def predict(self, X):
"""
This method makes predictions on test data 'X'.
Parameters
----------
X : numpy array
N x M numpy array; N = number of data points; M = number of
features.
"""
L = len(self.wts)
Z = arr(concat((np.ones((mat(X).shape[0], 1)), mat(X)),
axis=1)) # initialize to input features + constant
for l in range(L - 2):
Z = arr(
mat(Z) *
mat(self.wts[l]).T) # compute linear response of next layer
Z = arr(
concat((np.ones((mat(Z).shape[0], 1)), mat(self.sig(Z))),
axis=1)) # initialize to input features + constant
Z = arr(mat(Z) *
mat(self.wts[L - 1].T)) # compute output layer linear response
return self.sig_0(Z) # output layer activation function
def split(self, X, y, groups=None):
"""Generate indices to split data into training and test set.
Parameters
----------
X : array-like of shape (n_samples, n_features)
Training data, where n_samples is the number of samples
and n_features is the number of features.
y : array-like of shape (n_samples,), default=None
The target variable for supervised learning problems.
groups : array-like of shape (n_samples,)
Group labels for the samples used while splitting the dataset into
train/test set.
Yields
------
train : ndarray
The training set indices for that split.
test : ndarray
The testing set indices for that split.
"""
X, y, groups = indexable(X, y, groups)
indices = np.arange(_num_samples(X))
train_targets = {t: [] for t in np.unique(y)}
test_targets = {t: [] for t in np.unique(y)}
for target in targets.keys():
for test_i in self._iter_test_masks(X, y, groups, target):
train_targets[target].append(indices[np.logical_not(test_i)])
test_targets[target].append(indices[test_i])
train_index = np.concat(tuple(map(np.array, train_targets.values())))
test_index = np.concat(tuple(map(np.array, test_targets.values())))
yield train_index, test_index
def tile_sample(image, tile_shape):
height, width, n_channels = image.shape
hangover = width % tile_shape[1]
if hangover != 0:
pad_amount = tile_shape[1] - hangover
pad_shape = (height, pad_amount)
padding = np.zeros(pad_shape)
image = np.concat([image, padding], axis=2)
hangover = height % tile_shape[0]
if hangover != 0:
pad_amount = tile_shape[0] - hangover
pad_shape = list(image.shape)
pad_shape[1] = pad_amount
padding = np.zeros(pad_shape)
image = np.concat([image, padding], axis=1)
pad_height = tile_shape[0] - height % tile_shape[0]
pad_width = tile_shape[1] - width % tile_shape[1]
image = np.pad(image, ((0, pad_height), (0, pad_width), (0, 0)),
'constant')
H = int(height / tile_shape[0])
W = int(width / tile_shape[1])
slices = np.split(image, W, axis=1)
new_shape = (H, *tile_shape, n_channels)
slices = [np.reshape(s, new_shape) for s in slices]
new_images = np.concatenate(slices, axis=1)
new_images = new_images.reshape(H * W, *tile_shape, n_channels)
return new_images
def predict_data(model, data, steps):
it = iter(data)
x_test, y_test = zip(*[next(it) for _ in range(steps)])
yh_test = [model.predict(x) for x in x_test]
x_test = [np.concat(x) for x in zip(*x_test)]
y_test = np.concat(y_test).squeeze()
yh_test = np.concat(yh_test).squeeze()
return x_test, y_test, yh_test
def add_shape(self,
vertices,
normals,
triangles,
color,
atoms=None,
description=None):
"""Add shape to drawing
Parameters
----------
vertices : :py:class:`numpy.array` of coordinates
normals : :py:class:`numpy.array` of normals, one per vertex
triangles : :py:class:`numpy.array` of vertex indices, multiple of 3
color : either a single 4 element uint8 :py:class:`numpy.array`;
or an array of those values, one per vertex
atoms : a sequence of :py:class:`~chimerax.atomic.Atom`s
or an :py:class:`~chimerax.atomic.Atoms` collection.
description : a string describing the shape
The vertices, normals, and triangles can be custom or the results
from one of the :py:mod:`~chimerax.surface`'s geometry functions.
If the description is not given, it defaults to a list of the atoms.
