def auc_calc_beta(x_test, y_test, nn, N, perc, model, weightlist=None):
'''
Options for model "mc", "bbvi", "ensemble", "deterministic"
For BBVI, pass a list of weights.
'''
p = []
n_test = len(y_test)
if model != "deterministic":
if model == "mc":
p_allw, p_mean, entropymean = myentropy(nn, [nn.weights]*N, x_test.T, returnallp=True)
elif (model == "bbvi") and weightlist is not None:
p_allw, p_mean, entropymean = myentropy(nn, weightlist, x_test.T, returnallp=True)
elif model == "ensemble": # deterministic
p_allw, p_mean, entropymean = myentropy(nn, weightlist, x_test.T, returnallp=True)
#p_allw has dimension: NWeightSamples x NXData
idx = np.argsort(entropymean)
y_test = y_test[idx]
p_mean = p_mean[idx]
p_allw = p_allw[:, idx]
y_pred_retained_allw = p_allw[:, 0:int(perc*n_test)]
y_pred_retained = p_mean[0:int(perc*n_test)] # choosing samples with smallest entropy to evaluate
y_test_retained = y_test[0:int(perc*n_test)]
ypredmean = np.round(y_pred_retained)
ypred_allw = np.round(y_pred_retained_allw) #NW x NX
auc_allw = np.zeros(ypred_allw.shape[0])
for w in range(ypred_allw.shape[0]):
auc_allw[w] = np.count_nonzero(ypred_allw[w, :]==y_test_retained)/len(y_test_retained) * 100
return auc_allw
else:
auc = auc_calc_proba(x_test, y_test, nn, N, perc)
return auc #this only returns the mean accuracy
def load_and_pickle_binary_mnist():
N_data, train_images, train_labels, test_images, test_labels = load_mnist()
train_images = np.round(train_images)
test_images = np.round(test_images)
mnist_data = N_data, train_images, train_labels, test_images, test_labels
with open('mnist_binary_data.pkl', 'w') as f:
pickle.dump(mnist_data, f, 1)
def test_gp_missing_obs(bspline_data):
low, high = bspline_data['xlim']
num_bases = 5
bsplines_degree = 3
basis = BSplines(low, high, num_bases, bsplines_degree, boundaries='space')
n_clusters = 1
n_train = bspline_data['n_train']
truncated_time = bspline_data['truncated_time']
m = []
for i in range(n_clusters):
m.append(LinearWithBsplinesBasis(basis, no=i))
tr = []
tr.append((1.0, Treatment(2.0)))
gp = GP(m, linear_cov(basis), tr, ac_fn=None)
# modify dataset
gp.fit(make_missed_obs_samples(bspline_data['training2']),
options={'maxiter': 1})
print(gp.params)
assert np.round(gp.params['linear_with_bsplines_basis_mean_coef0'],
3).tolist() != [0.0] * num_bases
assert gp.params['treatment'].tolist() != [0.0]
assert np.round(gp.params['classes_prob_logit_F'], 0).tolist() == [1.0]
_test_gp_prediction(
gp, make_missed_obs_samples(bspline_data['testing1'][0:20]),
truncated_time)
def evaluate(self, X, Y, evaluation_type='regression'):
predicted_y = self.predict(X)
result = {}
result['error'] = numpy.mean(numpy.linalg.norm(predicted_y - Y,
axis=1))
if 'classification' in evaluation_type:
correct = 0
incorrect = 0
if self.last_outgoing_dimension == 1:
for py, y in zip(predicted_y, Y):
if (numpy.round(py) == numpy.round(y)):
correct += 1
else:
incorrect += 1
elif self.last_outgoing_dimension > 1:
for py, y in zip(predicted_y, Y):
if (numpy.argmax(py) == numpy.argmax(y)):
correct += 1
else:
incorrect += 1
result['accuracy'] = correct / (correct + incorrect)
return result
def HMMModelTest():
with np.errstate(under='ignore',
divide='raise',
over='raise',
invalid='raise'):
T = 10
D_latent = 5
D_obs = 4
meas = 2
size = 3
alpha_0 = np.random.random(D_latent) + 1
alpha = np.random.random((D_latent, D_latent)) + 1
L = np.random.random((D_latent, D_obs)) + 1
params = {'alpha_0': alpha_0, 'alpha': alpha, 'L': L}
hmm = HMMModel(**params)
_, ys = HMMModel.generate(T=T,
latentSize=D_latent,
obsSize=D_obs,
measurements=meas,
size=size)
hmm.fit(ys=ys,
method='gibbs',
nIters=500,
burnIn=200,
skip=2,
verbose=True)
marginal = hmm.state.ilog_marginal(ys)
print('\nParams')
for p in hmm.state.params:
print(np.round(p, decimals=3))
print()
print('MARGNIAL', marginal)
hmm.fit(ys=ys,
method='EM',
nIters=1000,
monitorMarginal=10,
verbose=False)
marginal = hmm.state.ilog_marginal(ys)
print('\nParams')
for p in hmm.state.params:
print(np.round(p, decimals=3))
print()
print('MARGNIAL', marginal)
hmm.fit(ys=ys, method='cavi', maxIters=1000, verbose=False)
elbo = hmm.state.iELBO(ys)
print('\nPrior mean field params')
for p in hmm.state.prior.mf_params:
print(np.round(p, decimals=3))
print()
print('ELBO', elbo)
def gradient_descent(g, w, x_train, x_val, alpha, max_its, batch_size,
**kwargs):
verbose = True
if 'verbose' in kwargs:
verbose = kwargs['verbose']
# flatten the input function, create gradient based on flat function
g_flat, unflatten, w = flatten_func(g, w)
grad = value_and_grad(g_flat)
# record history
num_train = x_train.shape[1]
num_val = x_val.shape[1]
w_hist = [unflatten(w)]
train_hist = [g_flat(w, x_train, np.arange(num_train))]
val_hist = [g_flat(w, x_val, np.arange(num_val))]
# how many mini-batches equal the entire dataset?
num_batches = int(np.ceil(np.divide(num_train, batch_size)))
# over the line
for k in range(max_its):
# loop over each minibatch
start = timer()
train_cost = 0
for b in range(num_batches):
# collect indices of current mini-batch
batch_inds = np.arange(b * batch_size,
min((b + 1) * batch_size, num_train))
# plug in value into func and derivative
cost_eval, grad_eval = grad(w, x_train, batch_inds)
grad_eval.shape = np.shape(w)
# take descent step with momentum
w = w - alpha * grad_eval
end = timer()
# update training and validation cost
train_cost = g_flat(w, x_train, np.arange(num_train))
val_cost = g_flat(w, x_val, np.arange(num_val))
# record weight update, train and val costs
w_hist.append(unflatten(w))
train_hist.append(train_cost)
val_hist.append(val_cost)
if verbose == True:
print('step ' + str(k + 1) + ' done in ' +
str(np.round(end - start, 1)) + ' secs, train cost = ' +
str(np.round(train_hist[-1][0], 4)) + ', val cost = ' +
str(np.round(val_hist[-1][0], 4)))
if verbose == True:
print('finished all ' + str(max_its) + ' steps')
#time.sleep(1.5)
#clear_output()
return w_hist, train_hist, val_hist
def callback(V_flat):
loss_val = loss(V_flat)
V = V_flat.reshape((N, d))
reg_val = ortho_reg_fn(V, lambda_param)
loss_no_reg = loss_val - reg_val
pi = -loss_no_reg
print("PI = " + str(np.round(pi, 4)) + " bits, reg = " +
str(np.round(reg_val, 4)))
def animate(k):
# clear panels
ax.cla()
lam = lams[k]
# print rendering update
if np.mod(k + 1, 25) == 0:
print('rendering animation frame ' + str(k + 1) + ' of ' +
str(num_frames))
if k == num_frames - 1:
print('animation rendering complete!')
time.sleep(1.5)
clear_output()
# run optimization
if algo == 'gradient_descent':
weight_history, cost_history = self.gradient_descent(
g, w, self.x, self.y, lam, alpha_choice, max_its,
batch_size)
if algo == 'RMSprop':
weight_history, cost_history = self.RMSprop(
g, w, self.x, self.y, lam, alpha_choice, max_its,
batch_size)
# choose set of weights to plot based on lowest cost val
ind = np.argmin(cost_history)
# classification? then base on accuracy
if 'counter' in kwargs:
# create counting cost history as well
counts = [
counter(v, self.x, self.y, lam) for v in weight_history
]
if k == 0:
ind = np.argmin(counts)
count = counts[ind]
acc = 1 - count / self.y.size
acc = np.round(acc, 2)
# save lowest misclass weights
w_best = weight_history[ind][1:]
# plot
ax.axhline(c='k', zorder=2)
# make bar plot
ax.bar(np.arange(0, len(w_best)), w_best, color='k', alpha=0.5)
# dress panel
title1 = r'$\lambda = ' + str(np.round(lam, 2)) + '$'
costval = cost_history[ind][0]
title2 = ', cost val = ' + str(np.round(costval, 2))
if 'counter' in kwargs:
title2 = ', accuracy = ' + str(acc)
title = title1 + title2
ax.set_title(title)
ax.set_xlabel('learned weights')
return artist,
def gradient_descent(g, w, a_train, s_train, alpha, max_its, verbose):
'''
A basic gradient descent module (full batch) for system identification training.
Inputs to gradient_descent function:
g - function to minimize
w - initial weights
a_train - training action sequence
s_train - training state sequence
alpha - steplength / learning rate
max_its - number of iterations to perform
verbose - print out update each step if verbose = True
'''
# flatten the input function, create gradient based on flat function
g_flat, unflatten, w = flatten_func(g, w)
grad = value_and_grad(g_flat)
# record history
# num_val = y_val.size
w_hist = [unflatten(w)]
train_hist = [g_flat(w, a_train, s_train)]
# over the line
alpha_choice = 0
for k in range(1, max_its + 1):
# take a single descent step
start = timer()
# plug in value into func and derivative
cost_eval, grad_eval = grad(w, a_train, s_train)
grad_eval.shape = np.shape(w)
# take descent step with momentum
w = w - alpha * grad_eval
end = timer()
# update training and validation cost
train_cost = g_flat(w, a_train, s_train)
val_cost = np.nan
# record weight update, train cost
w_hist.append(unflatten(w))
train_hist.append(train_cost)
if verbose == True:
print('step ' + str(k + 1) + ' done in ' +
str(np.round(end - start, 1)) + ' secs, train cost = ' +
str(np.round(train_hist[-1], 4)[0]))
if verbose == True:
print('finished all ' + str(max_its) + ' steps')
return w_hist, train_hist
def save_rotation_lookup(array_size, n_theta, dest_folder=None):
image_center = [np.floor(x / 2) for x in array_size]
coord0 = np.arange(array_size[0])
coord1 = np.arange(array_size[1])
coord2 = np.arange(array_size[2])
coord2_vec = np.tile(coord2, array_size[1])
coord1_vec = np.tile(coord1, array_size[2])
coord1_vec = np.reshape(coord1_vec, [array_size[1], array_size[2]])
coord1_vec = np.reshape(np.transpose(coord1_vec), [-1])
coord0_vec = np.tile(coord0, [array_size[1] * array_size[2]])
coord0_vec = np.reshape(coord0_vec, [array_size[1] * array_size[2], array_size[0]])
coord0_vec = np.reshape(np.transpose(coord0_vec), [-1])
# move origin to image center
coord1_vec = coord1_vec - image_center[1]
coord2_vec = coord2_vec - image_center[2]
# create matrix of coordinates
coord_new = np.stack([coord1_vec, coord2_vec]).astype(np.float32)
# create rotation matrix
theta_ls = np.linspace(0, 2 * np.pi, n_theta)
coord_old_ls = []
for theta in theta_ls:
m_rot = np.array([[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)]])
coord_old = np.matmul(m_rot, coord_new)
coord1_old = np.round(coord_old[0, :] + image_center[1]).astype(np.int)
coord2_old = np.round(coord_old[1, :] + image_center[2]).astype(np.int)
# clip coordinates
coord1_old = np.clip(coord1_old, 0, array_size[1]-1)
coord2_old = np.clip(coord2_old, 0, array_size[2]-1)
coord_old = np.stack([coord1_old, coord2_old], axis=1)
coord_old_ls.append(coord_old)
if dest_folder is None:
dest_folder = 'arrsize_{}_{}_{}_ntheta_{}'.format(array_size[0], array_size[1], array_size[2], n_theta)
if not os.path.exists(dest_folder):
os.mkdir(dest_folder)
for i, arr in enumerate(coord_old_ls):
np.save(os.path.join(dest_folder, '{:04}'.format(i)), arr)
coord1_vec = coord1_vec + image_center[1]
coord1_vec = np.tile(coord1_vec, array_size[0])
coord2_vec = coord2_vec + image_center[2]
coord2_vec = np.tile(coord2_vec, array_size[0])
for i, coord in enumerate([coord0_vec, coord1_vec, coord2_vec]):
np.save(os.path.join(dest_folder, 'coord{}_vec'.format(i)), coord)
return coord_old_ls
def test_mean_linear(linear_data):
m = Linear(1)
mp = m(params_only=True)
mp = m(mp, linear_data['training2'], params_only=True)
print(mp)
coef_ = np.round(mp['linear_mean_coef'], 2).tolist()
assert coef_[0] == 0.45
assert coef_[1] == -0.63 # bias affected the treatment
yhat = m(mp, np.array([1, 2, 3]))
assert np.round(yhat, 8).tolist() == [-0.17919259, 0.266663, 0.7125186]
def get_training_data(
data): #get all data input and output for training from the csv file
n = np.size(data, 0)
n = 2 * n / 3
n = int(np.round(n))
train_data = data[:n, :]
return train_data
def leapfrog(q, p, dVdq, path_len, step_size):
"""Leapfrog integrator for Hamiltonian Monte Carlo.
Parameters
----------
q : np.floatX
Initial position
p : np.floatX
Initial momentum
dVdq : callable
Gradient of the velocity
path_len : float
How long to integrate for
step_size : float
How long each integration step should be
Returns
-------
q, p : np.floatX, np.floatX
New position and momentum
"""
q, p = np.copy(q), np.copy(p)
p -= step_size * dVdq(q) / 2 # half step
for _ in np.arange(np.round(path_len / step_size) - 1):
q += step_size * p # whole step
p -= step_size * dVdq(q) # whole step
q += step_size * p # whole step
p -= step_size * dVdq(q) / 2 # half step
# momentum flip at end
return q, -p
def Accuracy(self, y, y_hat):
y_hat = np.round(y_hat)
temp = 0
for i in range(len(y)):
if y_hat[i] == y[i]:
temp += 1
return temp * 100 / (len(y))
def auc_calc(x_test, y_test, nn, N, perc, model, weightlist=None):
'''
Options for model "mc", "bbvi", "ensemble", "deterministic"
For BBVI, pass a list of weights.
'''
p = []
n_test = len(y_test)
if model != "deterministic":
if model == "mc":
p_mean, entropymean = myentropy(nn, [nn.weights]*N, x_test.T)
elif (model == "bbvi") and weightlist is not None:
p_mean, entropymean = myentropy(nn, weightlist, x_test.T)
elif model == "ensemble": # deterministic
nn_weights = []
for nn_here in nn:
nn_weights.append(nn_here.weights)
p_mean, entropymean = myentropy(nn, nn_weights, x_test.T)
idx = np.argsort(entropymean)
y_test = y_test[idx]
p_mean = p_mean[idx]
y_pred_retained = p_mean[0:int(perc*n_test)] # choosing samples with smallest entropy to evaluate
y_test_retained = y_test[0:int(perc*n_test)]
predict_proba = np.round(y_pred_retained)
auc = len(y_test_retained[predict_proba==y_test_retained]) / len(y_pred_retained) * 100
else:
auc = auc_calc_proba(x_test, y_test, nn, N, perc)
return auc
def p1_optimal(
p1=100,
p2=100,
alpha=[1, 0, 0., 0., 1., 0., 0., 0., 1., 0., 0., 0., 1, 0., 0., 0.],
n_clients_per_class=[50, 20, 10, 5]):
'''
Inputs :
------------
Output :
p1 : the best prize found for the first item
'''
bounds_p1 = so.Bounds(0, np.inf) # p1 prize of first item positive
x0 = [p1]
gradient_obj_fun = grad(obj_fun)
res = so.minimize(obj_fun,
x0=x0,
args=(p2, alpha, n_clients_per_class),
jac=gradient_obj_fun,
bounds=bounds_p1)
x_sol = res.x
return np.round(x_sol[0])
def get_test_data_result(
data): #get all data output for testing from the csv file
n = np.size(data, 0)
n = 2 * n / 3
n = int(np.round(n))
y = data[n:, -1]
return y
def classification_data(seed=0):
"""
Load 2D data. 2 Classes. Class labels generated from a 2-2-1 network.
