def test_matrix_difference_classwise():
print(
"\n-- 'matrix_difference' (cosine/cosine_ratio/l2), 'to_centering_beliefs' --"
)
X0 = np.array([
[2, 0, 0],
[2, 0, 2],
[0, 1, 0],
[0, 0, 3],
[0, 0, 3],
[1, 0, 2],
[0, 3, 3],
[0, 0, 0],
[0, 1, 0],
[0, 1, 0],
[9, 9, 9],
[9, 9, 9],
[100, 100, 100],
])
X1 = np.array([
[1, 1, 2],
[2, 1, 2],
[3, 4, 0],
[1, 1, 2],
[2, 1, 1],
[1, 2, 2],
[1, 2, 3],
[0, 0, 0],
[1, 0, 0],
[0, 2, 0],
[9, 9, 9],
[8, 9, 9],
[100, 100, 101],
])
print("X0:\n", X0)
print("X1:\n", X1)
# result = matrix_difference(X0, X1, similarity='accuracy', vector=True)
# print("accuracy:\n", result)
result = matrix_difference(X0, X1, similarity='accuracy', vector=False)
print("accuracy:\n", result)
result = get_class_indices(X0, 0)
print("indices for class 0 in X0:\n", result)
result = matrix_difference_classwise(
X0,
X1,
)
print("accuracy classwise:\n", result)
def calculate_accuracy(H,
X_train,
X_test,
train_ind,
test_ind,
W,
return_output,
s=0.5): # all that is needed to propagate
H0c = to_centering_beliefs(H)
eps_max = eps_convergence_linbp_parameterized(H0c,
W,
method='noecho',
alpha=alpha,
beta=beta,
gamma=gamma,
X=X2)
eps = s * eps_max
F, actualIt, actualPercentageConverged = \
linBP_symmetric_parameterized(X_train, W, H*eps,
method='noecho',
alpha=alpha, beta=beta, gamma=gamma,
numMaxIt=numMax,
convergencePercentage=0.99,
convergenceThreshold=0.99,
debug=2)
n, k = F.shape
for i in range(n):
if i not in test_ind:
F[i] = np.zeros(k)
accuracy_X = matrix_difference(X_test,
F,
ignore_rows=list(train_ind),
similarity='accuracy')
print("Holdout accuracy: {}".format(accuracy_X))
return_output.put(accuracy_X) ## For Parallel
def calculate_accuracy(H, X_train, X_test, train_ind, test_ind, W, s=0.5):
"""Propagates from X_train numMax times, calculates accuracy over X_test
"""
H0c = to_centering_beliefs(H)
eps_max = eps_convergence_linbp_parameterized(
H0c,
W, # TODO: an optimized version could attempt to calculate the spectral radius fewer times and re-use it for multiple splits
method='noecho',
alpha=alpha,
beta=beta,
gamma=gamma,
X=X2)
eps = s * eps_max
F, actualIt, actualPercentageConverged = linBP_symmetric_parameterized(
X_train,
W,
H * eps,
method='noecho',
alpha=alpha,
beta=beta,
gamma=gamma,
numMaxIt=numMax,
convergencePercentage=0.99,
convergenceThreshold=0.99,
debug=2)
n, k = F.shape
for i in range(n):
if i not in test_ind:
F[i] = np.zeros(k)
# TODO For label imbalance, better to use CLASSWISE (macro-averaging) here
accuracy_X = matrix_difference(X_test,
F,
ignore_rows=list(train_ind),
similarity='accuracy')
# print("accuracy now is {}".format(accuracy_X))
return accuracy_X
def run(choice, create_data=False, add_data=False, show_plot=False, create_pdf=False, show_pdf=False, shorten_length=False):
# -- Setup
CHOICE = choice
CREATE_DATA = create_data
ADD_DATA = add_data
SHOW_PDF = show_pdf
SHOW_PLOT = show_plot
CREATE_PDF = create_pdf
STD_FILL = True
csv_filename = 'Fig_End-to-End_accuracy_VaryK_{}.csv'.format(CHOICE)
header = ['currenttime',
'option',
'k',
'f',
'accuracy']
if CREATE_DATA:
save_csv_record(join(data_directory, csv_filename), header, append=False)
# -- Default Graph parameters
rep_SameGraph = 10 # iterations on same graph
initial_h0 = None # initial vector to start finding optimal H
distribution = 'powerlaw'
exponent = -0.3
length = 5
variant = 1
EC = True # Non-backtracking for learning
ymin = 0.3
ymax = 1
xmax = 8
xtick_lab = [2,3,4,5,6,7, 8]
xtick_labels = ['2', '3', '4', '5', '6', '7', '8']
ytick_lab = np.arange(0, 1.1, 0.1)
f_vec = [0.9 * pow(0.1, 1 / 5) ** x for x in range(21)]
k_vec = [3, 4, 5 ]
rep_DifferentGraphs = 10 # iterations on different graphs
err = 0
avoidNeighbors = False
gradient = False
pruneRandom = False
convergencePercentage_W = None
stratified = True
label_vec = ['*'] * 10
clip_on_vec = [False] * 10
draw_std_vec = range(10)
numberOfSplits = 1
linestyle_vec = ['dashed'] + ['solid'] * 10
linewidth_vec = [5, 4, 3, 3] + [3] * 10
marker_vec = [None, None, 'o', 'x', 'o', '^', 'o', 'x', 'o', '^', 'o', 'x', 'o', '^']
markersize_vec = [0, 0, 4, 8] + [6] * 10
facecolor_vec = ["#4C72B0", "#55A868", "#C44E52", "#8172B2", "#CCB974", "#64B5CD"]
# -- Options with propagation variants
if CHOICE == 500: ## 1k nodes
n = 1000
h = 8
d = 25
option_vec = ['opt1', 'opt2', 'opt3']
learning_method_vec = ['GS', 'MHE', 'DHE']
weight_vec = [10] * 3
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 2 + [True]
xmin = 3.
ymin = 0.
ymax = 1.
label_vec = ['GS', 'MCE', 'DCEr']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.03, 0.01, 0.001]
k_vec = [3, 4, 5, 6]
elif CHOICE == 501: ## 10k nodes
n = 10000
h = 8
d = 25
option_vec = ['opt1', 'opt2', 'opt3']
learning_method_vec = ['GT', 'MHE', 'DHE']
weight_vec = [10] * 3
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 2 + [True]
xmin = 2.
ymin = 0.
ymax = 1.
label_vec = ['GT', 'MCE', 'DCEr']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.03, 0.01, 0.001]
k_vec = [2, 3, 4, 5]
elif CHOICE == 502: ## 10k nodes
n = 10000
h = 8
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
ymin = 0.6
ymax = 1.
label_vec = ['GT', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.01]
k_vec = [2, 3, 4, 5, 6, 7, 8]
# option_vec = ['opt1', 'opt2', 'opt3', 'opt4']
# learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE']
# k_vec = [2, 3, 4, 5]
elif CHOICE == 503: ## 10k nodes
n = 10000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
ymin = 0.3
ymax = 0.9
label_vec = ['GT', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.01]
k_vec = [2, 3, 4, 5, 6, 7, 8]
# k_vec = [6, 7, 8]
clip_on_vec = [True] * 10
# option_vec = ['opt1', 'opt2', 'opt3', 'opt4']
# learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE']
# k_vec = [2, 3, 4, 5]
elif CHOICE == 504: ## 10k nodes
n = 10000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
xmax = 7
ymin = 0.2
ymax = 0.9
label_vec = ['GT', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.01]
# k_vec = [2, 3, 4, 5, 6, 7, 8]
k_vec = [7]
clip_on_vec = [True] * 10
elif CHOICE == 505: ## 10k nodes with f = 0.005
n = 10000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
xmax = 7
ymin = 0.2
ymax = 0.9
label_vec = ['GT', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.005]
k_vec = [2, 3, 4, 5, 6, 7]
# k_vec = [7]
clip_on_vec = [True] * 10
# elif CHOICE == 506: ## 10k nodes with f = 0.005
# n = 10000
# h = 3
# d = 25
# option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5']
# learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE']
# weight_vec = [10] * 10
# alpha_vec = [0] * 10
# beta_vec = [0] * 10
# gamma_vec = [0] * 10
# s_vec = [0.5] * 10
# numMaxIt_vec = [10] * 10
# randomize_vec = [False] * 4 + [True] + [False]
# xmin = 2
# xmax = 7
# ymin = 0.2
# ymax = 0.9
# label_vec = ['GT', 'LCE', 'MCE', 'DCE', 'DCEr']
# facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
# f_vec = [0.005]
# k_vec = [2,3,4,5,6,7]
# # k_vec = [7]
# clip_on_vec = [True] * 10
elif CHOICE == 506: ## 10k nodes
n = 10000
h = 8
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5']
learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE', 'DHE']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
xmax = 7
ymin = 0.2
ymax = 0.9
label_vec = ['GT', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.005]
k_vec = [2, 3, 4, 5, 6, 7, 8]
# k_vec = [5]
clip_on_vec = [True] * 10
rep_SameGraph = 1 # iterations on same graph
rep_DifferentGraphs = 1 # iterations on same graph
elif CHOICE == 507: ## 10k nodes with gradient and PruneRandom
n = 10000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['GS', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
ymin = 0.1
ymax = 0.9
label_vec = ['GS', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.01]
k_vec = [2, 3, 4, 5, 6, 7, 8]
# k_vec = [6, 7, 8]
clip_on_vec = [True] * 10
# option_vec = ['opt1', 'opt2', 'opt3', 'opt4']
# learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE']
# k_vec = [2, 3, 4, 5]
gradient = True
pruneRandom = True
elif CHOICE == 508: ## 10k nodes with gradient and PruneRandom
n = 1000
h = 3
d = 10
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['GS', 'LHE', 'MHE', 'DHE', 'DHE', 'Holdout']
weight_vec = [10] * 10
alpha_vec = [0] * 10
beta_vec = [0] * 10
gamma_vec = [0] * 10
s_vec = [0.5] * 10
numMaxIt_vec = [10] * 10
randomize_vec = [False] * 4 + [True] + [False]
xmin = 2
ymin = 0.1
ymax = 0.9
label_vec = ['GS', 'LCE', 'MCE', 'DCE', 'DCEr', 'Holdout']
facecolor_vec = ['black'] + ["#55A868", "#4C72B0", "#8172B2", "#C44E52", "#CCB974"] * 3
f_vec = [0.01]
k_vec = [2, 3, 4, 5, 6, 7, 8]
# k_vec = [6, 7, 8]
clip_on_vec = [True] * 10
# option_vec = ['opt1', 'opt2', 'opt3', 'opt4']
# learning_method_vec = ['GT', 'LHE', 'MHE', 'DHE']
# k_vec = [2, 3, 4, 5]
gradient = True
pruneRandom = True
rep_DifferentGraphs = 1
rep_SameGraph = 1
else:
raise Warning("Incorrect choice!")
RANDOMSEED = None # For repeatability
random.seed(RANDOMSEED) # seeds some other python random generator
np.random.seed(seed=RANDOMSEED) # seeds the actually used numpy random generator; both are used and thus needed
# print("CHOICE: {}".format(CHOICE))
# -- Create data
if CREATE_DATA or ADD_DATA:
for i in range(rep_DifferentGraphs): # create several graphs with same parameters
# print("\ni: {}".format(i))
for k in k_vec:
# print("\nk: {}".format(k))
H0 = create_parameterized_H(k, h, symmetric=True)
H0c = to_centering_beliefs(H0)
a = [1.] * k
alpha0 = np.array(a)
alpha0 = alpha0 / np.sum(alpha0)
W, Xd = planted_distribution_model_H(n, alpha=alpha0, H=H0, d_out=d,
distribution=distribution,
exponent=exponent,
directed=False,
debug=False)
X0 = from_dictionary_beliefs(Xd)
for j in range(rep_SameGraph): # repeat several times for same graph
# print("j: {}".format(j))
ind = None
for f in f_vec: # Remove fraction (1-f) of rows from X0 (notice that different from first implementation)
X1, ind = replace_fraction_of_rows(X0, 1-f, avoidNeighbors=avoidNeighbors, W=W, ind_prior=ind, stratified=stratified)
X2 = introduce_errors(X1, ind, err)
for option_index, (learning_method, alpha, beta, gamma, s, numMaxIt, weights, randomize) in \
enumerate(zip(learning_method_vec, alpha_vec, beta_vec, gamma_vec, s_vec, numMaxIt_vec, weight_vec, randomize_vec)):
# -- Learning
if learning_method == 'GT':
H2c = H0c
elif learning_method == 'Holdout':
H2 = estimateH_baseline_serial(X2, ind, W, numMax=numMaxIt,
# ignore_rows=ind,
numberOfSplits=numberOfSplits,
# method=learning_method, variant=1, distance=length,
EC=EC,
alpha=alpha, beta=beta, gamma=gamma)
H2c = to_centering_beliefs(H2)
elif learning_method != 'DHE':
H2 = estimateH(X2, W, method=learning_method, variant=1, distance=length, EC=EC, weights=weights, randomize=randomize)
H2c = to_centering_beliefs(H2)
else:
H2 = estimateH(X2, W, method=learning_method, variant=1, distance=length, EC=EC, weights=weights, randomize=randomize, gradient=gradient, randomrestarts=pruneRandom)
H2c = to_centering_beliefs(H2)
# -- Propagation
X2c = to_centering_beliefs(X2, ignoreZeroRows=True) # try without
eps_max = eps_convergence_linbp_parameterized(H2c, W,
method='noecho',
alpha=alpha, beta=beta, gamma=gamma,
X=X2)
eps = s * eps_max
try:
F, actualIt, actualPercentageConverged = \
linBP_symmetric_parameterized(X2, W, H2c * eps,
method='noecho',
alpha=alpha, beta=beta, gamma=gamma,
numMaxIt=numMaxIt,
convergencePercentage=convergencePercentage_W,
debug=2)
except ValueError as e:
print (
"ERROR: {} with {}: d={}, h={}".format(e, learning_method, d, h))
else:
accuracy_X = matrix_difference(X0, F, ignore_rows=ind)
tuple = [str(datetime.datetime.now())]
text = [option_vec[option_index],
k,
f,
accuracy_X]
# text = ['' if v is None else v for v in text] # TODO: test with vocabularies
# text = np.asarray(text) # without np, entries get ugly format
tuple.extend(text)
# print("option: {}, f: {}, actualIt: {}, accuracy: {}".format(option_vec[option_index], f, actualIt, accuracy_X))
save_csv_record(join(data_directory, csv_filename), tuple)
# -- Read, aggregate, and pivot data for all options
df1 = pd.read_csv(join(data_directory, csv_filename))
# print("\n-- df1: (length {}):\n{}".format(len(df1.index), df1.head(15)))
# -- Aggregate repetitions
df2 = df1.groupby(['option', 'k', 'f']).agg \
({'accuracy': [np.mean, np.std, np.size, np.median], # Multiple Aggregates
})
df2.columns = ['_'.join(col).strip() for col in df2.columns.values] # flatten the column hierarchy
df2.reset_index(inplace=True) # remove the index hierarchy
df2.rename(columns={'accuracy_size': 'count'}, inplace=True)
# print("\n-- df2 (length {}):\n{}".format(len(df2.index), df2.head(15)))
# -- Pivot table
df3 = pd.pivot_table(df2, index=['f', 'k'], columns=['option'], values=[ 'accuracy_mean', 'accuracy_std'] ) # Pivot
# print("\n-- df3 (length {}):\n{}".format(len(df3.index), df3.head(30)))
df3.columns = ['_'.join(col).strip() for col in df3.columns.values] # flatten the column hierarchy
df3.reset_index(inplace=True) # remove the index hierarchy
# df2.rename(columns={'time_size': 'count'}, inplace=True)
# print("\n-- df3 (length {}):\n{}".format(len(df3.index), df3.head(100)))
# X_f = k_vec
X_f = df3['k'].values # read k from values instead
Y_hash = defaultdict(dict)
Y_hash_std = defaultdict(dict)
for f in f_vec:
for option in option_vec:
Y_hash[f][option] = list()
Y_hash_std[f][option] = list()
for f in f_vec:
for option in option_vec:
Y_hash[f][option] = df3.loc[df3['f'] == f]['accuracy_mean_{}'.format(option)].values
Y_hash_std[f][option] = df3.loc[df3['f'] == f]['accuracy_std_{}'.format(option)].values
if CREATE_PDF or SHOW_PLOT or SHOW_PDF:
# -- Setup figure
fig_filename = 'Fig_End-to-End_accuracy_varyK_{}.pdf'.format(CHOICE)
mpl.rc('font', **{'family': 'sans-serif', 'sans-serif': [u'Arial', u'Liberation Sans']})
mpl.rcParams['axes.labelsize'] = 20
mpl.rcParams['xtick.labelsize'] = 16
mpl.rcParams['ytick.labelsize'] = 16
mpl.rcParams['legend.fontsize'] = 14
mpl.rcParams['grid.color'] = '777777' # grid color
mpl.rcParams['xtick.major.pad'] = 2 # padding of tick labels: default = 4
mpl.rcParams['ytick.major.pad'] = 1 # padding of tick labels: default = 4
mpl.rcParams['xtick.direction'] = 'out' # default: 'in'
mpl.rcParams['ytick.direction'] = 'out' # default: 'in'
mpl.rcParams['axes.titlesize'] = 16
mpl.rcParams['figure.figsize'] = [4, 4]
fig = figure()
ax = fig.add_axes([0.13, 0.17, 0.8, 0.8])
opt_f_vecs = [(option, f) for option in option_vec for f in f_vec]
for ((option, f), color, linewidth, clip_on, linestyle, marker, markersize) in \
zip(opt_f_vecs, facecolor_vec, linewidth_vec, clip_on_vec, linestyle_vec, marker_vec, markersize_vec):
# label = learning_method_vec[option_vec.index(option)]
label = label_vec[option_vec.index(option)]
# label = label + " " + str(f)
if STD_FILL:
# print((X_f))
# print(Y_hash[f][option])
ax.fill_between(X_f, Y_hash[f][option] + Y_hash_std[f][option], Y_hash[f][option] - Y_hash_std[f][option],
facecolor=color, alpha=0.2, edgecolor=None, linewidth=0)
ax.plot(X_f, Y_hash[f][option] + Y_hash_std[f][option], linewidth=0.5, color='0.8', linestyle='solid')
ax.plot(X_f, Y_hash[f][option] - Y_hash_std[f][option], linewidth=0.5, color='0.8', linestyle='solid')
ax.plot(X_f, Y_hash[f][option], linewidth=linewidth, color=color, linestyle=linestyle, label=label, zorder=4, marker=marker,
markersize=markersize, markeredgewidth=1, markeredgecolor='black', clip_on=clip_on)
if CHOICE==507:
Y_f = [1/float(i) for i in X_f]
ax.plot(X_f, Y_f, linewidth=2, color='black', linestyle='dashed',
label='Random', zorder=4, marker='x',
markersize=8, markeredgewidth=1, markeredgecolor='black', clip_on=clip_on)
# -- Title and legend
if distribution == 'uniform':
distribution_label = ',$uniform'
else:
distribution_label = '$'
if n < 1000:
n_label='{}'.format(n)
else:
n_label = '{}k'.format(int(n / 1000))
title(r'$\!\!\!n\!=\!{}, d\!=\!{}, h\!=\!{}, f\!=\!{}{}'.format(n_label, d, h, f, distribution_label))
handles, label_vec = ax.get_legend_handles_labels()
legend = plt.legend(handles, label_vec,
loc='upper right', # 'upper right'
handlelength=2,
labelspacing=0, # distance between label entries
handletextpad=0.3, # distance between label and the line representation
borderaxespad=0.2, # distance between legend and the outer axes
borderpad=0.3, # padding inside legend box
numpoints=1, # put the marker only once
)
# # legend.set_zorder(1)
frame = legend.get_frame()
frame.set_linewidth(0.0)
frame.set_alpha(0.9) # 0.8
# -- Figure settings and save
plt.xticks(xtick_lab, xtick_labels)
plt.yticks(ytick_lab, ytick_lab)
ax.yaxis.set_major_formatter(mpl.ticker.FormatStrFormatter('%.1f'))
# Only show ticks on the left and bottom spines
ax.yaxis.set_ticks_position('left')
ax.xaxis.set_ticks_position('bottom')
grid(b=True, which='major', axis='both', alpha=0.2, linestyle='solid', linewidth=0.5) # linestyle='dashed', which='minor', axis='y',
grid(b=True, which='minor', axis='both', alpha=0.2, linestyle='solid', linewidth=0.5) # linestyle='dashed', which='minor', axis='y',
xlabel(r'Number of Classes $(k)$', labelpad=0) # labelpad=0
ylabel(r'Accuracy', labelpad=0)
xlim(xmin, xmax)
ylim(ymin, ymax)
if CREATE_PDF:
savefig(join(figure_directory, fig_filename), format='pdf',
dpi=None,
edgecolor='w',
orientation='portrait',
transparent=False,
bbox_inches='tight',
pad_inches=0.05,
frameon=None)
if SHOW_PLOT:
plt.show()
if SHOW_PDF:
showfig(join(figure_directory, fig_filename))
def _f_worker_(X0, W, f, f_index):
RANDOMSEED = None # For repeatability
random.seed(RANDOMSEED) # seeds some other python random generator
np.random.seed(
seed=RANDOMSEED
) # seeds the actually used numpy random generator; both are used and thus needed
X1, ind = replace_fraction_of_rows(X0,
1 - f,
avoidNeighbors=avoidNeighbors,
W=W,
stratified=stratified)
X2 = introduce_errors(X1, ind, err)
for option_index, (label, select_lambda, learning_method, alpha, beta, gamma, s, numMaxIt, weights, randomize) in \
enumerate(zip(labels, select_lambda_vec, learning_method_vec, alpha_vec, beta_vec, gamma_vec, s_vec, numMaxIt_vec, weight_vec, randomize_vec)):
learn_time = -1
# -- Learning
if learning_method == 'GT':
H2c = H0c
elif learning_method == 'Heuristic':
# print('Heuristic')
H2c = H_heuristic
elif learning_method == 'Holdout':
# print('Holdout')
H2 = estimateH_baseline_serial(
X2,
ind,
W,
numMax=numMaxIt,
# ignore_rows=ind,
numberOfSplits=numberOfSplits,
# method=learning_method, variant=1,
# distance=length,
EC=EC,
alpha=alpha,
beta=beta,
gamma=gamma,
doubly_stochastic=doubly_stochastic)
H2c = to_centering_beliefs(H2)
else:
if "DCEr" in learning_method:
learning_method = "DCEr"
elif "DCE" in learning_method:
learning_method = "DCE"
# -- choose optimal lambda: allows to specify different lambda for different f
# print("option: ", option_index)
if select_lambda == True:
weight = lambda_vec[f_index]
# print("weight : ", weight)
else:
weight = weights
# -- learn H
learn_start = time.time()
H2 = estimateH(X2,
W,
method=learning_method,
variant=1,
distance=length,
EC=EC,
weights=weight,
randomrestarts=num_restarts,
randomize=randomize,
constraints=constraints,
gradient=gradient,
doubly_stochastic=doubly_stochastic)
learn_time = time.time() - learn_start
H2c = to_centering_beliefs(H2)
# if learning_method not in ['GT', 'GS']:
# print(FILENAMEZ, f, learning_method)
# print(H2c)
# -- Propagation
prop_start = time.time()
# X2c = to_centering_beliefs(X2, ignoreZeroRows=True) # try without
eps_max = eps_convergence_linbp_parameterized(H2c,
W,
method='noecho',
alpha=alpha,
beta=beta,
gamma=gamma,
X=X2)
eps = s * eps_max
# print("Max eps: {}, eps: {}".format(eps_max, eps))
# eps = 1
try:
F, actualIt, actualPercentageConverged = \
linBP_symmetric_parameterized(X2, W, H2c * eps,
method='noecho',
alpha=alpha, beta=beta, gamma=gamma,
numMaxIt=numMaxIt,
convergencePercentage=convergencePercentage_W,
debug=2)
prop_time = time.time() - prop_start
if Macro_Accuracy:
accuracy_X = matrix_difference_classwise(X0,
F,
ignore_rows=ind)
precision = matrix_difference_classwise(X0,
F,
similarity='precision',
ignore_rows=ind)
recall = matrix_difference_classwise(X0,
F,
similarity='recall',
ignore_rows=ind)
else:
accuracy_X = matrix_difference(X0, F, ignore_rows=ind)
precision = matrix_difference(X0,
F,
similarity='precision',
ignore_rows=ind)
recall = matrix_difference(X0,
F,
similarity='recall',
ignore_rows=ind)
result = [str(datetime.datetime.now())]
text = [
label, f, accuracy_X, precision, recall, learn_time, prop_time
]
result.extend(text)
# print("method: {}, f: {}, actualIt: {}, accuracy: {}, precision:{}, recall: {}, learning time: {}, propagation time: {}".format(label, f, actualIt, accuracy_X, precision, recall, learn_time, prop_time))
save_csv_record(join(data_directory, csv_filename), result)
except ValueError as e:
print("ERROR: {} with {}: d={}, h={}".format(
e, learning_method, d, h))
raise e
return 'success'
def test_matrix_difference():
print(
"\n-- 'matrix_difference' (cosine/cosine_ratio/l2), 'to_centering_beliefs' --"
)
X0 = np.array([
[2, 0, 0],
[2, 0, 2],
[0, 1, 0],
[0, 0, 3],
[0, 0, 3],
[1, 0, 2],
[0, 3, 3],
[0, 0, 0],
[0, 1, 0],
[0, 1, 0],
[9, 9, 9],
[9, 9, 9],
[100, 100, 100],
])
X1 = np.array([
[1, 1, 2],
[2, 1, 2],
[3, 4, 0],
[1, 1, 2],
[2, 1, 1],
[1, 2, 2],
[1, 2, 3],
[0, 0, 0],
[1, 0, 0],
[0, 2, 0],
[9, 9, 9],
[8, 9, 9],
[100, 100, 101],
])
print("X0:\n", X0)
print("X1:\n", X1)
result = matrix_difference(X0, X1, similarity='cosine', vector=True)
print("cosine:\n", result)
result = matrix_difference(X0, X1, similarity='cosine_ratio', vector=True)
print("cosine_ratio:\n", result)
result = matrix_difference(X0, X1, similarity='l2', vector=True)
print("l2:\n", result)
X0 = np.array([[1., 0., 0.], [0.30804075, 0.56206462, 0.12989463],
[0.32434628, 0.33782686, 0.33782686],
[0.30804075, 0.12989463, 0.56206462],
[0.14009173, 0.71981654, 0.14009173],
[0.32273419, 0.21860539, 0.45866042],
[0.33804084, 0.32391832, 0.33804084],
[0.45866042, 0.21860539, 0.32273419]])
X1 = np.array([[1., 0., 0.], [0.22382029, 0.45296374, 0.32321597],
[0.32434628, 0.33782686, 0.33782686],
[0.22382029, 0.32321597, 0.45296374],
[0.2466463, 0.5067074, 0.2466463],
[0.32273419, 0.21860539, 0.45866042],
[0.33804084, 0.32391832, 0.33804084],
[0.45866042, 0.21860539, 0.32273419]])
print("\nX0:\n", X0)
print("X1:\n", X1)
result = matrix_difference(X0, X1, similarity='cosine_ratio', vector=True)
print("cosine_ratio:\n", result)
# X0z = row_normalize_matrix(X0, norm='zscores')
# X1z = row_normalize_matrix(X1, norm='zscores')
X0z = to_centering_beliefs(X0)
X1z = to_centering_beliefs(X1)
print("\nX0z:\n", X0z)
print("X1z:\n", X1z)
result = matrix_difference(X0z,
X1z,
similarity='cosine_ratio',
vector=True)
print("cosine_ratio zscores:\n", result)
# actualPercentageConverged = matrix_convergence_percentage(X0z, X1z, threshold=convergenceCosineSimilarity)
X0 = np.array([1, 0, 0])
X1 = np.array([1, 1, 0])
print("\nX0:\n", X0)
print("X1:\n", X1)
result = matrix_difference(X0, X1, similarity='cosine_ratio', vector=True)
print("cosine_ratio zscores:\n", result)
X0 = np.array([-30, -15, 45])
X1 = np.array([-15, -30, 45])
print("\nX0:\n", X0)
print("X1:\n", X1)
result = matrix_difference(X0, X1, similarity='cosine_ratio', vector=True)
print("cosine_ratio zscores:\n", result)
def test_matrix_difference_with_accuracy_etc():
print(
"\n-- 'matrix_difference' (precision/recall/accuracy/cosine), 'max_binary_matrix' --"
)
X_true = np.array([[2, 0, 0], [2, 0, 2], [0, 1, 0], [0, 0, 3], [0, 0, 3],
[1, 0, 2], [0, 3, 3], [0, 3, 3], [0, 0, 3], [0, 0, 3]])
X_pred = np.array([
[1, 1, 2],
[2, 1, 2],
[3, 4, 0],
[1, 1, 2],
[2, 1, 1],
[1, 2, 2],
[1, 2, 3],
[1, 2.99, 3],
[1, 2.8, 3],
[1, 2.99, 3],
])
X_true_b = max_binary_matrix(X_true)
X_pred_b = max_binary_matrix(X_pred)
X_pred_b1 = max_binary_matrix(X_pred, threshold=0.1)
print("X_true:\n", X_true)
print("X_pred:\n", X_pred)
print("\nX_true binary:\n", X_true_b)
print("X_pred binary:\n", X_pred_b)
print("X_pred binary with threshold 0.1:\n", X_pred_b1)
ind = list([])
# ind = list([0, 1])
# ind = list([1, 2, 3, 4, 5])
# ind = list([0, 2, 3, 4, 5, 6])
print(
"\nPrecision:\n",
matrix_difference(X_true,
X_pred_b1,
ind,
vector=True,
similarity='precision'))
# print("*** type:", type (matrix_difference(X_true, X_pred_b1, ind, vector=True, similarity='precision')))
print(
"Recall:\n",
matrix_difference(X_true,
X_pred_b1,
ind,
vector=True,
similarity='recall'))
print(
"Accuracy:\n",
matrix_difference(X_true,
X_pred_b1,
ind,
vector=True,
similarity='accuracy'))
cosine_list = matrix_difference(X_true,
X_pred,
ind,
vector=True,
similarity='cosine')
print("Cosine:\n", cosine_list)
print("Cosine sorted:\n", sorted(cosine_list, reverse=True))
print("\nPrecision:\n",
matrix_difference(X_true, X_pred, ind, similarity='precision'))
print("Recall:\n",
matrix_difference(X_true, X_pred, ind, similarity='recall'))
print("Accuracy:\n", matrix_difference(X_true, X_pred, ind))
print("Cosine:\n",
matrix_difference(X_true, X_pred, ind, similarity='cosine'))
def test_matrix_difference_with_cosine_simililarity():
print("\n-- 'matrix_difference' (cosine), 'row_normalize_matrix' --")
print("k=3")
v1 = np.array([1, 0, 0])
v2 = np.array([0, 1, 0])
v3 = np.array([1, 1, 0])
print("Cosine with original:\n ", \
matrix_difference(v1,
v1, similarity='cosine'))
print("Cosine with original zscore:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v1, norm='zscores'), similarity='cosine'))
print("Cosine with zscore :\n ", \
matrix_difference(v1,
row_normalize_matrix(v1, norm='zscores'), similarity='cosine'))
print("Cosine with normal:\n ", \
matrix_difference(v1,
v2, similarity='cosine'))
print("Cosine with normal after both zscore:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v2, norm='zscores'), similarity='cosine'))
print("! Notice that average guessing leads to expectation of 0!")
print("Cosine v1, v3:\n ", \
matrix_difference(v1,
v3, similarity='cosine'))
print("Cosine v1, v3 after zscore:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v3, norm='zscores'), similarity='cosine'))
print("\nk=5")
v1 = np.array([1, 0, 0, 0, 0])
v2 = np.array([0, 1, 0, 0, 0])
v3 = np.array([1, 1, 0, 0, 0])
v4 = np.array([0, 0, 0, 0, 0])
print("Cosine with normal:\n ", \
matrix_difference(v1,
v2, similarity='cosine'))
print("Cosine with normal after both zscore:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v2, norm='zscores'), similarity='cosine'))
print("! Notice that average guessing leads to expectation of 0!")
