In [1]:
from images.scripts_helper import RobotStats
import cPickle as pickle
import matplotlib.pyplot as plt
import matplotlib.pylab as pylab
import matplotlib
import numpy as np
import inspect
from numbers import Number
%matplotlib inline
f = open('/Users/beijbom/Dropbox/dummyfigs/rs3.pkld', 'r')
db = pickle.load(f)
def best_per_source(db, metric):
source_ids = np.unique(map(lambda x: x.source_id, db))
dbnew = []
for source_id in source_ids:
db_sub =[(row) for row in db if row.source_id == source_id]
db_valid =[(row) for row in db_sub if isinstance(getattr(row, metric),(int, np.float32, np.float64))]
if len(db_valid) > 0:
dbnew.append(db_valid[np.argmax(map(lambda row: getattr(row, metric), db_valid))])
return dbnew
def remove_small(db, th):
sizes = np.unique(map(lambda x: x.nsamples_org, db))
return [(row) for row in db if row.nsamples_org > th]
def valid_alleviation(db):
return [(row) for row in db if isinstance(row.allevation_level, (int, np.float32, np.float64))]
(dir(db[0]), len(db))
Out[1]:
In [110]:
def train_size_plotter(metric):
yss = [1, 1]
for (itt, field) in enumerate(['target_nbr_samples_final', 'target_nbr_samples_hp']):
ys = []
target_vals = np.unique(map(lambda x: getattr(x, field), db))
for val in target_vals:
db_sub =[(row) for row in db if getattr(row, field) == val]
db_valid =[(row) for row in db_sub if isinstance(getattr(row, metric),(int, np.float32, np.float64))]
ys.append(map(lambda x:(getattr(x, metric)), db_valid))
plt.subplot(1, 2, itt+1)
plt.hist(ys, bins = np.linspace(0, 100, 10), rwidth = 3, histtype = 'barstacked', alpha = .85)
plt.title(metric + " " + field)
plt.legend(target_vals)
yss[itt] = ys
return yss
pylab.rcParams['figure.figsize'] = (12.0, 5.0)
train_size_plotter("allevation_level")
plt.show()
In [22]:
metric = "allevation_level"
target_vals = [2500, 4000, 7500]
bins = np.linspace(-5, 105, 8)
pylab.rcParams['figure.figsize'] = (12.0, 5.0)
colors = ['red', 'green', 'blue']
font = {'size' : 14, 'weight' : 'normal'}
matplotlib.rc('font', **font)
for (itt, val) in enumerate(target_vals):
db_sub =[(row) for row in db if getattr(row, "target_nbr_samples_final") == val]
db_valid =[(row) for row in db_sub if isinstance(getattr(row, metric),(int, np.float32, np.float64))]
ys = map(lambda x:(getattr(x, metric)), db_valid)
vals, a = np.histogram(ys, bins = bins)
vals = np.float32(vals) / sum(vals)
bincenters = 0.5*(bins[1:]+bins[:-1])
plt.plot(bincenters, vals, color = colors[itt], marker = '.', markersize = 10, linewidth = 2)
plt.fill_between(bincenters, vals, color = colors[itt], alpha = .3)
plt.legend(target_vals)
plt.grid(b = 'on', axis='y')
plt.xlabel('Level of alleviation (%)')
plt.ylabel('Pr(x)')
plt.savefig('alleviation_stats_trainsize_breakdown.png', dpi=400)
In [22]:
pylab.rcParams['figure.figsize'] = (12.0, 5.0)
target_vals = [2500, 4000, 7500]
def standard_barplot(db, metric, multiplier):
ys = []
for val in target_vals:
db_sub =[(row) for row in db if row.target_nbr_samples_final == val]
ys.append(map(lambda x:(multiplier*getattr(x, metric)), db_sub))
plt.hist(ys, bins = np.linspace(10,100,10), alpha = .75, histtype='barstacked')
plt.legend(target_vals)
plt.grid(b = 'on', axis='y')
plt.ylabel("Count")
font = {'size' : 14, 'weight' : 'normal'}
matplotlib.rc('font', **font)
plt.subplot(1, 2, 1)
standard_barplot(valid_alleviation(remove_small(db, 5000)), 'funckappa', 100)
plt.text(13, 22, 'A)', fontsize = 24)
plt.xlabel("Acc (Kappa func. group %)")
plt.subplot(1, 2, 2)
standard_barplot(valid_alleviation(remove_small(db, 5000)), 'allevation_level', 1)
plt.xlabel("Level of alleviation (%)")
plt.text(13, 15.8, 'B)', fontsize = 24)
plt.savefig('cnstats_acc_all.png', dpi=400)
In [46]:
def accuracy_scatter(db, xfield, yfield, ymult = 1):
for (i, val) in enumerate(target_vals):
db_sub =[(row) for row in db if row.target_nbr_samples_final == val]
yvals = map(lambda x:getattr(x, yfield)*ymult, db_sub)
xvals = map(lambda x:getattr(x, xfield), db_sub)
plt.scatter(xvals, yvals, color=colors[i])
In [62]:
colors = ['blue', 'green', 'red']
target_vals = [2500, 4000, 7500]
pylab.rcParams['figure.figsize'] = (12.0, 5.0)
font = {'size' : 14, 'weight' : 'normal'}
matplotlib.rc('font', **font)
plt.subplot(1, 2, 1)
accuracy_scatter(db, "nsamples_org", "funckappa", ymult= 100)
plt.legend(target_vals)
plt.xscale('log')
plt.xlim([500, 900000])
plt.text(1000, 97, 'A)', fontsize = 24)
plt.ylim([0, 110])
plt.ylabel('Acc. (Kappa func. group %)')
plt.xlabel('Number of samples')
plt.grid('on')
plt.subplot(1, 2, 2)
accuracy_scatter(db, "nlabels", "train_time", ymult = 1.0/3600)
plt.legend(target_vals)
# plt.xscale('log')
# plt.yscale('log')
plt.xlim([0, 120])
plt.ylim([0, 120])
plt.text(20, 105, 'B)', fontsize = 24)
plt.ylabel('Training time (h)')
plt.xlabel('Size of label-set')
plt.grid('on')
plt.savefig('cnstats_scatter.png', dpi=400)
In [124]:
pylab.rcParams['figure.figsize'] = (12.0, 5.0)
plt.subplot(1, 2, 1)
plt.scatter(map(lambda x:x.nlabels, db), map(lambda x:x.fullkappa*100, db))
plt.ylim([0, 100])
plt.ylabel('Acc. (Kappa full %)')
plt.xlabel('Size of labelset')
plt.grid('on')
plt.subplot(1, 2, 2)
plt.scatter(map(lambda x:x.nlabels, db), map(lambda x:x.funckappa*100, db))
plt.ylim([0, 100])
plt.ylabel('Acc. (Kappa func. group %)')
plt.xlabel('Size of labelset')
plt.grid('on')
plt.savefig('cnstats_labelset_scatter.pdf')
In [133]:
db_sub =[(x) for x in db if isinstance(x.allevation_level, (int, np.float32, np.float64))]
pylab.rcParams['figure.figsize'] = (12.0, 5.0)
plt.scatter(map(lambda x:x.nlabels, db_sub), map(lambda x:x.allevation_level, db_sub))
plt.xlabel('n samples')
plt.grid('on')
In [108]:
ns = map(lambda x:x.nsamples_org, db)
np.argmax(ns)
db[2].source_id
np.argmax([None, 'N.A.', 12])
len(remove_small(db, 1000))
Out[108]: