def get_parameters(self, parameter, features, actions, reinitialize):
""" Updates the parameters of the distribution
Args:
parameter (np.array): parameter of the distribution
features (np.array): observation features
actions (np.array): observation actions
reinitialize (bool): for bucketizer to be applied on new features
"""
if not self.feature_map.anchor_points_initialized:
self.feature_map.initialize_anchor_points(features, actions)
self.bins = self.feature_map.action_anchor_points
self.inds = np.digitize(actions, self.bins, right=True)
bucketized_actions = self.bins[self.inds]
fm_a = np.concatenate([self.indicator(bucketized_actions - b)[:, np.newaxis] for b in self.bins],
axis=1)
fm_c = self.feature_map.contextual_feature_map(features)
self.fm_c_shape = fm_c.shape[1]
self.representation = np.einsum('ij, ik -> ikj', fm_c, fm_a)
if reinitialize:
self.inds = np.digitize(actions, self.bins, right=True)
self.inds[self.inds == np.max(self.inds)] = np.max(self.inds) -1
bucketized_actions = self.bins[self.inds]
fm_a = np.concatenate([self.indicator(bucketized_actions - b)[:, np.newaxis] for b in self.bins],
axis=1)
fm_c = self.feature_map.contextual_feature_map(features)
self.representation = np.einsum('ij, ik -> ikj', fm_c, fm_a)
itcp, param = parameter[0], parameter[1:]
param = param.reshape(self.bins.shape[0], self.fm_c_shape)
return np.einsum('ikl, kl-> ik', self.representation, param) + itcp, self.inds
def assign_to_modal_uparams(this_uparam, modal_uparam):
try:
mid_pts = 0.5 * (modal_uparam[1:] + modal_uparam[:-1])
bins = np.concatenate(((-np.inf, ), mid_pts, (np.inf, )))
inds_in_modal = np.digitize(this_uparam, bins) - 1
numerical = True
except:
print('non-numerical parameter')
numerical = False
if numerical:
uinds = np.unique(inds_in_modal)
inds_in_this = np.zeros((0, ), dtype='int')
for uind in uinds:
candidates = np.where(inds_in_modal == uind)[0]
dist_from_modal = np.abs(this_uparam[candidates] -
modal_uparam[uind])
to_keep = candidates[np.argmin(dist_from_modal)]
inds_in_this = np.concatenate((inds_in_this, (to_keep, )))
inds_in_modal = inds_in_modal[inds_in_this]
bool_in_this = np.zeros((len(this_uparam), ), dtype='bool')
bool_in_modal = np.zeros((len(modal_uparam), ), dtype='bool')
bool_in_this[inds_in_this] = True
bool_in_modal[inds_in_modal] = True
else:
assert (np.all(this_uparam == modal_uparam))
bool_in_this, bool_in_modal = [
np.ones(this_uparam.shape, dtype='bool') for iparam in range(2)
]
return bool_in_this, bool_in_modal
def fit_gaussian(S, Cin, param_means, param_sigmas):
'''
generates predictions using parameters sampled form the distribution, then outputs
a gaussian fitted to the predictions
'''
output_samples = sample_output(sol, Cin, true_params, prior_sigmas) * 100
# generate approximate gaussian
bins = np.array(range(1000))
indeces = np.digitize(output_samples, bins)
count1 = np.bincount(indeces[:, 0], minlength=1001)[1:]
print(bins.shape)
print(count1.shape)
print(count1.shape)
print(count1)
count2 = np.bincount(indeces[:, 1], minlength=1001)[1:]
# fit gaussians
'''
popt1 = curve_fit(gaussian, bins, count1)
popt2 = curve_fit(gaussian, bins, count2)
print('popt1', popt1)
print('popt2', popt2)
'''
plt.plot(bins, count1)
plt.plot(bins, count2)
plt.show()
sys.exit()
def __init__(self,
counts,
lengths,
ploidy,
multiscale_factor=1,
constraint_lambdas=None,
constraint_params=None):
self.lengths = np.array(lengths)
self.lengths_lowres = decrease_lengths_res(lengths, multiscale_factor)
self.ploidy = ploidy
self.multiscale_factor = multiscale_factor
if constraint_lambdas is None:
self.lambdas = {}
else:
self.lambdas = constraint_lambdas
if constraint_params is None:
self.params = {}
else:
self.params = constraint_params
torm = find_beads_to_remove(counts=counts,
nbeads=self.lengths_lowres.sum() * ploidy)
self.torm_3d = np.repeat(torm.reshape(-1, 1), 3, axis=1)
self.row, self.col = _constraint_dis_indices(
counts=counts,
n=self.lengths_lowres.sum(),
lengths=self.lengths_lowres,
ploidy=ploidy)
# Calculating distances for neighbors, which are on the off diagonal
# line - i & j where j = i + 1
row_adj = ag_np.unique(self.row).astype(int)
row_adj = row_adj[ag_np.isin(row_adj + 1, self.col)]
# Remove if "neighbor" beads are actually on different chromosomes or
# homologs
self.row_adj = row_adj[ag_np.digitize(
row_adj,
np.tile(lengths, ploidy).cumsum()) == ag_np.digitize(
row_adj + 1,
np.tile(lengths, ploidy).cumsum())]
self.col_adj = self.row_adj + 1
self.check()
def resampling(w, rs):
"""
Stratified resampling with "nograd_primitive" to ensure autograd
takes no derivatives through it.
