def make_2d_plot(p, noise=None, overlap=None):
"""
Make a 2D plot for the specified slice.
@param p: The full path to the results file.
@param noise: The degree of noise to use, if none all are used.
@param overlap: The degree of overlap to use, if none all are used.
"""
# Get the data
with open(p, 'rb') as f:
sp_x, sp_y, svm_x, svm_y, param = cPickle.load(f)
# Fix the messed up sort order
ix = np.array(pd.DataFrame(param).sort_values([0, 1],
ascending=[True, True]).index).astype('i')
param = param[ix]
sp_x, sp_y = sp_x[ix], sp_y[ix]
svm_x, svm_y = svm_x[ix], svm_y[ix]
# Refactor the data
ix = []
if noise is not None:
id = 0
id2 = 1
val = noise
func = float
name = 'overlap'
term = 100. / 40
elif overlap is not None:
id = 1
id2 = 0
val = overlap
func = int
name = 'noise'
term = 100.
else:
raise 'noise and overlap are exclusive parameters'
for i in xrange(len(param)):
if func(param[i][id]) == val: ix.append(i)
x = param[ix][:, id2] * term
# Make the plots
dir = os.path.dirname(p)
plot_error((x, x), (np.median(sp_x[ix], 1), np.median(svm_x[ix], 1)),
('SP', 'SVM'), (compute_err(sp_x[ix]), compute_err(svm_x[ix])),
'% {0}'.format(name.capitalize()), '% Error', xlim=(-5, 105),
ylim=(-5, 105), show=False,
out_path=os.path.join(dir, 'train-{0}.png'.format(name)))
plot_error((x, x), (np.median(sp_y[ix], 1), np.median(svm_y[ix], 1)),
('SP', 'SVM'), (compute_err(sp_y[ix]), compute_err(svm_y[ix])),
'% {0}'.format(name.capitalize()), '% Error', xlim=(-5, 105),
ylim=(-5, 105), show=False,
out_path=os.path.join(dir, 'test-{0}.png'.format(name)))
def make_2d_plot(p, noise=None, overlap=None):
"""
Make a 2D plot for the specified slice.
@param p: The full path to the results file.
@param noise: The degree of noise to use, if none all are used.
@param overlap: The degree of overlap to use, if none all are used.
"""
# Get the data
with open(p, 'rb') as f:
sp_x, sp_y, svm_x, svm_y, param = cPickle.load(f)
# Fix the messed up sort order
ix = np.array(pd.DataFrame(param).sort_values([0, 1],
ascending=[True, True]).index).astype('i')
param = param[ix]
sp_x, sp_y = sp_x[ix], sp_y[ix]
svm_x, svm_y = svm_x[ix], svm_y[ix]
# Refactor the data
ix = []
if noise is not None:
id = 0
id2 = 1
val = noise
func = float
name = 'overlap'
term = 100. / 40
elif overlap is not None:
id = 1
id2 = 0
val = overlap
func = int
name = 'noise'
term = 100.
else:
raise 'noise and overlap are exclusive parameters'
for i in xrange(len(param)):
if func(param[i][id]) == val: ix.append(i)
x = param[ix][:, id2] * term
# Make the plots
dir = os.path.dirname(p)
plot_error((x, x), (np.median(sp_x[ix], 1), np.median(svm_x[ix], 1)),
('SP', 'SVM'), (compute_err(sp_x[ix]), compute_err(svm_x[ix])),
'% {0}'.format(name.capitalize()), '% Error', xlim=(-5, 105),
ylim=(-5, 105), show=False,
out_path=os.path.join(dir, 'train-{0}.png'.format(name)))
plot_error((x, x), (np.median(sp_y[ix], 1), np.median(svm_y[ix], 1)),
('SP', 'SVM'), (compute_err(sp_y[ix]), compute_err(svm_y[ix])),
'% {0}'.format(name.capitalize()), '% Error', xlim=(-5, 105),
ylim=(-5, 105), show=False,
out_path=os.path.join(dir, 'test-{0}.png'.format(name)))
def main(base_path, ntrials=4, seed=123456789):
"""
Run the experiments.
@param base_path: The full path to where the data should be stored.
@param ntrials: The number of trials to use for the experiment.
@param seed: The seed for the random number generator.