:param seed: random number seed
:return:
"""
npr.seed(seed)
data = np.load("./data/2D_toy_data_linear.npz")
x = data['x']
y = data['y']
ids = np.arange(x.shape[0])
npr.shuffle(ids)
# 75/25 split
num_train = int(np.round(0.01 * x.shape[0]))
x_train = x[ids[:num_train]]
y_train = y[ids[:num_train]]
x_test = x[ids[num_train:]]
y_test = y[ids[num_train:]]
mu = np.mean(x_train, axis=0)
std = np.std(x_train, axis=0)
x_train = (x_train - mu) / std
x_test = (x_test - mu) / std
train_stats = dict()
train_stats['mu'] = mu
train_stats['sigma'] = std
return x_train, y_train, x_test, y_test, train_stats
def __init__(self,
Wdict,
rate_f=calnet.utils.f_miller_troyer,
rate_fprime=calnet.utils.fprime_m_miller_troyer,
u_fn=calnet.utils.u_fn_WW):
# Wmx,Wmy,Wsx,Wsy,s02,k,kappa,XX,XXp,Eta,Xi
for key in Wdict:
setattr(self, key, Wdict[key])
self.rate_f = rate_f
self.rate_fprime = rate_fprime
self.u_fn = u_fn
self.nP = self.Wmx.shape[0]
self.nQ = self.Wmy.shape[0]
self.nS = int(np.round(self.Eta.shape[1] / self.nQ))
self.nT = 1
self.nN = self.Eta.shape[0]
wws = ['WWmx', 'WWmy', 'WWsx', 'WWsy']
ws = ['Wmx', 'Wmy', 'Wsx', 'Wsy']
for w, ww in zip(ws, wws):
W = getattr(self, w)
WW = calnet.utils.gen_Weight_k_kappa(W, self.k, self.kappa)
setattr(self, ww, WW)
self.YY = self.compute_f_(self.Eta, self.Xi, self.s02)
self.resEta = self.Eta - self.u_fn_m(self.XX, self.YY)
self.resXi = self.Xi - self.u_fn_s(self.XX, self.YY)
def get_metrics(y_real, y_pred):
for how_many_last in [100, 20, 10]:
real = y_real[-how_many_last:]
pred = y_pred[-how_many_last:]
print('On last {} steps'.format(how_many_last))
print('test acc', np.mean(real == np.round(pred)))
print('test auc', roc_auc_score(real, pred))
def step_generator(t, **kwargs):
'''
Makes a random piecewise constant step frequency input. Note
that the steps always take on integer values.
'''
# lets make random piecewise frequency
num_steps = 2
if 'num_steps' in kwargs:
num_steps = kwargs['num_steps'] - 1
# pick num_pieces random locations for step ledges
r = np.random.permutation(len(t))[:num_steps]
r = np.sort(r)
# generate random level per step
levels = np.round(5 * np.random.rand(num_steps + 1)) + 1
f = np.ones((t.size, 1))
# set each chunk to appropriate level
f[:r[0]] *= levels[0] # set first chunk to appropriate level
for n in range(1, len(r) - 1):
f[r[n - 1]:r[n]] *= levels[n]
f[r[-1]:] *= levels[-1]
return f
def realign_image(arr, shift):
"""
Translate and rotate image via Fourier
Parameters
----------
arr : ndarray
Image array.
shift: tuple
Mininum and maximum values to rescale data.
angle: float, optional
Mininum and maximum values to rescale data.
Returns
-------
ndarray
Output array.
"""
# if both shifts are integers, do circular shift; otherwise perform Fourier shift.
if np.count_nonzero(np.abs(np.array(shift) - np.round(shift)) < 0.01) == 2:
temp = np.roll(arr, int(shift[0]), axis=0)
temp = np.roll(temp, int(shift[1]), axis=1)
temp = temp.astype('float32')
else:
temp = fourier_shift(np.fft.fftn(arr), shift)
temp = np.fft.ifftn(temp)
temp = np.abs(temp).astype('float32')
return temp
def test_sgp(bspline_data):
low, high = bspline_data['xlim']
num_bases = 5
bsplines_degree = 3
basis = BSplines(low, high, num_bases, bsplines_degree, boundaries='space')
n_clusters = 1
random_basis = np.random.multivariate_normal(np.zeros(num_bases),
0.1 * np.eye(num_bases),
n_clusters)
n_train = bspline_data['n_train']
truncated_time = bspline_data['truncated_time']
m = []
for i in range(n_clusters):
m.append(LinearWithBsplinesBasis(basis, no=i, init=random_basis[i]))
tr = []
tr.append((1.0, Treatment(2.0)))
mcgp = GP(m, linear_cov(basis), tr, ac_fn=None)
mcgp.fit(bspline_data['training2'], options={'maxiter': 1})
print(mcgp.params)
assert mcgp.params['treatment'].tolist() != [0.0]
assert np.round(mcgp.params['classes_prob_logit_F'], 0).tolist() == [1.0]
_test_gp_prediction(mcgp, bspline_data['testing1'][0:20], truncated_time)
def gradient_descent(g, alpha, max_its, w, num_pts, train_portion,**kwargs):
