def sparse_sift(image, fraction=1.0):
""" sparse oriented SIFT at Harris-LaPlace and Difference of Gaussians keypoints
use VLFEAT vl_covdet through octave; expects a grayscale image
"""
octave.eval("addpath ~/CC/vlfeat/vlfeat-0.9.20/toolbox")
octave.eval("vl_setup")
octave.push("im", image)
octave.eval("im = single(im);")
octave.eval(
"[kp,sift_hl] = vl_covdet(im, 'method', 'HarrisLaplace', 'EstimateOrientation', true); "
)
octave.eval(
"[kp,sift_dog] = vl_covdet(im, 'method', 'DoG', 'EstimateOrientation', true); "
)
octave.eval("descrs = [sift_hl, sift_dog];")
descriptors = octave.pull("descrs")
kp = octave.pull("kp")
# flip from column-major to row-major
descriptors = descriptors.T
if fraction < 1.0:
descriptors = random_sample(descriptors, fraction)
return kp, descriptors
def set(self, expid, variables, values):
'''
Writes one or more variables to the workspace of the current Octave session
'''
n = len(variables)
for i in range(n):
try:
octave.push(variables[i], values[i])
except:
pass
def fit(self, K, y):
"""Learn a low-rank kernel approximation.
:param K: (``numpy.ndarray``) or of (``Kinterface``). The kernel to be approximated with G.
:param y: (``numpy.ndarray``) Class labels :math:`y_i \in {-1, 1}` or regression targets.
"""
# Convert to explicit form
K = K[:, :]
y = y.reshape((len(y), 1))
# Call original implementation
octave.push(["K", "y", "rank", "centering", "kappa", "delta", "tol"], [
K, y, self.rank, self.centering, self.kappa, self.delta, self.tol
])
octave.eval(
"[G, P, Q, R, error1, error2, error, predicted_gain, true_gain] = csi(K, y, rank, centering, kappa, delta, tol)",
verbose=False)
G, P, Q, R, error1, error2, error, predicted_gain, true_gain = \
octave.pull(["G", "P", "Q", "R", "error1", "error2", "error", "predicted_gain", "true_gain"])
R = np.atleast_2d(np.array(R))
# Octave indexes from 1
P = P.ravel().astype(int) - 1
# Resort rows to respect the order
n, k = K.shape[0], self.rank
self.I = self.active_set_ = list(P[:k])
Go = np.zeros((n, k))
Qo = np.zeros((n, k))
Ro = np.zeros((k, k))
km = min(k, G.shape[1])
Go[P, :km] = G[:, :km]
Qo[P, :km] = Q[:, :km]
Ro[:km, :km] = R[:km, :km]
self.G = Go[:, :self.rank]
self.P = P[:self.rank]
self.Q = Qo[:, :]
self.R = Ro[:, :self.rank]
self.error1 = error1
self.error1 = error2
self.error = error
self.predicted_gain = predicted_gain
self.true_gain = true_gain
self.trained = True
self.active_set_ = self.I[:self.rank]
def plot_calibration_mapping(calibration_model,
min_score,
max_score,
resolution=1000,
file_name='calibration_mapping.png'):
# Function for plotting what probabilities different scores get mapped to.
# "General purpose prediction function"
# PERHAPS ADD PROBABILITY DISTRIBUTION OF TRAINING DATA (OR TESTING?)?
# WOULD INDICATE HOW MANY SAMPLES FALL INTO ONE BIN.
diff = max_score - min_score
scores = [
min_score + i * diff / float(resolution) for i in range(resolution + 1)
]
try: # IR model
probabilities = calibration_model.predict(scores)
except:
try: # ENIR
import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
enir = importr('enir')
r = robjects.r
# Automatic conversion or numpy arrays to R-vectors
import rpy2.robjects.numpy2ri
rpy2.robjects.numpy2ri.activate()
# ENIR-MODEL MIGHT NEED TO BE PUSHED TO R-ENVIRONMENT?
probabilities = enir.enir_predict(calibration_model,
robjects.FloatVector(scores))
probabilities = np.array(probabilities)
except:
try: # BBQ
from oct2py import octave
octave.eval("addpath('./calibration/BBQ/')", verbose=False)
octave.push('scores', scores, verbose=False)
octave.push('calibration_model',
calibration_model,
verbose=False)
octave.eval(
'probabilities = predict(calibration_model, scores, 1)',
verbose=False)
probabilities = octave.pull('probabilities', verbose=False)
probabilities = np.array([item[0] for item in probabilities])
except:
pass # Continue with BIR and WABIR? RCIR?