"""
# X-Axis data
npoints = 11
pct_noises = np.linspace(0, 1, npoints)
x = (pct_noises * 100, pct_noises * 100)
# Run the experiment
results = Parallel(n_jobs=-1)(
delayed(base_experiment)(os.path.join(base_path, 'run-{0}'.format(i)),
seed2)
for i, seed2 in enumerate(generate_seeds(ntrials, seed), 1))
u_sp = np.zeros((len(results), npoints))
u_ip = np.zeros((len(results), npoints))
o_sp = np.zeros((len(results), npoints))
o_ip = np.zeros((len(results), npoints))
for i, (a, b, c, d) in enumerate(results):
u_sp[i], u_ip[i], o_sp[i], o_ip[i] = a, b, c, d
# Save the results
with open(os.path.join(base_path, 'results.pkl'), 'wb') as f:
cPickle.dump((u_sp, u_ip, o_sp, o_ip), f, cPickle.HIGHEST_PROTOCOL)
# Make some plots
e = (compute_err(u_sp, axis=0), compute_err(u_ip, axis=0))
y = (np.median(u_sp, 0), np.median(u_ip, 0))
plot_error(x,
y, ('SP Output', 'Raw Data'),
e,
'% Noise',
'Uniqueness [%]',
xlim=False,
ylim=(-5, 105),
out_path=os.path.join(base_path, 'uniqueness.png'),
show=False)
e = (compute_err(o_sp, axis=0), compute_err(o_ip, axis=0))
y = (np.median(o_sp, 0), np.median(o_ip, 0))
plot_error(x,
y, ('SP Output', 'Raw Data'),
e,
'% Noise',
'Normalized Overlap [%]',
xlim=False,
ylim=(-5, 105),
out_path=os.path.join(base_path, 'overlap.png'),
show=False)
def plot_single_run(bp):
"""
Create an error plot for a single run.
@param bp: The base path.
"""
def read(p):
"""
Read in the data.
@param p: The path to the file to read.
@return: The results.
"""
with open(p, 'rb') as f:
reader = csv.reader(f)
data = []
for row in reader:
data.append(float(row[1]))
return np.array(data) * 100
def get_data(p):
"""
Get all of the results.
@param p: The directory to obtain the data in.
@return: The results.
"""
permanence = []
for d in os.listdir(p):
npath = os.path.join(p, d)
if os.path.isdir(npath):
permanence.append(
read(os.path.join(npath, 'permanence_boost.csv')))
return np.array(permanence)
data = get_data(bp)
plot_error(show=True,
x_series=(np.arange(data.shape[1]), ),
y_series=(np.median(data, 0), ),
y_errs=(compute_err(data, axis=0), ),
xlim=(0, 20),
ylim=(0, 100),
x_label='Iteration',
y_label='% Columns Boosted')
def plot_single_run(bp): """ Create an error plot for a single run. @param bp: The base path. """ def read(p): """ Read in the data. @param p: The path to the file to read. @return: The results. """ with open(p, 'rb') as f: reader = csv.reader(f) data = [] for row in reader: data.append(float(row[1])) return np.array(data) * 100 def get_data(p): """ Get all of the results. @param p: The directory to obtain the data in. @return: The results. """ permanence = [] for d in os.listdir(p): npath = os.path.join(p, d) if os.path.isdir(npath): permanence.append(read(os.path.join(npath, 'permanence_boost.csv'))) return np.array(permanence) data = get_data(bp) plot_error(show=True, x_series=(np.arange(data.shape[1]),), y_series=(np.median(data, 0),), y_errs=(compute_err(data, axis=0), ), xlim=(0, 20), ylim=(0, 100), x_label='Iteration', y_label='% Columns Boosted' )
def main(base_path, ntrials=4, seed=123456789):
"""
Run the experiments.
@param base_path: The full path to where the data should be stored.
@param ntrials: The number of trials to use for the experiment.
@param seed: The seed for the random number generator.