"""
# extend drawing's vertices, normals, vertex_colors, and triangles
# atoms is a molarray.Atoms collection
# description is what shows up when hovered over
asarray = numpy.asarray
concat = numpy.concatenate
if color.ndim == 1 or color.shape[0] == 1:
colors = numpy.empty((vertices.shape[0], 4), dtype=numpy.uint8)
colors[:] = color
else:
colors = color.asarray(color, dtype=numpy.uint8)
assert colors.shape[1] == 4 and colors.shape[0] == vertices.shape[0]
if self.vertices is None:
if atoms is not None:
self._add_handler_if_needed()
self.set_geometry(asarray(vertices, dtype=numpy.float32),
asarray(normals, dtype=numpy.float32),
asarray(triangles, dtype=numpy.int32))
self.vertex_colors = colors
s = _AtomicShape(range(0, self.triangles.shape[0]), description,
atoms)
self._shapes.append(s)
return
offset = self.vertices.shape[0]
start = self.triangles.shape[0]
new_vertex_colors = concat((self.vertex_colors, colors))
self.set_geometry(
asarray(concat((self.vertices, vertices)), dtype=numpy.float32),
asarray(concat((self.normals, normals)), dtype=numpy.float32),
asarray(concat((self.triangles, triangles + offset)),
dtype=numpy.int32))
self.vertex_colors = new_vertex_colors
s = _AtomicShape(range(start, self.triangles.shape[0]), description,
atoms)
self._shapes.append(s)
def expand_buffer(self, block_size=None):
if block_size is not None:
self.block_size = block_size
self.value_history = np.concat(self.value_history,
np.empty(self.block_size))
self.ratio_history = np.concat(self.ratio_history,
np.empty(self.block_size))
self.max_history_size += self.block_size
def opponent_colour(mat):
r = mat[:, :, 0]
g = mat[:, :, 1]
b = mat[:, :, 2]
#probably a better way to do this
I = r + g + b
RG = r - g
BY = b - (r + g) / 2
# hopefully this concat works!
return np.concat(np.concat(I, RG), BY)
def predict(self, img_input):
init_attention_map = np.zeros(img_input.shape[:-1] + (1, ))
init_attention_map.fill(0.5)
init_cell_state = np.zeros(img_input.shape[:-1] + (32, ))
zeros_mask = np.zeros(img_input.shape[:-1] + (1, ))
attention_map, lstm_feats, attention_maps = self.attentive_rnn.predict(
[img_input, init_attention_map, init_cell_state, zeros_mask])
np.concat([attention_map, img_input], axis=-1)
skip_output_1, skip_output_2, skip_output_3 = self.autoencoder.predict(
auto_encoder_input)
return skip_output_3
def get_batch(self, batch_size):
probs = self.priorities ** self.per_alpha
probs /= probs.sum()
self.indices = np.random.choice(self.capacity, batch_size, p=probs)
samples = [self.buffer[idx] for idx in self.indices]
# Importance Sampling reweighting
weights = (len(self.buffer) * probs[self.indices]) ** (-self.per_beta)
weights /= weights.max()
self.weights = np.array(weights, dtype=np.float32)
state, action, reward, next_state, done = zip(*samples)
return concat(state), action, reward, concat(next_state), done
def smooth_left_right(inputMap, minFrag=500, boundary_frags=True):
from numpy import array, concatenate as concat, sum
if len(inputMap.frags) < 2:
return inputMap
frags = np.array(inputMap.frags)
# Remove boundary fragments - ignore them while smoothing
if boundary_frags:
bleft = frags[0]
bright = frags[-1]
frags = frags[1:-1]
while True:
num_frags = frags.shape[0]
if num_frags < 2:
break
i = np.argmin(frags)
f = frags[i]
last_ind = num_frags - 1
if f > minFrag:
break
if (i == 0):
# This is the first frag, merge right
frags = concat(([frags[0] + frags[1]], frags[2:]))
elif (i == last_ind):
# This is the laste frag, merge left
frags = concat((frags[:-2], [frags[-1] + frags[-2]]))
else:
# This is an interior frag, merge with min
left_frag = frags[i - 1]
right_frag = frags[i + 1]
if left_frag < right_frag:
frags = concat(
(frags[0:i - 1], [frags[i - 1] + frags[i]], frags[i + 1:]))
else:
frags = concat(
(frags[0:i], [frags[i] + frags[i + 1]], frags[i + 2:]))
# Reattach boundary fragments
if boundary_frags:
frags = concat(([bleft], frags, [bright]))
output_map = MalignerMap(frags=frags, mapId=inputMap.mapId)
return output_map
def smooth_left_right(inputMap, minFrag=500, boundary_frags = True):
from numpy import array, concatenate as concat, sum
if len(inputMap.frags) < 2:
return inputMap
frags = np.array(inputMap.frags)
# Remove boundary fragments - ignore them while smoothing
if boundary_frags:
bleft = frags[0]
bright = frags[-1]
frags = frags[1:-1]
while True:
num_frags = frags.shape[0]
if num_frags < 2:
break
i = np.argmin(frags)
f = frags[i]
last_ind = num_frags - 1
if f > minFrag:
break
if (i == 0):
# This is the first frag, merge right
frags = concat(([frags[0] + frags[1]], frags[2:]))
elif (i == last_ind):
# This is the laste frag, merge left
frags = concat((frags[:-2], [frags[-1] + frags[-2]]))
else:
# This is an interior frag, merge with min
left_frag = frags[i-1]
right_frag = frags[i+1]
if left_frag < right_frag:
frags = concat((frags[0:i-1], [frags[i-1] + frags[i]], frags[i+1:]))
else:
frags = concat((frags[0:i], [frags[i] + frags[i+1]], frags[i+2:]))
# Reattach boundary fragments
if boundary_frags:
frags = concat(([bleft], frags, [bright]))
output_map = MalignerMap(frags = frags, mapId = inputMap.mapId)
return output_map
def add_shapes(self, shape_info):
"""Add multiple shapes to drawing
Parameters
----------
shape_info: sequence of :py:class:`AtomicShapeInfo`
There must be no initial geometry.
"""
from numpy import empty, float32, int32, uint8, concatenate as concat
num_shapes = len(shape_info)
all_vertices = [None] * num_shapes
all_normals = [None] * num_shapes
all_triangles = [None] * num_shapes
all_colors = [None] * num_shapes
all_shapes = [None] * num_shapes
num_vertices = 0
num_triangles = 0
has_atoms = False
for i, info in enumerate(shape_info):
vertices, normals, triangles, color, atoms, description = info
all_vertices[i] = vertices
all_normals[i] = normals
all_triangles[i] = triangles + num_vertices
if color.ndim == 1 or color.shape[0] == 1:
colors = empty((vertices.shape[0], 4), dtype=uint8)
colors[:] = color
else:
colors = color.asarray(color, dtype=uint8)
assert colors.shape[1] == 4 and colors.shape[
0] == vertices.shape[0]
all_colors[i] = colors
has_atoms = has_atoms or (atoms is not None)
new_num_triangles = num_triangles + len(triangles)
all_shapes[i] = _AtomicShape(
range(num_triangles, new_num_triangles), description, atoms)
num_vertices += len(vertices)
num_triangles = new_num_triangles
if has_atoms:
self._add_handler_if_needed()
vertices = empty((num_vertices, 3), dtype=float32)
normals = empty((num_vertices, 3), dtype=float32)
triangles = empty((num_triangles, 3), dtype=int32)
self.set_geometry(concat(all_vertices, out=vertices),
concat(all_normals, out=normals),
concat(all_triangles, out=triangles))
self.vertex_colors = concat(all_colors)
self._shapes = all_shapes
def vstack(v):
if len(v) == 0:
return np.array([])
if v[0].ndim == 1:
return np.concat(v)
else:
return np.vstack(v)
def rbind(*dfs):
# Add assertion for all sets of columns being equal
result = pd.concat(dfs[0], axis=0, ignore_index=True)
return result
def cap_n_bound(self, traces, bound, Tg, decay=False):
""" take 2d array and set values >= upper boundary as np.nan
return dataframe of capped ndarray, assumes all traces meet
the criteria trials[any(val >= a)]
::Arguments::
traces (ndarray):
go decision traces for single cond
bound (float):
boundary for given condition, all
values>=bound --> np.nan
decay (bool <False>):