# flatten the input function, create gradient based on flat function
g_flat, unflatten, w = flatten_func(g, w)
grad = value_and_grad(g_flat)
# containers for histories
weight_hist = []
train_ind_hist = []
test_ind_hist = []
# store first weights
weight_hist.append(unflatten(w))
# pick random proportion of training indecies
train_num = int(np.round(train_portion*num_pts))
inds = np.random.permutation(num_pts)
train_inds = inds[:train_num]
test_inds = inds[train_num:]
# record train / test inds
train_ind_hist.append(train_inds)
test_ind_hist.append(test_inds)
# over the line
for k in range(max_its):
# plug in value into func and derivative
cost_eval,grad_eval = grad(w,train_inds)
grad_eval.shape = np.shape(w)
# take descent step with momentum
w = w - alpha*grad_eval
# record weight update
weight_hist.append(unflatten(w))
#### pick new train / test split ####
# pick random proportion of training indecies
train_num = int(np.round(train_portion*num_pts))
inds = np.random.permutation(num_pts)
train_inds = inds[:train_num]
test_inds = inds[train_num:]
# record train / test inds
train_ind_hist.append(train_inds)
test_ind_hist.append(test_inds)
return weight_hist,train_ind_hist,test_ind_hist
def __init__(self, frame, sky_coord, observations):
"""Source intialized with a single pixel
Parameters
----------
frame: `~scarlet.Frame`
The frame of the model
sky_coord: tuple
Center of the source
observations: instance or list of `~scarlet.Observation`
Observation(s) to initialize this source
"""
C, Ny, Nx = frame.shape
self.center = np.array(frame.get_pixel(sky_coord), dtype="float")
# initialize SED from sky_coord
try:
iter(observations)
except TypeError:
observations = [observations]
# determine initial SED from peak position
# SED in the frame for source detection
seds = []
for obs in observations:
_sed = get_psf_sed(sky_coord, obs, frame)
seds.append(_sed)
sed = np.concatenate(seds).reshape(-1)
if np.any(sed <= 0):
# If the flux in all channels is <=0,
# the new sed will be filled with NaN values,
# which will cause the code to crash later
msg = "Zero or negative SED {} at y={}, x={}".format(
sed, *sky_coord)
if np.all(sed <= 0):
logger.warning(msg)
else:
logger.info(msg)
# set up parameters
sed = Parameter(
sed,
name="sed",
step=partial(relative_step, factor=1e-2),
constraint=PositivityConstraint(),
)
center = Parameter(self.center, name="center", step=1e-1)
# define bbox
pixel_center = tuple(np.round(center).astype("int"))
front, back = 0, C
bottom = pixel_center[0] - frame.psf.shape[1] // 2
top = pixel_center[0] + frame.psf.shape[1] // 2
left = pixel_center[1] - frame.psf.shape[2] // 2
right = pixel_center[1] + frame.psf.shape[2] // 2
bbox = Box.from_bounds((front, back), (bottom, top), (left, right))
super().__init__(frame, sed, center, self._psf_wrapper, bbox=bbox)
def _test_gp_prediction(m, data, truncated_time, exclude_ac=[]):
_samples = make_predict_samples(data, None, None, truncated_time)
s = 0.0
for (y, x), (_y, _x, _x_star) in zip(data, _samples):
yhat, cov_hat = m.predict(_x_star, _y, _x, exclude_ac)
_t_star, _rx_star = _x_star
idx = _t_star > truncated_time
s += np.sum((yhat - y)[idx]**2) / len(y[idx])
p_a, p_mix = m.class_posterior(_y, _x, exclude_ac)
if exclude_ac: assert len(p_a) == 2 - len(exclude_ac)
assert np.round(np.sum(p_a), 0) == 1.0
assert np.round(np.sum(p_mix), 0) == 1.0
mse = s / len(data)
print(mse)
assert True
def PrintPerf(Params, iter, _):
if iter == 0:
print(" Epoch | Train cost ")
if iter % 5 == 0:
Cost = ObjectiveFunWrap(Params, iter)
Gradient = flatten(ObjectiveGrad(Params, iter))
print(
str(iter) + ' ' + str(np.round(Cost, 6)) + ' ' +
str(np.square(Gradient[0]).sum()))
def MovingWinFeats(x, xLen, fs, winLen, winDisp, featFn):
y = np.zeros((1, np.floor(((xLen - winLen * fs) / (winDisp * fs)) + 1)))
y[0] = featFn(x[0:np.round(winLen * fs)])
a = np.arange(2, np.floor(((xLen - winLen * fs) / (winDisp * fs)) + 1))
for i in a:
y[i] = featFn(x[np.ceil(winDisp * fs *
(i - 1)):np.ceil(winDisp * fs * (i - 1) +
winLen * fs)])