"""
# X-Axis data
npoints = 11
pct_noises = np.linspace(0, 1, npoints)
x = (pct_noises * 100, pct_noises * 100)
# Run the experiment
results = Parallel(n_jobs=-1)(delayed(base_experiment)(
os.path.join(base_path, 'run-{0}'.format(i)), seed2) for i, seed2 in
enumerate(generate_seeds(ntrials, seed), 1))
u_sp = np.zeros((len(results), npoints))
u_ip = np.zeros((len(results), npoints))
o_sp = np.zeros((len(results), npoints))
o_ip = np.zeros((len(results), npoints))
for i, (a, b, c, d) in enumerate(results):
u_sp[i], u_ip[i], o_sp[i], o_ip[i] = a, b, c, d
# Save the results
with open(os.path.join(base_path, 'results.pkl'), 'wb') as f:
cPickle.dump((u_sp, u_ip, o_sp, o_ip), f, cPickle.HIGHEST_PROTOCOL)
# Make some plots
e = (compute_err(u_sp, axis=0), compute_err(u_ip, axis=0))
y = (np.median(u_sp, 0), np.median(u_ip, 0))
plot_error(x, y, ('SP Output', 'Raw Data'), e, '% Noise', 'Uniqueness [%]',
xlim=False, ylim=(-5, 105), out_path=os.path.join(base_path,
'uniqueness.png'), show=False)
e = (compute_err(o_sp, axis=0), compute_err(o_ip, axis=0))
y = (np.median(o_sp, 0), np.median(o_ip, 0))
plot_error(x, y, ('SP Output', 'Raw Data'), e, '% Noise',
'Normalized Overlap [%]', xlim=False, ylim=(-5, 105),
out_path=os.path.join(base_path, 'overlap.png'), show=False)
def main2(base_path):
"""
@param base_path: Full path to pickle file to work with.
"""
# Mapping of independent variables to indexes
data_index = {
'fit_time':0,
'learn_fit_time':1,
'pred_fit_time':2,
'input_uniqueness':3,
'input_overlap':4,
'input_correlation':5,
'sp_uniqueness':6,
'sp_overlap':7,
'sp_correlation':8
}
# Get the data
with open(base_path, 'rb') as f:
x, y = cPickle.load(f)
x = sorted(set(x[-1])) # For now work with 1D
# Pull out data for this plot
y1 = (y[data_index['input_uniqueness']], y[data_index['sp_uniqueness']])
y2 = (y[data_index['input_overlap']], y[data_index['sp_overlap']])
y3 = (y[data_index['input_correlation']], y[data_index['sp_correlation']])
# Refactor the data
x_series = (x, x, x)
med = lambda y: np.median(y, axis=1) * 100
err = lambda y: compute_err(y, axis=1) * 100
y1_series = map(med, y1)
y1_errs = map(err, y1)
y2_series = map(med, y2)
y2_errs = map(err, y2)
y3_series = map(med, y3)
y3_errs = map(err, y3)
# Make the main plot
fig = plt.figure(figsize=(21, 20), facecolor='white')
ax = fig.add_subplot(111)
ax.spines['top'].set_color('none')
ax.spines['bottom'].set_color('none')
ax.spines['left'].set_color('none')
ax.spines['right'].set_color('none')
ax.tick_params(labelcolor='w', top='off', bottom='off', left='off',
right='off')
# Make subplots
ax1 = fig.add_subplot(311)
plot_error(show=False, legend=False, ax=ax1, title='Uniqueness',
x_series=x_series, y_series=y1_series, y_errs=y1_errs, ylim=(-5, 105))
ax2 = fig.add_subplot(312, sharex=ax1, sharey=ax1)
plot_error(show=False, legend=False, ax=ax2, title='Overlap',
x_series=x_series, y_series=y2_series, y_errs=y2_errs, ylim=(-5, 105))
ax3 = fig.add_subplot(313, sharex=ax1, sharey=ax1)
plot_error(show=False, legend=False, ax=ax3, title='Correlation',
x_series=x_series, y_series=y3_series, y_errs=y3_errs, ylim=(-5, 105))
plt.tight_layout(h_pad=2)
plt.show()
def main(ntrials=10, seed=123456798):
"""
Vary the amount of noise and overlap and see how the system performs.
@param ntrials: The number of trials to perform.
@param seed: The seed for the random number generator.