# Plot score vs. probability:
plt.plot(scores, probabilities)
plt.title("Calibration mapping")
plt.savefig(file_name)
plt.gcf().clear()
def dense_sift(image, fraction=1.0):
""" dense SIFT
use VLFEAT vl_phow through octave; expects a grayscale image
"""
octave.push("im", image)
octave.eval("im = single(im);")
octave.eval("[kp,siftd] = vl_phow(im); ")
descriptors = octave.pull("siftd")
# flip from column-major to row-major
descriptors = descriptors.T
if fraction < 1.0:
descriptors = random_sample(descriptors, fraction)
return descriptors
# Get good trials indices
goodtrials = np.where(df['badtrial'] == 0)[0]
df = df.iloc[goodtrials]
mod = sm.OLS(df["amp_['POz']_0.4-0.8"],
sm.add_constant(df[variable]),
hasconst=True).fit()
bic_erps.loc[p, variable] = mod.bic
bic_beta.loc[p, variable] = mod.params[variable]
bic_erps.to_csv(opj(outpath, 'bic_erps.csv'))
# Use octave to run the VBA-toolbox
octave.push('L', np.asarray(bic_erps.transpose()) * -1)
octave.addpath('/matlab/vbatoolbox')
octave.addpath('/matlab/vbatoolbox/core')
octave.addpath('/matlab/vbatoolbox/core/display')
octave.addpath('/matlab/vbatoolbox/utils')
octave.eval("options.DisplayWin = 0")
p, out = octave.eval("VBA_groupBMC(L, options)", nout=2)
# Save to plot
file = open(opj(outpath, 'erps_olsmean_VBAmodelcomp.pkl'), "wb")
pickle.dump(out, file)
# ########################################################################
# Mass univariate regression
###########################################################################
# Load model data
label="Credible intervals")
# Create reliability diagram:
isotonic.plot_reliability_diagram(isotonic_regression_model,
test_scores,
test_class,
file_name="./" + test_description +
"/reliability_diagram" + ".png",
label="Reliability diagram")
# Naeini's model
# Comparison to Naeini's model (it's a matlab model, using Octave).
print(
"Training Naeini-model. This might take a while (~20 minutes on a laptop)."
)
octave.push('training_scores', naeini_training_scores, verbose=False)
octave.push('training_class', naeini_training_class, verbose=False)
octave.eval("options.N0 = 2", verbose=False)
octave.eval(
"BBQ_model = build(training_scores', training_class', options)",
verbose=False)
# In the following, '1' indicates model averaging, as done in the paper by Naeini & al.
octave.eval("training_bbq_prob = predict(BBQ_model, training_scores, 1)",
verbose=False)
training_bbq_prob = octave.pull("training_bbq_prob", verbose=False)
training_bbq_prob = np.array([item[0] for item in training_bbq_prob])
octave.push('test_scores', test_scores, verbose=False)
octave.eval("test_bbq_prob = predict(BBQ_model, test_scores, 0)",
verbose=False)
test_bbq_prob = octave.pull("test_bbq_prob", verbose=False)
test_bbq_prob = np.array([item[0] for item in test_bbq_prob])
if args.n_features_keep:
weights = weights[..., sorted_inds_keep]
# get the shape of the weights and make sure they satisfy certain conditions
w_x, w_y, w_in, w_out = weights.shape
# compute the new number of features you will have and make placeholder to store them
nx = args.input_w - (w_x - 1)
ny = args.input_h - (w_y - 1)
w_out_new = (nx // args.stride_x) * (ny // args.stride_y) * w_out
nonshared = np.zeros([args.input_w, args.input_h, w_in, w_out_new],
dtype=np.float64)
# fill in the original features in the simple cell tensor
count = 0
for k in range(w_out):
for i in range(0, nx, args.stride_x):
for j in range(0, ny, args.stride_y):
nonshared[i:i + w_x, j:j + w_y, :, count] = weights[:, :, :, k]
count += 1
# write the new features
write_fpath = os.path.join(args.save_dir, os.path.split(fpath)[1])
feat_data[0]['values'][0] = nonshared
octave.push(['write_fpath', 'feat_data'], [write_fpath, feat_data])
octave.eval('writepvpsharedweightfile(write_fpath, feat_data)')
logging.info('NONSHARED GRID SIZE IS {}x{}x{}.'.format(ny // args.stride_y,
nx // args.stride_x,
w_out))
# Create ENIR model using R:
enir_model = enir.enir_build(
robjects.FloatVector(training_scores.tolist()),
robjects.BoolVector(training_class.tolist()))
enir_prob = enir.enir_predict(enir_model,
robjects.FloatVector(test_scores.tolist()))
# Convert to numpy.array:
enir_prob = np.array(enir_prob)
# Kill session?
# Create isotonic regression model
ir_model = IsotonicRegression(y_min=0, y_max=1, out_of_bounds='clip')
ir_model.fit(X=training_scores, y=training_class)
# Create BBQ-model
octave.push('training_scores', training_scores, verbose=False)
octave.push('training_class', training_class, verbose=False)
octave.eval('options.N0 = 2', verbose=False)
octave.eval(
"bbq_model = build(training_scores', training_class', options)",
verbose=False)
octave.push('test_scores', test_scores)
octave.eval("test_prob = predict(bbq_model, test_scores, 1)",
verbose=False)
bbq_test_prob = octave.pull('test_prob', verbose=False)
bbq_test_prob = [item[0] for item in bbq_test_prob]