concatenate mirror image of input array
along axis 1 (i.e., bold up, decay down)
::Returns::
traces_df (DataFrame):
all go traces
"""
traces[traces >= bound] = np.nan
if decay:
traces_capped = concat((traces, traces[:, ::-1]), axis=1)
else:
traces_capped = traces
traces_df = pd.DataFrame(traces_capped.T)
traces_df.iloc[Tg:, :] = np.nan
return traces_df
def set_robot_blk_states(self, rState, bStates):
"""Set simulator internal states
:param rState: 1D np array (x, y, theta)
:param bStates: 2D np array [(x, y)]
"""
state = np.concat((rState, bStates.flat))
return self.set_state(state)
def better_matching(workers_vec, contract_vec):
_, UN = contract_vec.shape
assert _ == 1
_, WN = workers_vec.shape
assert _ == 1
matching = np.zeros((WN, UN))
probe_matrix = abs(contract_vec - workers_vec.T)
probe_matrix = abs(contract_vec) - probe_matrix
assert (WN, UN) == probe_matrix.shape
if WN == 1:
hr = workers_vec[0, 0]
i = np.argmax(probe_matrix)
matching[0, i] = 1
target = abs(contract_vec[i] - hr)
return matching, target
best_target = 100500100500
best_matching = None
for i in range(WN):
_workers = np.concat([workers_vec[:, :i], workers_vec[:, i + 1:]])
for j in range(UN):
_contracts = contract_vec.copy
_contracts[j] -= workers_vec[i]
small_target, small_matching = better_matching(
_workers, _contracts)
if small_target + 0:
pass
return best_matching, best_target
def vstack(v):
if len(v) == 0:
return np.array([])
if v[0].ndim == 1:
return np.concat(v)
else:
return np.vstack(v)
def generate_dpm_traces(self):
""" Get go,ssrt using same function as proactive then use the indices for
correct go and correct stop trials to slice summed, but momentary, dvg/dvs evidence vectors
(not accumulated over time just a basic sum of the vectors)
"""
# ensure parameters are all vectorized
self.p = self.vectorize_params(self.p)
Pg, Tg = self.__update_go_process__(self.p)
Ps, Ts = self.__update_stop_process__(self.p)
self.bound = self.p['a']
self.onset = self.p['tr']
ncond = self.ncond
ntot = self.ntot
dx = self.dx
base = self.base
self.nss_all = self.nss
nssd = self.nssd
ssd = self.ssd
nss = self.nss_all / self.nssd
xtb = self.xtb
get_ssbase = lambda Ts, Tg, DVg: array([[DVc[:nss / nssd, ix] for ix in np.where(Ts < Tg[i], Tg[i] - Ts, 0)] for i, DVc in enumerate(DVg)])[:, :, :, na]
self.gomoments = xtb[:, na] * np.where((rs((ncond, ntot, Tg.max())).T < Pg), dx, -dx).T
self.ssmoments = np.where(rs((ncond, nssd, nss, Ts.max())) < Ps, dx, -dx)
DVg = base[:, na] + xtb[:, na] * np.cumsum(self.gomoments, axis=2)
self.dvg = DVg[:, :nss_all, :]
self.dvs = self.get_ssbase(Ts, Tg, DVg) + np.cumsum(self.ssmoments, axis=3)
dg = self.gomoments[:, nss_all:, :].reshape(ncond, nssd, nss, Tg.max())
ds = self.ssmoments.copy()
ss_list, go_list = [], []
for i, (s, g) in enumerate(zip(ds, dg)):
diff = Ts - Tg[i]
pad_go = diff[diff > 0]
pad_ss = abs(diff[diff < 0])
go_list.extend([concat((np.zeros((nss, gp)), g[i]), axis=1) for gp in pad_go])
go_list.extend([g[x] for x in range(len(pad_ss))])
ss_list.extend([concat((np.zeros((nss, abs(spad))), s[i, :, -(Tg[i] - spad):]), axis=1) for spad in pad_ss])
ss_list.extend([s[i, :, -tss:] for tss in Ts[:len(pad_go)]])
r = map((lambda x: np.cumsum((x[0] + x[1]), axis=1)), zip(go_list, ss_list))
def interpolate(x0s, x, y, axis=0): """ Interpolate y(x) onto points in x0s using piecewise linear interpolation If y is multidimensional then `axis` specifies the axis along which x varies """ y0s = [interpolate1(x0, x, y, axis) for x0 in x0s] return np.concat([unsqueeze(y0, y.ndim, axis) for y0 in y0s], axis)
def construct_observation(observation, delta):
'''Construct an observation from metrics plus deltas
Takes raw values, standardizes, and concats
'''
return np.concat(
[metric_standardize(observation),
metric_standardize(delta)])
def dobeshi(signal):
S = signal
n = len(S)
s1 = S[::2] + sqrt(3) * S[1::2]
d1 = S[1::2] - sqrt(3)/4 * s1 - (sqrt(3) - 2) / 4 * concat(((s1[-1],), s1[:-1]))
s2 = s1 - concat((d1[1:n//2], (d1[0],)))
s = (sqrt(3)-1) / sqrt(2) * s2
d = -(sqrt(3)+1) / sqrt(2) * d1
d1 = d * ((sqrt(3)-1)/sqrt(2))
s2 = s * ((sqrt(3)+1)/sqrt(2))
s1 = s2 + roll(d1, -1)
S[1::2] = d1 + sqrt(3)/4*s1 + (sqrt(3)-2)/4*roll(s1, 1)
S[::2] = s1 - sqrt(3) * S[1::2]
return d, S
def phase_plane(func, params, fig_name, coords, func_vars = [0,1], bounds = (), \
param_tuple = (), fig_title = '', arrows = False, interval_cond = (0, 5000, 100000),
traj_type = 'reg', mod_param = 0, xlim = None, ylim = None, xlabel = None, ylabel = None,
time_series = False, pplane_dir = '', tseries_dir = ''):
graph_specifications(graph_type = 'phase plane', xlabel = xlabel, ylabel = ylabel, param_tuple = param_tuple, \
xlim = xlim, ylim = ylim, title = fig_title)
if arrows:
#creating an array of points corresponding to the 2d grid where vector field arrows are going to be graphed
corner, r, sampling = bounds
xy = neighborhood(*bounds)
t = np.linspace(0, 1, 10)
phi_integrated_arr = np.empty((10, 2))
vector_coords = np.repeat([coords[0]], xy.shape[0], axis=0)
vector_coords[:, func_vars] = xy
for phi in vector_coords:
solution = odeint(func, phi, t, args=params)[:, func_vars]
phi_integrated_arr = concat((phi_integrated_arr, solution), axis=1)
#finding the direction for the vectors in forms of differences to plot them with plt.arrow
sol_arr = phi_integrated_arr[:, 2:]
difference = sol_arr - np.roll(sol_arr, 1, axis=0)
sol_arr = sol_arr[0]
difference = difference[1]
scaling_factor = r / sampling
diff_normalized = diff_normalization(difference) * scaling_factor * 0.5
size = sol_arr.size
sol = sol_arr.reshape((int(size / 2), 2))
for start, change in zip(sol, diff_normalized):
plt.arrow(start[0],
start[1],
change[0],
change[1],
width=scaling_factor * 0.1,
color="#808080")
phase_plane_traj(func, coords, params, interval_cond, func_vars = func_vars, traj_type = traj_type, \
mod_param = mod_param)
pplane_name = pplane_dir + "PhasePlane_" + fig_name
plt.savefig("./{}".format(pplane_name))
plt.show()
if time_series:
start, end, sample = interval_cond
trajectories, t = selected_traj(func, params, coords, interval_cond)
if traj_type == 'mod':
traj1 = mod(trajectories[sample - 100:sample, func_vars[0]],
mod_param)
else:
traj1 = trajectories[sample - 100:sample, func_vars[0]]
t = t[sample - 100:sample]
pltime_series(traj1, t, 't', xlabel, fig_name, tseries_dir=tseries_dir)
return 'vector_field'
def __init__(self, *args):
multiples = [np.unique(ids) for ids in args]
unique = np.concat(multiples)
if len(multiples) > 1:
unique = np.unique(unique)
unique.sort()
self._index = pd.Index(unique)
self._dtype = np.min_scalar_type(len(self._index))
def predict(self, X): """ This method makes predictions on test data 'X'. Parameters ---------- X : numpy array N x M numpy array; N = number of data points; M = number of features. """ L = len(self.wts) Z = arr(concat((np.ones((mat(X).shape[0],1)), mat(X)), axis=1)) # initialize to input features + constant for l in range(L - 2): Z = arr(mat(Z) * mat(self.wts[l]).T) # compute linear response of next layer Z = arr(concat((np.ones((mat(Z).shape[0],1)), mat(self.sig(Z))), axis=1)) # initialize to input features + constant Z = arr(mat(Z) * mat(self.wts[L - 1].T)) # compute output layer linear response return self.sig_0(Z) # output layer activation function
def combine_meshes(meshes):
"""
Combine the iterable of (vertices, faces) into a single (vertices, faces).