"""
# Amount to vary each type by
pct_noises = np.linspace(0, 1, 101)
noverlap_bits = np.arange(0, 41)
# Vary the noise
results = Parallel(n_jobs=-1)(delayed(base_experiment)(
noise, 0, 'noise-{0}'.format(i), ntrials, False, seed)
for i, noise in enumerate(pct_noises, 1))
noise_x, noise_y, svm_noise_x, svm_noise_y = [], [], [], []
noise_x_err1, noise_x_err2, noise_y_err1, noise_y_err2 = [], [], [], []
svm_noise_x_err1, svm_noise_x_err2 = [], []
svm_noise_y_err1, svm_noise_y_err2 = [], []
for sp_x, sp_y, svm_x, svm_y in results:
noise_x.append(np.median(sp_x))
noise_y.append(np.median(sp_y))
svm_noise_x.append(np.median(svm_x))
svm_noise_y.append(np.median(svm_y))
e = compute_err(sp_x, axis=None)
noise_x_err1.append(e[0])
noise_x_err2.append(e[1])
e = compute_err(sp_y, axis=None)
noise_y_err1.append(e[0])
noise_y_err2.append(e[1])
e = compute_err(svm_x, axis=None)
svm_noise_x_err1.append(e[0])
svm_noise_x_err2.append(e[1])
e = compute_err(svm_y, axis=None)
svm_noise_y_err1.append(e[0])
svm_noise_y_err2.append(e[1])
noise_x_err = (svm_noise_x_err1, svm_noise_x_err2)
noise_y_err = (svm_noise_y_err1, svm_noise_y_err2)
svm_noise_x_err = (svm_noise_x_err1, svm_noise_x_err2)
svm_noise_y_err = (svm_noise_y_err1, svm_noise_y_err2)
# Vary the overlaps
results = Parallel(n_jobs=-1)(
delayed(base_experiment)(0.35, overlap, 'overlap-{0}'.format(i),
ntrials, False, seed)
for i, overlap in enumerate(noverlap_bits, 1))
overlap_x, overlap_y, svm_overlap_x, svm_overlap_y = [], [], [], []
overlap_x_err1, overlap_x_err2 = [], []
overlap_y_err1, overlap_y_err2 = [], []
svm_overlap_x_err1, svm_overlap_x_err2 = [], []
svm_overlap_y_err1, svm_overlap_y_err2 = [], []
for sp_x, sp_y, svm_x, svm_y in results:
overlap_x.append(np.median(sp_x))
overlap_y.append(np.median(sp_y))
svm_overlap_x.append(np.median(svm_x))
svm_overlap_y.append(np.median(svm_y))
e = compute_err(sp_x, axis=None)
overlap_x_err1.append(e[0])
overlap_x_err2.append(e[1])
e = compute_err(sp_y, axis=None)
overlap_y_err1.append(e[0])
overlap_y_err2.append(e[1])
e = compute_err(svm_x, axis=None)
svm_overlap_x_err1.append(e[0])
svm_overlap_x_err2.append(e[1])
e = compute_err(svm_y, axis=None)
svm_overlap_y_err1.append(e[0])
svm_overlap_y_err2.append(e[1])
overlap_x_err = (svm_overlap_x_err1, svm_overlap_x_err2)
overlap_y_err = (svm_overlap_y_err1, svm_overlap_y_err2)
svm_overlap_x_err = (svm_overlap_x_err1, svm_overlap_x_err2)
svm_overlap_y_err = (svm_overlap_y_err1, svm_overlap_y_err2)
# # Save the results
p = os.path.join(os.path.expanduser('~'), 'scratch', 'novelty_experiments')
with open(os.path.join(p, 'results.pkl'), 'wb') as f:
cPickle.dump((noise_x, noise_y), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_noise_x, svm_noise_y), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((noise_x_err, noise_y_err), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_noise_x_err, svm_noise_y_err), f,
cPickle.HIGHEST_PROTOCOL)
cPickle.dump((overlap_x, overlap_y), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_overlap_x, svm_overlap_y), f,
cPickle.HIGHEST_PROTOCOL)
cPickle.dump((overlap_x_err, overlap_y_err), f,
cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_overlap_x_err, svm_overlap_y_err), f,
cPickle.HIGHEST_PROTOCOL)
# Make the plots
plot_error((pct_noises * 100, pct_noises * 100), (noise_x, svm_noise_x),
('SP', 'SVM'), (noise_x_err, svm_noise_x_err),
'% Noise',
'% Error',
'Noise: Base Class',
out_path=os.path.join(p, 'noise_base.png'),
xlim=(-5, 105),
ylim=(-5, 105),
show=False)
plot_error((pct_noises * 100, pct_noises * 100), (noise_y, svm_noise_y),
('SP', 'SVM'), (noise_y_err, svm_noise_y_err),
'% Noise',
'% Error',
'Noise: Novelty Class',
out_path=os.path.join(p, 'noise_novelty.png'),
xlim=(-5, 105),
ylim=(-5, 105),
show=False)
noverlap_pct = noverlap_bits / 40. * 100
plot_error((noverlap_pct, noverlap_pct), (overlap_x, svm_overlap_x),
('SP', 'SVM'), (overlap_x_err, svm_overlap_x_err),
'% Overlapping Bits',
'% Error',
'Overlap: Base Class',
out_path=os.path.join(p, 'overlap_base.png'),
xlim=(-5, 105),
ylim=(-5, 105),
show=False)
plot_error((noverlap_pct, noverlap_pct), (overlap_y, svm_overlap_y),
('SP', 'SVM'), (overlap_y_err, svm_overlap_y_err),
'% Overlapping Bits',
'% Error',
'Overlap: Novelty Class',
out_path=os.path.join(p, 'overlap_novelty.png'),
xlim=(-5, 105),
ylim=(-5, 105),
show=False)
def main(ntrials=10, seed=123456798):
"""
Vary the amount of noise and overlap and see how the system performs.