def draw_fit_trainval(self,ax,run,plot_fit):
# set plotting limits
xmax = np.max(copy.deepcopy(self.x))
xmin = np.min(copy.deepcopy(self.x))
xgap = (xmax - xmin)*0.1
xmin -= xgap
xmax += xgap
ymax = np.max(copy.deepcopy(self.y))
ymin = np.min(copy.deepcopy(self.y))
ygap = (ymax - ymin)*0.3
ymin -= ygap
ymax += ygap
####### plot total model on original dataset #######
# scatter original data - training and validation sets
train_inds = run.train_inds
valid_inds = run.val_inds
ax.scatter(self.x[:,train_inds],self.y[:,train_inds],color = self.colors[1],s = 40,edgecolor = 'k',linewidth = 0.9)
ax.scatter(self.x[:,valid_inds],self.y[:,valid_inds],color = self.colors[0],s = 40,edgecolor = 'k',linewidth = 0.9)
if plot_fit == True:
# plot fit on residual
s = np.linspace(xmin,xmax,2000)[np.newaxis,:]
# plot total fit
t = 0
# get current run
cost = run.cost
model = run.model
feat = run.feature_transforms
normalizer = run.normalizer
cost_history = run.train_cost_histories[0]
weight_history = run.weight_histories[0]
# get best weights
win = np.argmin(cost_history)
w = weight_history[win]
t = model(normalizer(s),w)
ax.plot(s.T,t.T,linewidth = 4,c = 'k')
ax.plot(s.T,t.T,linewidth = 2,c = 'r')
lam = run.lam
ax.set_title( 'lam = ' + str(np.round(lam,2)) + ' and fit to original',fontsize = 14)
if plot_fit == False:
ax.set_title('test',fontsize = 14,color = 'w')
### clean up panels ###
ax.set_xlim([xmin,xmax])
ax.set_ylim([ymin,ymax])
# label axes
ax.set_xlabel(r'$x$', fontsize = 14)
ax.set_ylabel(r'$y$', rotation = 0,fontsize = 14,labelpad = 15)
def split_data(self, folds):
# split data into k equal (as possible) sized sets
L = np.size(self.y)
order = np.random.permutation(L)
c = np.ones((L, 1))
L = int(np.round((1 / folds) * L))
for s in np.arange(0, folds - 2):
c[order[s * L:(s + 1) * L]] = s + 2
c[order[(folds - 1) * L:]] = folds
return c
def test_pow():
fun = lambda x, y : to_scalar(x ** y)
d_fun_0 = lambda x, y : to_scalar(grad(fun, 0)(x, y))
d_fun_1 = lambda x, y : to_scalar(grad(fun, 1)(x, y))
make_positive = lambda x : np.abs(x) + 1.1 # Numeric derivatives fail near zero
for arg1, arg2 in arg_pairs():
arg1 = make_positive(arg1)
arg2 = np.round(arg2)
check_grads(fun, arg1, arg2)
check_grads(d_fun_0, arg1, arg2)
check_grads(d_fun_1, arg1, arg2)
def sample_invwishart(S, nu):
n = S.shape[0]
chol = np.linalg.cholesky(S)
if (nu <= 81 + n) and (nu == np.round(nu)):
x = npr.randn(nu, n)
else:
x = np.diag(np.sqrt(np.atleast_1d(chi2.rvs(nu - np.arange(n)))))
x[np.triu_indices_from(x, 1)] = npr.randn(n*(n-1)//2)
R = np.linalg.qr(x, 'r')
T = solve_triangular(R.T, chol.T, lower=True).T
return np.dot(T, T.T)
def toArray(dic,h,w):
"""
Convert dicts of arrays to arrays for the visualize function.
"""
outList = []
for k in dic:
if k != -1:
outList.append(dic[k].tolist())
outC = np.round(np.array(outList),2)
outC = outC.T
out = np.reshape(outC,(h,w))
return out
def __init__(self, band,
filename = None,
fits_file_template = None,
timg = None,
exposure_num = 0,
calib = None,
gain = None,
darkvar = None,
sky = None,
frame = None,
fits_table = None):
self.band = band
if fits_file_template:
self.band_file = fits_file_template%band
self.img = fitsio.FITS(self.band_file)[exposure_num].read()
header = fitsio.read_header(self.band_file, ext=exposure_num)
elif filename is not None:
self.band_file = filename
self.img = fitsio.FITS(self.band_file)[exposure_num].read()
header = fitsio.read_header(self.band_file, ext=exposure_num)
elif timg:
self.band_file = None
self.img = timg[0].getImage()
header = timg[1]['hdr']
self.timg = timg[0]
self.invvar = self.timg.getInvvar()
else:
pass
self.header = header
self.frame = frame
self.fits_table = fits_table
# Compute the number of electrons, resource:
# http://data.sdss3.org/datamodel/files/BOSS_PHOTOOBJ/frames/RERUN/RUN/CAMCOL/frame.html
# (Neither of these look like integers)
if fits_file_template or filename:
self.dn = self.img / header["CALIB"] + header["SKY"]
self.nelec = np.round(self.dn * header["GAIN"])
else:
# TODO(awu): what are CALIB and GAIN?
self.dn = self.img / calib + sky #timg[0].getSky().val
self.nelec = np.round(self.dn * gain)
# make nelec immutable - it is constant data!!
self.nelec.flags.writeable = False
self.shape = self.nelec.shape
self.pixel_grid = self.make_pixel_grid() # keep pixel grid around
# reference points
# TODO: Does CRPIX1 refer to the first axis of self.img ??
self.rho_n = np.array([header['CRPIX1'], header['CRPIX2']]) - 1 # PIXEL REFERENCE POINT (fits stores it 1-based indexing)
self.phi_n = np.array([header['CRVAL1'], header['CRVAL2']]) # EQUA REFERENCE POINT
self.Ups_n = np.array([[header['CD1_1'], header['CD1_2']], # MATRIX takes you into EQUA TANGENT PLANE
[header['CD2_1'], header['CD2_2']]])
self.Ups_n_inv = np.linalg.inv(self.Ups_n)
#astrometry wcs object for "exact" x,y to equa ra,dec conversion
import astropy.wcs as wcs
self.wcs = wcs.WCS(self.header)
self.use_wcs = False
# set image specific KAPPA and epsilon
if fits_file_template:
self.kappa = header['GAIN'] # TODO is this right??
self.epsilon = header['SKY'] * self.kappa # background rate
self.epsilon0 = self.epsilon # background rate copy (for debuggin)
self.darkvar = header['DARKVAR'] # also eventually contributes to mean?
self.calib = header['CALIB'] # dn = nmaggies / calib, calib is NMGY
else:
self.kappa = gain
self.epsilon = timg[0].sky.val * self.kappa
self.epsilon0 = self.epsilon
self.darkvar = darkvar
self.calib = calib
# point spread function
if fits_file_template:
psfvec = [header['PSF_P%d'%i] for i in range(18)]
else:
psfvec = [psf for psf in timg[0].getPsf()]
self.weights = np.array(psfvec[0:3])
self.means = np.array(psfvec[3:9]).reshape(3, 2) # one comp mean per row
covars = np.array(psfvec[9:]).reshape(3, 3) # [var_k(x), var_k(y), cov_k(x,y)] per row
self.covars = np.zeros((3, 2, 2))
self.invcovars = np.zeros((3, 2, 2))
self.logdets = np.zeros(3)
for i in range(3):
self.covars[i,:,:] = np.array([[ covars[i,0], covars[i,2]],
[ covars[i,2], covars[i,1]]])
# cache inverse covariance
self.invcovars[i,:,:] = np.linalg.inv(self.covars[i,:,:])
# cache log determinant
sign, logdet = np.linalg.slogdet(self.covars[i,:,:])
self.logdets[i] = logdet
self.psf_mog = MixtureOfGaussians(means = self.means, covs = self.covars, pis = self.weights)
# for a point source in this image, calculate the radius such that
# at least 99% of photons from that source will fall within
ERROR = 0.001
self.R = calc_bounding_radius(self.weights,
self.means,
self.covars,
ERROR)
def test_poisson_logpmf_broadcast(): combo_check(stats.poisson.logpmf, [1])([np.round(R(4, 3)**2)], [R(4, 1)**2 + 1.1]) def test_poisson_pmf_broadcast(): combo_check(stats.poisson.pmf, [1])([np.round(R(4, 3)**2)], [R(4, 1)**2 + 1.1])
def test_poisson_logpmf(): combo_check(stats.poisson.logpmf, [1])([np.round(R(4)**2)], [R(4)**2 + 1.1]) def test_poisson_pmf(): combo_check(stats.poisson.pmf, [1])([np.round(R(4)**2)], [R(4)**2 + 1.1])
def test_poisson_pmf_broadcast(): combo_check(stats.poisson.pmf, [1])([np.round(R(4, 3)**2)], [R(4, 1)**2 + 1.1])
def test_t_pdf(): combo_check(stats.t.pdf, [0,1,2,3])([R(4)], [R(4)**2 + 2.1], [R(4)], [R(4)**2 + 2.1])