Args:
meshes: iterable of (vertices, faces)
Returns:
vertices: (nv, 3) float numpy array of vertex coordinates
faces: (nf, 3) int numpy array of face vertex indices.
"""
vertices = []
faces = []
nv = 0
for v, f in meshes:
vertices.append(v)
faces.append(f + nv)
nv += len(v)
return np.concat(vertices, axis=0), np.concat(faces, axis=0)
def __responses(self, wts, X, sig, sig_0): """ This is a helper method that gets the linear sum from previous layer (A) and saturated activation responses (Z) for a data point. Used in: train """ L = len(wts) + 1 A = [None for i in range(L)] Z = arr([None for i in range(L)], dtype=object).ravel() A[0] = arr([1]) Z[0] = arr(concat((np.ones((mat(X).shape[0],1)), mat(X)), axis=1)) # compute linear combination of inputs for l in range(1, L - 1): A[l] = arr(mat(Z[l - 1]) * mat(wts[l - 1]).T) # compute linear combination of previous layer Z[l] = arr(concat((np.ones((mat(X).shape[0],1)), mat(sig(A[l]))), axis=1)) # pass through activation function and add constant feature A[L - 1] = arr(mat(Z[L - 2]) * mat(wts[L - 2]).T) Z[L - 1] = sig_0(A[L - 1]) # output layer return (A,Z)
def fit(self, x, y):
one = np.array([1])
if self.W is None:
self.shapes[0] = (x.shape[1] + 1, self.shapes[0][1])
self.W = [
np.random.rand(*shape).T / shape[0] for shape in self.shapes
]
for i in range(y.shape[0]):
o = reduce(
lambda o, w: o +
[1 / (1 + np.exp(-np.dot(w, concat(
(o[-1], one)))))], self.W, [x[i]])
dLu = (o[-1] - y[i]) * o[-1] * (1 - o[-1])
for j in range(self.num_layers, 0, -1):
dLW = np.dot(dLu.reshape(-1, 1),
concat((o[j - 1], one)).reshape(-1, 1).T)
dLu = np.dot(self.W[j - 1][:, :-1].T,
dLu) * o[j - 1] * (1 - o[j - 1])
self.W[j - 1] -= 0.1 * dLW
return self
def read_attn(x, xhat, h_dec_prev, Fx, Fy, gamma, N):
Fx_t = np.transpose(Fx, perm=[0, 2, 1])
x = np.reshape(x, [1, 128, 128])
xhat = np.reshape(xhat, [1, 128, 128])
FyxFx_t = np.reshape(np.matmul(Fy, np.matmul(x, Fx_t)), [-1, N * N])
FyxhatFx_t = np.reshape(np.matmul(Fy, np.matmul(x, Fx_t)), [-1, N * N])
return gamma * np.concat([FyxFx_t, FyxhatFx_t], 1)
def inputs_preprocssing(X, mode):
# X \in [0, 255], RGB
if mode == 'inception':
return normalize_range(X, [0, 255], [-1, 1])
elif mode == 'vgg':
vgg_bgr_mean = [103.939, 116.779, 123.68]
X_r, X_g, X_b = np.split(X, indices_or_sections=3, axis=3)
X_b -= vgg_bgr_mean[0]
X_g -= vgg_bgr_mean[1]
X_r -= vgg_bgr_mean[2]
return np.concat([X_b, X_g, X_r], axis=3)
def combine_triangles(triangle_info):
from numpy import empty, float32, int32, concatenate as concat
all_vertices = []
all_normals = []
all_triangles = []
num_vertices = 0
num_triangles = 0
for i, info in enumerate(triangle_info):
vertices, normals, triangles = info
all_vertices.append(vertices)
all_normals.append(normals)
all_triangles.append(triangles + num_vertices)
num_vertices += len(vertices)
num_triangles += len(triangles)
vertices = empty((num_vertices, 3), dtype=float32)
normals = empty((num_vertices, 3), dtype=float32)
triangles = empty((num_triangles, 3), dtype=int32)