@param ntrials: The number of trials to perform.
@param seed: The seed for the random number generator.
"""
# Amount to vary each type by
pct_noises = np.linspace(0, 1, 101)
noverlap_bits = np.arange(0, 41)
# Vary the noise
results = Parallel(n_jobs=-1)(delayed(base_experiment)(noise, 0,
'noise-{0}'.format(i), ntrials, False, seed) for i, noise in enumerate(
pct_noises, 1))
noise_x, noise_y, svm_noise_x, svm_noise_y = [], [], [], []
noise_x_err1, noise_x_err2, noise_y_err1, noise_y_err2 = [], [], [], []
svm_noise_x_err1, svm_noise_x_err2 = [], []
svm_noise_y_err1, svm_noise_y_err2 = [], []
for sp_x, sp_y, svm_x, svm_y in results:
noise_x.append(np.median(sp_x))
noise_y.append(np.median(sp_y))
svm_noise_x.append(np.median(svm_x))
svm_noise_y.append(np.median(svm_y))
e = compute_err(sp_x, axis=None)
noise_x_err1.append(e[0])
noise_x_err2.append(e[1])
e = compute_err(sp_y, axis=None)
noise_y_err1.append(e[0])
noise_y_err2.append(e[1])
e = compute_err(svm_x, axis=None)
svm_noise_x_err1.append(e[0])
svm_noise_x_err2.append(e[1])
e = compute_err(svm_y, axis=None)
svm_noise_y_err1.append(e[0])
svm_noise_y_err2.append(e[1])
noise_x_err = (svm_noise_x_err1, svm_noise_x_err2)
noise_y_err = (svm_noise_y_err1, svm_noise_y_err2)
svm_noise_x_err = (svm_noise_x_err1, svm_noise_x_err2)
svm_noise_y_err = (svm_noise_y_err1, svm_noise_y_err2)
# Vary the overlaps
results = Parallel(n_jobs=-1)(delayed(base_experiment)(0.35, overlap,
'overlap-{0}'.format(i), ntrials, False, seed) for i, overlap in
enumerate(noverlap_bits, 1))
overlap_x, overlap_y, svm_overlap_x, svm_overlap_y = [], [], [], []
overlap_x_err1, overlap_x_err2 = [], []
overlap_y_err1, overlap_y_err2 = [], []
svm_overlap_x_err1, svm_overlap_x_err2 = [], []
svm_overlap_y_err1, svm_overlap_y_err2 = [], []
for sp_x, sp_y, svm_x, svm_y in results:
overlap_x.append(np.median(sp_x))
overlap_y.append(np.median(sp_y))
svm_overlap_x.append(np.median(svm_x))
svm_overlap_y.append(np.median(svm_y))
e = compute_err(sp_x, axis=None)
overlap_x_err1.append(e[0])
overlap_x_err2.append(e[1])
e = compute_err(sp_y, axis=None)
overlap_y_err1.append(e[0])
overlap_y_err2.append(e[1])
e = compute_err(svm_x, axis=None)
svm_overlap_x_err1.append(e[0])
svm_overlap_x_err2.append(e[1])
e = compute_err(svm_y, axis=None)
svm_overlap_y_err1.append(e[0])
svm_overlap_y_err2.append(e[1])
overlap_x_err = (svm_overlap_x_err1, svm_overlap_x_err2)
overlap_y_err = (svm_overlap_y_err1, svm_overlap_y_err2)
svm_overlap_x_err = (svm_overlap_x_err1, svm_overlap_x_err2)
svm_overlap_y_err = (svm_overlap_y_err1, svm_overlap_y_err2)
# # Save the results
p = os.path.join(os.path.expanduser('~'), 'scratch', 'novelty_experiments')
with open(os.path.join(p, 'results.pkl'), 'wb') as f:
cPickle.dump((noise_x, noise_y), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_noise_x, svm_noise_y), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((noise_x_err, noise_y_err), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_noise_x_err, svm_noise_y_err), f,
cPickle.HIGHEST_PROTOCOL)
cPickle.dump((overlap_x, overlap_y), f, cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_overlap_x, svm_overlap_y), f,
cPickle.HIGHEST_PROTOCOL)
