse_table['p_val'] > threshold]
if len(keepers) > 0:
df_x_best = df_x[keepers.index.values]
my_y = vectors.FactorVector(s_y)
# make summary plots
ro.r['plot.summary.vis'](df_x_best, my_y)
# plot small multiples
ro.r['small.multiples'](df_x_best, my_y)
else:
print "WARNING: 0 features with p_val > " + str(
threshold) + "; skipping summary plots"
palette = ro.StrVector(group_metadata['color'].tolist())
palette.names = ro.StrVector(group_metadata.index.tolist())
# make stability plots
plot_stability = ro.r['plot.stability']
plot_stability(importance=ro.FloatVector(refuse_table['VIM_mean']),
stability=ro.FloatVector(refuse_table['VIM_var']),
labels=ro.StrVector(refuse_table.index.values),
factor_colour=ro.StrVector(refuse_table['group']),
palette=palette,
main="Feature Importance vs Sensitivity",
info=run_info)
plot_stability(importance=ro.FloatVector(refuse_table['p_val']),
stability=ro.FloatVector(refuse_table['VIM_var']),
labels=ro.StrVector(refuse_table.index.values),
factor_colour=ro.StrVector(refuse_table['group']),
palette=palette,
main="Feature p-value vs Sensitivity",
xlabel="Likelihood that a Feature is Relevant",
info=run_info)
# the following portion requires refuse_table and vim_reps
if vim_reps is not None:
# make box plots
plot_importance_box = ro.r['plot.importance.box']
# in addition to vim_reps DF, need to pass version of refuse_table
vim_reps_supplement = refuse_table.loc[vim_reps.columns.values,
['group', 'p_val']]
vim_reps_supplement.sort(columns='p_val',
ascending=False,
inplace=True)
samples = outer_reps * inner_reps
if iss:
samples *= len(df_x.index)
plot_importance_box(
df_vim_reps=vim_reps,
df_supplement=vim_reps_supplement,
palette=palette,
title=data_label + ' Feature Importance Box Plot',
type="box",
info=run_info + "\n# of samples: " + str(samples))
else:
print "WARNING: Missing 'vim_reps.csv'"
else:
print "WARNING: Missing 'REFUSE_Table.csv'"