print("Cosine v1, v3:\n ", \
matrix_difference(v1,
v3, similarity='cosine'))
print("Cosine v1, v3 after zscore:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v3, norm='zscores'), similarity='cosine'))
print("Average Cos similarity partly zscore:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v3, norm='zscores'), similarity='cosine'))
print("Cosine with 0-vector:\n ", \
matrix_difference(row_normalize_matrix(v1, norm='zscores'),
row_normalize_matrix(v4, norm='zscores'), similarity='cosine'))
print()
X = np.array([[1, 0, 0, 0, 0], [1, 0, 0, 0, 0], [1, 0, 0, 0, 0],
[1, 0, 0, 0, 0], [1, 1, 0, 0, 0]])
Y = np.array([[1, 0, 0, 0, 0], [1, 1, 0, 0, 0], [0, 0, 0, 0, 0],
[0, 1, 0, 0, 0], [1, 1.1, 0, 0, 0]])
print("X\n", X)
print("Y\n", Y)
Xs = row_normalize_matrix(X, norm='zscores')
Ys = row_normalize_matrix(Y, norm='zscores')
print("Xs\n", Xs)
print("Ys\n", Ys)
print("\nCosine original:\n ", \
matrix_difference(X,
Y, vector=True, similarity='cosine'))
print("Cosine zscore:\n ", \
matrix_difference(Xs,
Ys, vector=True, similarity='cosine'))
print("Average cosine zscore:\n ", \
matrix_difference(X,
Y, similarity='cosine'))
def run(choice, create_data=False, add_data=False, show_plot=False, create_pdf=False, show_pdf=False, shorten_length=False,
show_arrows=False):
# -- Setup
CHOICE = choice
CREATE_DATA = create_data
ADD_DATA = add_data
SHOW_PLOT = show_plot
SHOW_PDF = show_pdf
CREATE_PDF = create_pdf
SHOW_STD = True ## FALSE for just scatter plot points
SHOW_ARROWS = show_arrows
# -- Default Graph parameters
rep_SameGraph = 1 # iterations on same graph
distribution = 'powerlaw'
exponent = -0.3
length = 5
variant = 1
EC = False
numberOfSplits = 1
scaling_vec = [None]*10
ymin = 0.3
ymax = 1
xmin = 1e-3
xmax = 1e3
xtick_lab = [1e-3, 0.01, 0.1, 1, 10, 100, 1000]
xtick_labels = [r'$10^{-3}$', r'$10^{-2}$', r'$10^{-1}$', r'$1$', r'$10$', r'$10^{2}$', r'$10^{3}$']
ytick_lab = np.arange(0, 1.1, 0.1)
k = 3
a = 1
rep_DifferentGraphs = 1 # iterations on different graphs
err = 0
avoidNeighbors = False
convergencePercentage_W = 0.99
facecolor_vec = ["#4C72B0", "#55A868", "#8172B2", "#C44E52", "#CCB974", "#64B5CD"]
label_vec = ['MCE', 'LCE', 'DCE', 'Holdout']
linewidth_vec = [4, 3, 1, 2, 2, 1]
# clip_ons = [True, True, True, True, True, True]
FILEZNAME = 'Fig_timing_accuracy_learning'
marker_vec = ['s', '^', 'v', 'o', 'x', '+', 'None'] #'^'
length_vec = [5]
stratified = True
f = 0.01
numMaxIt_vec = [10]*7
alpha_vec = [0] * 7
beta_vec = [0] * 7 # TODO: LinBP does not use beta. Also SSLH uses alpha, but not beta for W^row! Now fixed
gamma_vec = [0] * 7
s_vec = [0.5] * 7
# -- Main Options
if CHOICE == 1: # Main graph
n = 1000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['Holdout'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'Holdout', 'GS']
randomize_vec = [False]*3 + [True] + [None]*2
scaling_vec = [None]*2 + [10, 100] + [None]*2
splits_vec = [1, 2, 4, 8]
elif CHOICE == 2:
n = 1000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'GS']
randomize_vec = [False]*3 + [True] + [None]
scaling_vec = [None]*2 + [10, 100] + [None]
elif CHOICE == 3:
n = 1000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'GS']
randomize_vec = [False]*3 + [True] + [None]
scaling_vec = [None]*2 + [10, 100] + [None]
f = 0.02
elif CHOICE == 4: # TODO: Overnight Wolfgang
n = 1000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['Holdout'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'Holdout', 'GS']
randomize_vec = [False]*3 + [True] + [None]*2
scaling_vec = [None]*2 + [10, 100] + [None]*2
splits_vec = [1, 2, 4, 8, 16]
elif CHOICE == 5: # Toy graph with 100 nodes
n = 100
h = 3
d = 8
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['Holdout'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'Holdout', 'GS']
randomize_vec = [False]*3 + [True] + [None]*2
scaling_vec = [None]*2 + [10, 100] + [None]*2
splits_vec = [1, 2, 4, 8]
f=0.05
elif CHOICE == 6: # To be run by Prakhar on Cluster
n = 10000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['Holdout'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCEr', 'Holdout', 'GS']
randomize_vec = [False]*3 + [True] + [None]*2
scaling_vec = [None]*2 + [10, 100] + [None]*2
splits_vec = [1, 2, 4, 8]
f=0.003
xmin = 1e-2
# ymax = 0.9
ymin = 0.2
ymax = 0.9
xmin = 1e-2
xmax = 1e3
elif CHOICE == 7:
n = 1000
h = 3
d = 25
option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['Holdout'] + ['GS']
label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'Holdout', 'GS']
randomize_vec = [False]*3 + [True] + [None]*2
scaling_vec = [None]*2 + [10, 100] + [None]*2
splits_vec = [1, 2, 4, 8, 16]
f=0.009
# elif CHOICE == 8: # not working well
# n = 1000
# h = 3
# d = 25
# option_vec = ['opt1', 'opt2', 'opt3', 'opt4', 'opt5', 'opt6']
# learning_method_vec = ['MHE'] + ['LHE'] + ['DHE'] + ['DHE'] + ['Holdout'] + ['GS']
# label_vec = ['MCE', 'LCE', 'DCE', 'DCE r', 'Holdout', 'GS']
# randomize_vec = [False]*3 + [True] + [None]*2
# scaling_vec = [None]*2 + [10, 100] + [None]*2
# splits_vec = [1, 2, 4, 8, 16]
# f=0.005
else:
raise Warning("Incorrect choice!")