cPickle.dump((overlap_x_err, overlap_y_err), f,
cPickle.HIGHEST_PROTOCOL)
cPickle.dump((svm_overlap_x_err, svm_overlap_y_err), f,
cPickle.HIGHEST_PROTOCOL)
# Make the plots
plot_error((pct_noises * 100, pct_noises * 100), (noise_x, svm_noise_x),
('SP', 'SVM'), (noise_x_err, svm_noise_x_err), '% Noise', '% Error',
'Noise: Base Class', out_path=os.path.join(p, 'noise_base.png'),
xlim=(-5, 105), ylim=(-5, 105), show=False)
plot_error((pct_noises * 100, pct_noises * 100), (noise_y, svm_noise_y),
('SP', 'SVM'), (noise_y_err, svm_noise_y_err), '% Noise', '% Error',
'Noise: Novelty Class', out_path=os.path.join(p, 'noise_novelty.png'),
xlim=(-5, 105), ylim=(-5, 105), show=False)
noverlap_pct = noverlap_bits / 40. * 100
plot_error((noverlap_pct, noverlap_pct), (overlap_x, svm_overlap_x),
('SP', 'SVM'), (overlap_x_err, svm_overlap_x_err),
'% Overlapping Bits', '% Error', 'Overlap: Base Class',
out_path=os.path.join(p, 'overlap_base.png'), xlim=(-5, 105),
ylim=(-5, 105),show=False)
plot_error((noverlap_pct, noverlap_pct), (overlap_y, svm_overlap_y),
('SP', 'SVM'), (overlap_y_err, svm_overlap_y_err),
'% Overlapping Bits', '% Error', 'Overlap: Novelty Class',
out_path=os.path.join(p, 'overlap_novelty.png'), xlim=(-5, 105),
ylim=(-5, 105),show=False)
def plot_single_run(bp1, bp2):
"""Create an error plot for a single run.
@param bp1: The base path for global inhibition results.
@param bp2: The base path for local inhibition results.
"""
def read(p):
"""Read in the data.
@param p: The path to the file to read.
@return: The results.
"""
with open(p, 'r') as f:
reader = csv.reader(f)
data = []
for row in reader:
data.append(float(row[1]))
return np.array(data) * 100
def get_data(p):
"""Get all of the results.
@param p: The directory to obtain the data in.
@return: The results.
"""
permanence = []
for d in os.listdir(p):
npath = os.path.join(p, d)
if os.path.isdir(npath):
permanence.append(
read(os.path.join(npath, 'permanence_boost.csv')))
return np.array(permanence)
# Get the data
data = [get_data(bp1)]
data.append(get_data(bp2))
# Build the series
x_series = (np.arange(data[0].shape[1]), )
# Make the main plot
fig = plt.figure(figsize=(21, 20), facecolor='white')
ax = fig.add_subplot(111)
ax.spines['top'].set_color('none')
ax.spines['bottom'].set_color('none')
ax.spines['left'].set_color('none')
ax.spines['right'].set_color('none')
ax.tick_params(labelcolor='w',
top='off',
bottom='off',
left='off',
right='off')
ax.set_xlabel('Iteration')
ax.set_ylabel('% Columns Boosted')
# Make subplots
ax1 = fig.add_subplot(211)
plot_error(show=False,
legend=False,
ax=ax1,
title='Global Inhibition',
x_series=x_series,
y_series=(np.median(data[0], 0), ),
y_errs=(compute_err(data[0], axis=0), ),
xlim=(0, 200),
ylim=(0, 100))
ax2 = fig.add_subplot(212, sharex=ax1, sharey=ax1)
plot_error(show=False,
ax=ax2,
title='Local Inhibition',
legend=False,
x_series=x_series,
y_series=(np.median(data[1], 0), ),
y_errs=(compute_err(data[1], axis=0), ),
xlim=(0, 200),
ylim=(0, 100))
# Save it
# plt.show()
plt.subplots_adjust(bottom=0.15, hspace=0.3)
plt.savefig('boost_permanence.png',
format='png',
facecolor=fig.get_facecolor(),
edgecolor='none')
def plot_density_results(bp, bp2=None):
"""Average the results.