return (concat(all_vertices, out=vertices), concat(all_normals,
out=normals),
concat(all_triangles, out=triangles))
def fir_pre_phase(b, x, n_ramp=None):
# applies a FIR filter with a pre-phase ramp
# to reduce ripple
#
# Arguments:
# * b: FIR coefficients
# * x: input signal
# * n_ramp: number of samples in pre-ramp
# (default = len(b))
signal = np.concat((np.linspace(-x[0], x[0], n_ramp), x))
y = np.lfilter(b, 1, signal)
return y[n_ramp + 1 :]
def generate_features(img_path):
# Returns relative size of axes, and normalized sum of the rows and column
rc_ratio, row_avg, col_avg = efg.get_img_edge_data(img_path, blur=3)
# Combine the later
row_col_arr = np.concatenate((row_avg, col_avg))
# Run pca to collapse to 1/20 the original size, 85% variance
pca_vals = _RC_PCA.transform(row_col_arr)
return np.concat((np.array(rc_ratio), pca_vals))
def fir_pre_phase(b, x, n_ramp=None):
# applies a FIR filter with a pre-phase ramp
# to reduce ripple
#
# Arguments:
# * b: FIR coefficients
# * x: input signal
# * n_ramp: number of samples in pre-ramp
# (default = len(b))
signal = np.concat((np.linspace(-x[0],x[0], n_ramp), x))
y = np.lfilter(b,1,signal)
return y[n_ramp+1:]
def get_batch(self, batch_size):
probs = self.priority_tree[-self.num_leaf_nodes:] ** self.per_alpha
probs /= probs.sum()
self.indices = np.zeros((batch_size,), dtype=np.int32)
priority_segment = self.total_priority / batch_size
for i in range(batch_size):
seg_start = priority_segment * i
seg_end = priority_segment * (i + 1)
# Sampling is linear in batch size, but logarithmic in buffer size
tree_idx = self.extract_index(np.random.uniform(seg_start, seg_end))
# convert tree_idx to buffer_idx which equals number of leaf nodes
self.indices[i] = tree_idx - self.num_parent_nodes
samples = [self.buffer[idx] for idx in self.indices]
# Importance Sampling reweighting
weights = (len(self.buffer) * probs[self.indices]) ** (-self.per_beta)
weights /= weights.max()
self.weights = np.array(weights, dtype=np.float32)
state, action, reward, next_state, done = zip(*samples)
return concat(state), action, reward, concat(next_state), done
def makeBatch(self, params):
size = np.random.randint(
low=params['lowSize'],
high=params['highSize'] + 1,
size=[params['batchSize'], params['length'], 1])
weight = np.random.randint(
low=params['lowWeight'],
high=params['highWeight'] + 1,
size=[params['batchSize'], params['length'], 1])
self.capacity = params['capacity']
return np.concat([size, weight], axis=2)
def extend_shape(self, vertices, normals, triangles, color=None):
"""Extend previous shape
Parameters
----------
vertices : :py:class:`numpy.array` of coordinates
normals : :py:class:`numpy.array` of normals, one per vertex
triangles : :py:class:`numpy.array` of vertex indices, multiple of 3
color : either None, a single 4 element uint8 :py:class:`numpy.array`;
or an array of those values, one per vertex. If None, then the
color is same as the last color of the existing shape.