csv_filename = '{}_{}.csv'.format(FILEZNAME, CHOICE)
header = ['currenttime',
'option',
'lensplit',
'f',
'accuracy',
'timetaken']
if CREATE_DATA:
save_csv_record(join(data_directory, csv_filename), header, append=False)
alpha0 = np.array([a, 1., 1.])
alpha0 = alpha0 / np.sum(alpha0)
H0 = create_parameterized_H(k, h, symmetric=True)
H0c = to_centering_beliefs(H0)
RANDOMSEED = None # For repeatability
random.seed(RANDOMSEED) # seeds some other python random generator
np.random.seed(seed=RANDOMSEED) # seeds the actually used numpy random generator; both are used and thus needed
# print("CHOICE: {}".format(CHOICE))
# -- Create data
if CREATE_DATA or ADD_DATA:
for i in range(rep_DifferentGraphs): # create several graphs with same parameters
# print("\ni: {}".format(i))
W, Xd = planted_distribution_model_H(n, alpha=alpha0, H=H0, d_out=d,
distribution=distribution,
exponent=exponent,
directed=False,
debug=False)
X0 = from_dictionary_beliefs(Xd)
for j in range(rep_SameGraph): # repeat several times for same graph
# print("j: {}".format(j))
ind = None
X1, ind = replace_fraction_of_rows(X0, 1-f, avoidNeighbors=avoidNeighbors, W=W, ind_prior=ind, stratified = stratified) # TODO: stratified sampling option = True
X2 = introduce_errors(X1, ind, err)
for option_index, (learning_method, alpha, beta, gamma, s, numMaxIt, weight, randomize, option) in \
enumerate(zip(learning_method_vec, alpha_vec, beta_vec, gamma_vec, s_vec, numMaxIt_vec, scaling_vec, randomize_vec, option_vec)):
# weight = np.array([np.power(scaling, i) for i in range(5)]) # TODO: now enough to specify weight as a scalar!
H_est_dict = {}
timeTaken_dict = {}
# -- Learning
if learning_method == 'Holdout' :
for numberOfSplits in splits_vec:
prev_time = time.time()
H_est_dict[numberOfSplits] = estimateH_baseline_serial(X2, ind, W, numMax=numMaxIt,
# ignore_rows=ind,
numberOfSplits=numberOfSplits,
# method=learning_method, variant=1, distance=length,
EC=EC,
weights=weight, alpha=alpha, beta=beta, gamma=gamma)
timeTaken = time.time() - prev_time
timeTaken_dict[numberOfSplits] = timeTaken
elif learning_method in ['LHE', 'MHE', 'DHE']: # TODO: no smartInit, just randomization as option
for length in length_vec:
prev_time = time.time()
H_est_dict[length] = estimateH(X2, W, method=learning_method, variant=1, randomize=randomize, distance=length, EC=EC, weights=weight)
timeTaken = time.time() - prev_time
timeTaken_dict[length] = timeTaken
elif learning_method == 'GS':
H_est_dict['GS'] = H0
for key in H_est_dict:
H_est = H_est_dict[key]
H2c = to_centering_beliefs(H_est)
# print("H_estimated by {} is \n".format(learning_method), H_est)
# print("H0 is \n", H0)
# print("randomize was: ", randomize)
# Propagation
X2c = to_centering_beliefs(X2, ignoreZeroRows=True) # try without
eps_max = eps_convergence_linbp_parameterized(H2c, W,
method='noecho',
alpha=alpha, beta=beta, gamma=gamma,
X=X2)
eps = s * eps_max
# print("Max Eps ", eps_max)
try:
F, actualIt, actualPercentageConverged = \
linBP_symmetric_parameterized(X2, W, H2c * eps,
method='noecho',
alpha=alpha, beta=beta, gamma=gamma,
numMaxIt=numMaxIt,
convergencePercentage=convergencePercentage_W,
convergenceThreshold=0.99,
debug=2)
except ValueError as e:
print(
"ERROR: {} with {}: d={}, h={}".format(e, learning_method, d, h))
else:
accuracy_X = matrix_difference(X0, F, ignore_rows=ind)
tuple = [str(datetime.datetime.now())]
if learning_method == 'Holdout':
text = [option,"split{}".format(key), f, accuracy_X, timeTaken_dict[key]]
elif learning_method in ['MHE', 'DHE', 'LHE']:
text = [option, "len{}".format(key), f, accuracy_X, timeTaken_dict[key]]
elif learning_method == 'GS':
text = [option, 0, f, accuracy_X, 0]
tuple.extend(text)
# print("option: {}, f: {}, actualIt: {}, accuracy: {}".format(option, f, actualIt, accuracy_X))
save_csv_record(join(data_directory, csv_filename), tuple)
# -- Read, aggregate, and pivot data for all options
df1 = pd.read_csv(join(data_directory, csv_filename))
# print("\n-- df1: (length {}):\n{}".format(len(df1.index), df1.head(15)))
# Aggregate repetitions
df2 = df1.groupby(['option', 'lensplit', 'f']).agg \
({'accuracy': [np.mean, np.std, np.size], # Multiple Aggregates
})
df2.columns = ['_'.join(col).strip() for col in df2.columns.values] # flatten the column hierarchy
df2.reset_index(inplace=True) # remove the index hierarchy
df2.rename(columns={'accuracy_size': 'count'}, inplace=True)
# print("\n-- df2 (length {}):\n{}".format(len(df2.index), df2.head(15)))
df3 = df1.groupby(['option', 'lensplit', 'f']).agg({'timetaken': [np.median] })
df3.columns = ['_'.join(col).strip() for col in df3.columns.values] # flatten the column hierarchy
df3.reset_index(inplace=True) # remove the index hierarchy
# resultdf3 = df3.sort(['timetaken'], ascending=1)
# print("\n-- df3 (length {}):\n{}".format(len(df3.index), df3.head(15)))
X_time_median_dict = {}
Y_acc_dict = {}
Y_std_dict = {}
for option in option_vec:
Y_acc_dict[option] = df2.loc[(df2['option'] == option), "accuracy_mean"].values
Y_std_dict[option] = df2.loc[(df2['option'] == option), "accuracy_std"].values
X_time_median_dict[option] = df3.loc[(df3['option'] == option), "timetaken_median"].values
# print("option: ", option)
# print("Y_acc_dict[option]: ", Y_acc_dict[option])
# print("Y_std_dict[option]: ", Y_std_dict[option])
# print("X_time_median_dict[option]: ", X_time_median_dict[option])
# -- Setup figure
fig_filename = '{}_{}.pdf'.format(FILEZNAME, CHOICE)