@param bp: The base path.
@param bp2: The second base path.
"""
def average(p):
"""Compute the average activations for each density.
@param p: The path to the file.
@return: The average.
"""
with open(p, 'r') as f:
reader = csv.reader(f)
data = []
for row in reader:
data.append(float(row[1]))
return np.mean(data) * 100
def get_data(p):
"""Get the data for a single run.
@param p: The path.
@return: A tuple containing the overlap and permanences.
"""
overlap, permanence = [], []
for d in os.listdir(p):
npath = os.path.join(p, d)
if os.path.isdir(npath):
overlap.append(
average(os.path.join(npath, 'overlap_boost.csv')))
permanence.append(
average(os.path.join(npath, 'permanence_boost.csv')))
return np.array(overlap), np.array(permanence)
def get_all_data(bp):
"""Get the data for all runs.
@param bp: The base path.
@return: A tuple containing the sparsity, overlap, and permanences.
"""
overlap, permanence, sparsity = [], [], []
for d in sorted([int(x) for x in os.listdir(bp)]):
sparsity.append((1 - (d / 100.)) * 100)
o, p = get_data(os.path.join(bp, str(d)))
overlap.append(o)
permanence.append(p)
return np.array(sparsity[::-1]), np.array(overlap[::-1]), \
np.array(permanence[::-1])
def make_plot_params(sparsity, overlap, permanence, title=None):
"""Generate the parameters for the plot.
@param sparsity: The sparsity array.
@param overlap: The overlap array.
@param permanence: The permanence array.
@param title: The title for the plot.
@return: A dictionary with the parameters.
"""
return {
'x_series': (sparsity, sparsity),
'y_series': (np.median(overlap, 1), np.median(permanence, 1)),
'series_names': ('Overlap Boosting', 'Permanence Boosting'),
'y_errs': (compute_err(overlap), compute_err(permanence)),
'xlim': (0, 100),
'ylim': (0, 45),
'title': title
}
data = get_all_data(bp)
if bp2 is None:
plot_error(**make_plot_params(*data))
else:
# Make main plot
fig = plt.figure(figsize=(21, 20), facecolor='white')
ax = fig.add_subplot(111)
ax.spines['top'].set_color('none')
ax.spines['bottom'].set_color('none')
ax.spines['left'].set_color('none')
ax.spines['right'].set_color('none')
ax.tick_params(labelcolor='w',
top='off',
bottom='off',
left='off',
right='off')
ax.set_xlabel('Sparsity [%]')
ax.set_ylabel('% Columns Boosted')
# Make subplots
ax1 = fig.add_subplot(211)
plot_error(show=False,
legend=False,
ax=ax1,
**make_plot_params(*data, title='Global Inhibition'))
data2 = get_all_data(bp2)
ax2 = fig.add_subplot(212, sharex=ax1, sharey=ax1)
plot_error(show=False,
ax=ax2,
**make_plot_params(*data2, title='Local Inhibition'))
# Save it
# plt.show()
plt.subplots_adjust(bottom=0.15, hspace=0.3)
plt.savefig('boost_sparseness.png',
format='png',
facecolor=fig.get_facecolor(),
edgecolor='none')
def plot_single_run(bp1, bp2):
"""
Create an error plot for a single run.
@param bp1: The base path for global inhibition results.
@param bp2: The base path for local inhibition results.
"""
def read(p):
"""
Read in the data.
@param p: The path to the file to read.
@return: The results.
"""
with open(p, 'rb') as f:
reader = csv.reader(f)
data = []
for row in reader:
data.append(float(row[1]))
return np.array(data) * 100
def get_data(p):
"""
Get all of the results.
@param p: The directory to obtain the data in.
@return: The results.