The associated atoms and description are that of the extended shape.
"""
if self.vertices is None:
raise ValueError("no shape to extend")
asarray = numpy.asarray
concat = numpy.concatenate
if color is None or color.ndim == 1 or color.shape[0] == 1:
colors = numpy.empty((vertices.shape[0], 4), dtype=numpy.uint8)
if color is None:
colors[:] = self.vertex_colors[-1]
else:
colors[:] = color
else:
colors = color.asarray(color, dtype=numpy.uint8)
assert colors.shape[1] == 4 and colors.shape[0] == vertices.shape[0]
offset = self.vertices.shape[0]
new_vertex_colors = concat((self.vertex_colors, colors))
self.set_geometry(
asarray(concat((self.vertices, vertices)), dtype=numpy.float32),
asarray(concat((self.normals, normals)), dtype=numpy.float32),
asarray(concat((self.triangles, triangles + offset)),
dtype=numpy.int32))
self.vertex_colors = new_vertex_colors
s = self._shapes[-1]
s.triangle_range = range(s.triangle_range.start,
self.triangles.shape[0])
def fill_to_end_of_year(df):
last = df.tail(1).index[0]
year = last.year
end = datetime.datetime(year + 1, 1, 1)
index_lst = []
i = 1
while last + datetime.timedelta(days=i) < end:
index_lst.append(last + datetime.timedelta(days=i))
i += 1
filler = np.empty((len(index_lst), len(df.columns)))
rest = DataFrame(filler, columns=df.columns)
rest.index = index_lst
return np.concat((df, rest))
def build_K_matrix(self):
zeros = [
[0,0,0],
[0,0,0],
[0,0,0]
]
self.L2G = concat((concat((self.Tm,zeros),axis=1),concat((zeros,self.Tm),axis=1)),axis=0)
self.G2L = concat((concat((self.TmT,zeros),axis=1),concat((zeros,self.TmT),axis=1)),axis=0)
self.K = np.dot(np.dot(self.L2G,self.Km),self.G2L)
def __responses(self, wts, X_in, sig, sig_0): """ Helper function that gets linear sum from previous layer (A) and saturated activation responses (Z) for a data point. Used in: train """ L = len(wts) constant_feat = np.ones((mat(X_in).shape[0],1)).flatten() # constant feature # compute linear combination of inputs A = [arr([1])] Z = [concat((constant_feat, X_in))] for l in range(1, L): A.append(Z[l - 1].dot(wts[l - 1].T)) # compute linear combination of previous layer # pass through activation function and add constant feature Z.append(cols((np.ones((mat(A[l]).shape[0],1)),sig(A[l])))) A.append(arr(mat(Z[L - 1]) * mat(wts[L - 1]).T)) Z.append(arr(sig_0(A[L]))) # output layer (saturate for classifier, not regressor) return A,Z
def get_layers(self): S = arr([mat(self.wts[i]).shape[1] - 1 for i in range(len(self.wts))]) S = concat((S, [mat(self.wts[-1]).shape[0]])) return S
m1i = m1k
m0i = m0k
for j, r in zip(s2, rk):
if j<0: continue
m1i[j] += r
m0i[j] += 1
return m1, m0, m1k, m0k
if __name__ == "__main__":
from numpy.random import binomial
from numpy import concatenate as concat
numpy.random.seed(123)
d = 5
phi = [[0.1, 0.7, 0.2], [0.1, 0.3, 0.9], [0.8, 0.1, 0.2]]
orgR = concat([concat([binomial(1, p, size=(d,d)) for p in pp], axis=1) for pp in phi])
i = numpy.arange(orgR.shape[0])
numpy.random.shuffle(i)
R = orgR[i,:]
i = numpy.arange(orgR.shape[1])
numpy.random.shuffle(i)
R = R[:,i]
model = IRM(R, alpha=1.0, a=1.0, b=1.0)
maxv = -1e9
for i in range(200):
model.update()
v = model.log_posterior()
if v > maxv:
maxv = v
maxm = model.clone()