# Create RCIR model with d = .20 (?)
# QUICK HACK TO TEST RCIR-CV INSTEAD OF RCIR WITH SOME d:
# NOTE THAT RCIR-CV GETS _MORE_ _DATA_ THAN
# rcir_model = isotonic.train_rcir_cv(training_scores, training_class , d=.1)
def snmf_fhals(A, k, init='normal'):
"""
Nonnegative Matrix Factorization.
Hierarchical alternating least squares algorithm
for computing the approximate low-rank nonnegative matrix factorization of
a square `(m, m)` matrix `A`. Given the target rank `k << m`,
the input matrix `A` is factored as `A = W Wt`. The nonnegative factor
matrices `W` and `Wt` are of dimension `(m, k)` and `(k, m)`, respectively.
Parameters
----------
A : array_like, shape `(m, m)`.
Real nonnegative input matrix.
k : integer, `k << m`.
Target rank.
init : str `{'normal'}`.
'normal' : Factor matrices are initialized with nonnegative
Gaussian random numbers.
tol : float, default: `tol=1e-4`.
Tolerance of the stopping condition.
maxiter : integer, default: `maxiter=100`.
Number of iterations.
verbose : boolean, default: `verbose=False`.
The verbosity level.
Returns
-------
W: array_like, `(m, k)`.
Solution to the non-negative least squares problem.
"""
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Error catching
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
m, n = A.shape
assert m == n
if (A < 0).any():
raise ValueError("Input matrix with nonnegative elements is required.")
if A.dtype == sci.float32:
data_type = sci.float32
elif A.dtype == sci.float64:
data_type = sci.float64
else:
raise ValueError("A.dtype is not supported.")
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Initialization methods for factor matrices W and H
# 'normal': nonnegative standard normal random init
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if init == 'normal':
n, _ = A.shape
H = 2 * np.sqrt(np.mean(np.mean(A)) / k) * np.random.rand(n, k)
maxiter = 10000
tol = 1e-3
alpha = np.max(H)**2
W = H.copy()
I_k = alpha * np.identity(k)
else:
raise ValueError('Initialization method is not supported.')
#End if
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Iterate the HALS algorithm until convergence or maxiter is reached
# i) Update factor matrix H and normalize columns
# ii) Update low-dimensional factor matrix W
# iii) Compute fit log( ||A-WH|| )
# -> break if fit <-5 or fit_change < tol
#~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
projnorm = float('inf')
left = H.T.dot(H)
right = A.dot(H)
octave.push('A', A)
octave.push('k', k)
for niter in range(maxiter):
octave.push('W', W)
octave.push('H', H)
octave.push('left', left)
octave.push('right', right)
octave.eval("[W, H, left, right, violation] = symnmf_anls_iter(A, W, H, left, right, k)", verbose=False)
W = octave.pull("W")
H = octave.pull("H")
left = octave.pull("left")
right = octave.pull("right")
violation = octave.pull("violation")
if niter == 0:
initgrad = violation
else:
projnorm = violation
if projnorm < tol * initgrad:
break
return H
for frame_num, lca_fpath in enumerate(lca_fpaths):
# read in the pvp files and get the lca features
pvp_data = octave.readpvpfile(lca_fpath)
weights = pvp_data[0]['values'][0]
w_out_original = weights.shape[-1]
# take out the current frame from the strf and add to the weights
mouse_strfs_frame = mouse_strfs[:, :, frame_num, :]
mouse_strfs_frame = mouse_strfs_frame[:, :, None, :]
try: # if shapes don't line up
weights = np.concatenate((weights, mouse_strfs_frame), 3)
except ValueError:
raise
# add the new weights back to the pvp_data structure and push to octave session
pvp_data[0]['values'][0] = weights
write_fpath = os.path.join(args.save_dir, os.path.split(lca_fpath)[1])
octave.push(['pvp_data', 'write_fpath'], [pvp_data, write_fpath])
# write the new pvp weight file
octave.eval('writepvpsharedweightfile(write_fpath, pvp_data)')
print('[INFO] THE NEW NUMBER OF FEATURES IS {}.'.format(weights.shape[-1]))
# write out the indices for each cell in the dictionary
cell_ids = [cell_id.split('_')[1] for cell_id in cell_ids]
cell_inds = [cell_ind + w_out_original for cell_ind in cell_inds]
df = pd.DataFrame(zip(cell_ids, cell_inds), columns=['CellID', 'FeatureIndex'])
df.to_csv(os.path.join(args.save_dir, 'cell_inds.txt'), index=False)