"""
N = w.shape[0]
bins = np.cumsum(w)
ind = np.arange(N)
u = (ind + rs.rand(N)) / N
return np.digitize(u, bins)
def sim_q(prop_params, model_params, y, smc_obj, rs, verbose=False):
"""
Simulates a single sample from the VSMC approximation.
Requires an SMC object with 2 member functions:
-- sim_prop(t, x_{t-1}, y, prop_params, model_params, rs)
-- log_weights(t, x_t, x_{t-1}, y, prop_params, model_params)
"""
# Extract constants
T = y.shape[0]
Dx = smc_obj.Dx
N = smc_obj.N
# Initialize SMC
X = np.zeros((N, T, Dx))
logW = np.zeros(N)
W = np.zeros((N, T))
ESS = np.zeros(T)
for t in range(T):
# Resampling
if t > 0:
ancestors = resampling(W[:, t - 1], rs)
X[:, :t, :] = X[ancestors, :t, :]
# Propagation
X[:, t, :] = smc_obj.sim_prop(t, X[:, t - 1, :], y, prop_params,
model_params, rs)
# Weighting
logW = smc_obj.log_weights(t, X[:, t, :], X[:, t - 1, :], y,
prop_params, model_params)
max_logW = np.max(logW)
W[:, t] = np.exp(logW - max_logW)
W[:, t] /= np.sum(W[:, t])
ESS[t] = 1. / np.sum(W[:, t]**2)
# Sample from the empirical approximation
bins = np.cumsum(W[:, -1])
u = rs.rand()
B = np.digitize(u, bins)
if verbose:
print('Mean ESS', np.mean(ESS) / N)
print('Min ESS', np.min(ESS))
return X[B, :, :]
def _setup(self):
""" Setup the experiments and creates the data
"""
# Actions
features, y = self.get_X_y_by_name()
potentials = self._get_potentials(y)
actions = self.rng.lognormal(mean=self.start_mu,
sigma=self.start_sigma,
size=potentials.shape[0])
rewards = self.get_rewards_from_actions(potentials, actions)
if self.discrete:
from scipy.stats import lognorm
rv = lognorm(s=self.start_sigma, scale=np.exp(self.start_mu))
quantiles = np.quantile(actions,
np.linspace(0, 1, self.discrete + 1))
action_anchors = np.pad(quantiles,
1,
'constant',
constant_values=(1e-7, np.inf))
bins = action_anchors[:-1]
inds = np.digitize(actions, bins, right=True)
inds_1 = inds - 1
inds_1[inds_1 == -1] = 0
pi_logging = rv.cdf(bins[inds]) - rv.cdf(bins[inds_1])
else:
pi_logging = Dataset.logging_policy(actions, self.start_mu,
self.start_sigma)
# Test train split
self.actions_train, self.actions_test, self.features_train, self.features_test, self.reward_train, \
self.reward_test, self.pi_0_train, self.pi_0_test, self.potentials_train, self.potentials_test, \
self.l_train, self.l_test = train_test_split(actions, features, rewards, pi_logging, potentials, y,
train_size=self.train_size, random_state=42)
self.actions_train, self.actions_valid, self.features_train, self.features_valid, self.reward_train, \
self.reward_valid, self.pi_0_train, self.pi_0_valid, self.potentials_train, self.potentials_valid, \
self.l_train, self.l_valid = train_test_split(self.actions_train, self.features_train, self.reward_train,
self.pi_0_train, self.potentials_train, self.l_train,
train_size=self.val_size, random_state=42)
min_max_scaler = MinMaxScaler(feature_range=(0, 1))
self.features_train = min_max_scaler.fit_transform(self.features_train)
self.features_valid = min_max_scaler.transform(self.features_valid)
self.features_test = min_max_scaler.transform(self.features_test)
self.baseline_reward_valid = np.mean(self.reward_valid)
self.baseline_reward_test = np.mean(self.reward_test)
def get_binned(fbin, tbin, lim):
#lim = 0.8
num = 10
bins = np.linspace(0,lim,num = num) + np.random.uniform(-0.05, 0.05)
ind = np.digitize(fbin, bins = bins)
y = []
se = []
x = []
for i in np.arange(num):
i += 1
rvals = tbin[ind == i]
if rvals.size:
x.append(bins[i-1])
y.append(np.mean(rvals))
se.append(np.std(rvals))
return (x, y, se)