"""
permanence = []
for d in os.listdir(p):
npath = os.path.join(p, d)
if os.path.isdir(npath):
permanence.append(read(os.path.join(npath,
'permanence_boost.csv')))
return np.array(permanence)
# Get the data
data = [get_data(bp1)]
data.append(get_data(bp2))
# Build the series
x_series = (np.arange(data[0].shape[1]), )
# Make the main plot
fig = plt.figure(figsize=(21, 20), facecolor='white')
ax = fig.add_subplot(111)
ax.spines['top'].set_color('none')
ax.spines['bottom'].set_color('none')
ax.spines['left'].set_color('none')
ax.spines['right'].set_color('none')
ax.tick_params(labelcolor='w', top='off', bottom='off', left='off',
right='off')
ax.set_xlabel('Iteration')
ax.set_ylabel('% Columns Boosted')
# Make subplots
ax1 = fig.add_subplot(211)
plot_error(show=False, legend=False, ax=ax1, title='Global Inhibition',
x_series=x_series, y_series=(np.median(data[0], 0), ),
y_errs=(compute_err(data[0], axis=0),), xlim=(0, 200), ylim=(0, 100))
ax2 = fig.add_subplot(212, sharex=ax1, sharey=ax1)
plot_error(show=False, ax=ax2, title='Local Inhibition', legend=False,
x_series=x_series, y_series=(np.median(data[1], 0), ),
y_errs=(compute_err(data[1], axis=0),), xlim=(0, 200), ylim=(0, 100))
# Save it
plt.subplots_adjust(bottom=0.15, hspace=0.3)
plt.savefig('boost_permanence.png', format='png',
facecolor=fig.get_facecolor(), edgecolor='none')
def plot_density_results(bp, bp2=None):
"""
Average the results.
@param bp: The base path.
@param bp2: The second base path.
"""
def average(p):
"""
Compute the average activations for each density.
@param p: The path to the file.
@return: The average.
"""
with open(p, 'rb') as f:
reader = csv.reader(f)
data = []
for row in reader:
data.append(float(row[1]))
return np.mean(data) * 100
def get_data(p):
"""
Get the data for a single run.
@param p: The path.
@return: A tuple containing the overlap and permanences.
"""
overlap, permanence = [], []
for d in os.listdir(p):
npath = os.path.join(p, d)
if os.path.isdir(npath):
overlap.append(average(os.path.join(npath,
'overlap_boost.csv')))
permanence.append(average(os.path.join(npath,
'permanence_boost.csv')))
return np.array(overlap), np.array(permanence)
def get_all_data(bp):
"""
Get the data for all runs.
@param bp: The base path.
@return: A tuple containing the sparsity, overlap, and permanences.
"""
overlap, permanence, sparsity = [], [], []
for d in sorted([int(x) for x in os.listdir(bp)]):
sparsity.append((1 - (d / 100.)) * 100)
o, p = get_data(os.path.join(bp, str(d)))
overlap.append(o)
permanence.append(p)
return np.array(sparsity[::-1]), np.array(overlap[::-1]), \
np.array(permanence[::-1])
def make_plot_params(sparsity, overlap, permanence, title=None):
"""
Generate the parameters for the plot.
@param sparsity: The sparsity array.
@param overlap: The overlap array.
@param permanence: The permanence array.
@param title: The title for the plot.
@return: A dictionary with the parameters.
"""
return {'x_series':(sparsity, sparsity),
'y_series':(np.median(overlap, 1), np.median(permanence, 1)),
'series_names':('Overlap Boosting', 'Permanence Boosting'),
'y_errs':(compute_err(overlap), compute_err(permanence)),
'xlim':(0, 100), 'ylim':(0, 45), 'title':title
}
data = get_all_data(bp)
if bp2 is None:
plot_error(**make_plot_params(*data))
else:
# Make main plot
fig = plt.figure(figsize=(21, 20), facecolor='white')
ax = fig.add_subplot(111)
ax.spines['top'].set_color('none')
ax.spines['bottom'].set_color('none')
ax.spines['left'].set_color('none')
ax.spines['right'].set_color('none')
ax.tick_params(labelcolor='w', top='off', bottom='off', left='off',
right='off')
ax.set_xlabel('Sparsity [%]')
ax.set_ylabel('% Columns Boosted')
# Make subplots
ax1 = fig.add_subplot(211)
plot_error(show=False, legend=False, ax=ax1, **make_plot_params(*data,
title='Global Inhibition'))
data2 = get_all_data(bp2)
ax2 = fig.add_subplot(212, sharex=ax1, sharey=ax1)
plot_error(show=False, ax=ax2, **make_plot_params(*data2,
title='Local Inhibition'))
# Save it
plt.subplots_adjust(bottom=0.15, hspace=0.3)
plt.savefig('boost_sparseness.png', format='png',
facecolor=fig.get_facecolor(), edgecolor='none')