mpl.rc('font', **{'family': 'sans-serif', 'sans-serif': [u'Arial', u'Liberation Sans']})
mpl.rcParams['axes.labelsize'] = 18
mpl.rcParams['xtick.labelsize'] = 16
mpl.rcParams['ytick.labelsize'] = 16
mpl.rcParams['axes.titlesize'] = 16
mpl.rcParams['legend.fontsize'] = 14
mpl.rcParams['grid.color'] = '777777' # grid color
mpl.rcParams['xtick.major.pad'] = 2 # padding of tick labels: default = 4
mpl.rcParams['ytick.major.pad'] = 1 # padding of tick labels: default = 4
mpl.rcParams['xtick.direction'] = 'out' # default: 'in'
mpl.rcParams['ytick.direction'] = 'out' # default: 'in'
mpl.rcParams['figure.figsize'] = [4, 4]
fig = figure()
ax = fig.add_axes([0.13, 0.17, 0.8, 0.8])
SHOW_ARROWS = True
for choice, color, learning_method, label, linewidth, marker in \
zip(option_vec, facecolor_vec, learning_method_vec, label_vec, linewidth_vec, marker_vec):
if learning_method == 'Holdout':
# Draw std
X1 = X_time_median_dict[choice]
s = X1.argsort()
X1 = X1[s]
Y1 = Y_acc_dict[choice][s]
Y2 = Y_std_dict[choice][s]
if SHOW_STD:
ax.fill_between(X1, Y1 + Y2, Y1 - Y2, facecolor=color, alpha=0.2, edgecolor=None, linewidth=0)
ax.plot(X1, Y1 + Y2, linewidth=0.5, color='0.8', linestyle='solid')
ax.plot(X1, Y1 - Y2, linewidth=0.5, color='0.8', linestyle='solid')
ax.set_ylim(bottom=ymin)
ax.plot(X1, Y1, linewidth=linewidth, color=color, linestyle='solid', label=label, zorder=20, marker='x', markersize=linewidth + 5, markeredgewidth=1)
ax.annotate(np.round(X1[1], decimals=1), xy=(X1[1], Y1[1] - 0.05), color=color, va='center', annotation_clip=False, zorder=5)
else:
ax.scatter(list(X1), list(Y1),
color=color, label=label, marker='x', s=42)
elif learning_method == 'GS':
ax.plot([1e-4, 1e4], [Y_acc_dict[choice], Y_acc_dict[choice]],
linewidth=1, color='black',
linestyle='dashed', zorder=0,
marker=None,
label=label,
)
else: # For all other
if SHOW_STD:
ax.errorbar(list(X_time_median_dict[choice]), list(Y_acc_dict[choice]), yerr=Y_std_dict[choice],
fmt='-o', linewidth=2, color=color,
label=label, marker=marker, markersize=8)
ax.annotate(np.round(X_time_median_dict[choice], decimals=2), xy=(X_time_median_dict[choice], Y_acc_dict[choice]-0.05), color=color, va='center',
annotation_clip=False, zorder=5)
else:
ax.scatter(list(X_time_median_dict[choice]), list(Y_acc_dict[choice]),
color=color, label=label, marker=marker, s=42)
if SHOW_ARROWS:
dce_opt = 'opt4'
holdout_opt = 'opt5'
ax.annotate(s='', xy=(X_time_median_dict[dce_opt], Y_acc_dict[dce_opt]-0.3), xytext=(X_time_median_dict[holdout_opt][2]+0.02, Y_acc_dict[dce_opt]-0.3), arrowprops=dict(arrowstyle='<->'))
ax.annotate(str(int(np.round(X_time_median_dict[holdout_opt][2] / X_time_median_dict[dce_opt]))) + 'x', xy=((X_time_median_dict[dce_opt] + X_time_median_dict[holdout_opt][2])/100, Y_acc_dict[dce_opt]-0.28),
color='black', va='center',
# bbox = dict(boxstyle="round,pad=0.3", fc="w"),
annotation_clip=False, zorder=5)
# -- Title and legend
title(r'$\!\!\!n\!=\!{}\mathrm{{k}}, d\!=\!{}, h\!=\!{}, f\!=\!{}$'.format(int(n / 1000), d, h, f))
handles, label_vec = ax.get_legend_handles_labels()
for i, (h, learning_method) in enumerate(zip(handles, learning_method_vec)): # remove error bars in legend
if isinstance(handles[i], collections.Container):
handles[i] = handles[i][0]
# plt.legend(loc='upper left', numpoints=1, ncol=3, fontsize=8, bbox_to_anchor=(0, 0))
SHOW_STD = False
legend = plt.legend(handles, label_vec,
loc='upper right', # 'upper right'
handlelength=2,
fontsize=12,
labelspacing=0.2, # distance between label entries
handletextpad=0.3, # distance between label and the line representation
borderaxespad=0.2, # distance between legend and the outer axes
borderpad=0.3, # padding inside legend box
numpoints=1, # put the marker only once
)
if not(SHOW_STD):
legend = plt.legend(handles, label_vec,
loc='upper right', # 'upper right'
handlelength=2,
fontsize=10,
labelspacing=0.2, # distance between label entries
handletextpad=0.3, # distance between label and the line representation
borderaxespad=0.2, # distance between legend and the outer axes
borderpad=0.3, # padding inside legend box
numpoints=1, # put the marker only once
scatterpoints=1 # display only one-scatter point in legend
)
# # legend.set_zorder(1)
frame = legend.get_frame()
frame.set_linewidth(0.0)
frame.set_alpha(0.9) # 0.8
# -- Figure settings and save
plt.xscale('log')
plt.xticks(xtick_lab, xtick_labels)
plt.yticks(ytick_lab, ytick_lab)
ax.yaxis.set_major_formatter(mpl.ticker.FormatStrFormatter('%.1f'))
ax.yaxis.set_ticks_position('left')
ax.xaxis.set_ticks_position('bottom')
ax.set_ylim(bottom=ymin)
grid(b=True, which='major', axis='both', alpha=0.2, linestyle='solid', linewidth=0.5) # linestyle='dashed', which='minor', axis='y',
grid(b=True, which='minor', axis='both', alpha=0.2, linestyle='solid', linewidth=0.5) # linestyle='dashed', which='minor', axis='y',
xlim(xmin, xmax)
ylim(ymin, ymax)
xlabel(r'Time Median (sec)', labelpad=0) # labelpad=0
ylabel(r'Accuracy', labelpad=0)
if CREATE_PDF:
savefig(join(figure_directory, fig_filename), format='pdf',
dpi=None,
edgecolor='w',
orientation='portrait',
transparent=False,
bbox_inches='tight',
pad_inches=0.05,
frameon=None)
if SHOW_PDF:
showfig(join(figure_directory, fig_filename))
if SHOW_PLOT:
plt.show()