def vis_hybrid_image(hybrid_image):
"""
%visualize a hybrid image by progressively downsampling the image and
%concatenating all of the images together.
"""
scales = 5 #how many downsampled versions to create
padding = 5 #how many pixels to pad.
original_height = hybrid_image.shape[0]
num_colors = hybrid_image.shape[2]
#counting how many color channels the input has
output = hybrid_image
cur_image = hybrid_image
for i in range(2, scales + 1):
gap_array = np.ones((original_height, padding, num_colors))
output = cat((output, gap_array),
axis=1) # wrongly done in the sample code
cur_image = resize(cur_image,
(cur_image.shape[0] // 2, cur_image.shape[1] // 2),
anti_aliasing=True)
vertical_gap = np.ones((original_height - cur_image.shape[0],
cur_image.shape[1], num_colors))
tmp = cat((vertical_gap, cur_image), axis=0)
output = cat((output, tmp), axis=1)
return (output)
def concatenate(data, axis):
"""Concatenate multiple trials into one trials, according to any dimension.
Parameters
----------
data : instance of DataTime, DataFreq, or DataTimeFreq
axis : str
axis that you want to concatenate (it can be 'trial')
Returns
-------
instace of same class as input
the data will always have only one trial
Notes
-----
If axis is 'trial', it will add one more dimension, and concatenate based
on it. It will then create a new axis, called 'trial_axis' (not 'trial'
because that axis is hard-coded).
If you want to concatenate across trials, you need:
>>> expand_dims(data1.data[0], axis=1).shape
"""
output = data._copy(axis=False)
for dataaxis in data.axis:
output.axis[dataaxis] = empty(1, dtype='O')
if dataaxis == axis:
output.axis[dataaxis][0] = cat(data.axis[dataaxis])
else:
output.axis[dataaxis][0] = data.axis[dataaxis][0]
if len(unique(output.axis[dataaxis][0])) != len(
output.axis[dataaxis][0]):
lg.warning('Axis ' + dataaxis + ' does not have unique values')
output.data = empty(1, dtype='O')
if axis == 'trial':
# create new axis
new_axis = empty(1, dtype='O')
n_trial = data.number_of('trial')
trial_name = ['trial{0:06}'.format(x) for x in range(n_trial)]
new_axis[0] = asarray(trial_name, dtype='U')
output.axis['trial_axis'] = new_axis
# concatenate along the extra dimension
all_trial = []
for one_trial in data.data:
all_trial.append(expand_dims(one_trial, -1))
output.data[0] = cat(all_trial, axis=-1)
else:
output.data[0] = cat(data.data, axis=output.index_of(axis))
return output
def pos_to_mat(pos, rot_mat=None):
"""
Given an (x, y, z) position, create a 4x4 homogeneous transformation matrix.
"""
mat = eye(3) if rot_mat is None else rot_mat
mat = cat((mat, A(*(pos,)).T), axis=1)
mat = cat((mat, A((0,0,0,1))))
return mat
def concatenate(data, axis):
"""Concatenate multiple trials into one trials, according to any dimension.
Parameters
----------
data : instance of DataTime, DataFreq, or DataTimeFreq
axis : str
axis that you want to concatenate (it can be 'trial')
Returns
-------
instace of same class as input
the data will always have only one trial
Notes
-----
If axis is 'trial', it will add one more dimension, and concatenate based
on it. It will then create a new axis, called 'trial_axis' (not 'trial'
because that axis is hard-coded).
If you want to concatenate across trials, you need:
>>> expand_dims(data1.data[0], axis=1).shape
"""
output = data._copy(axis=False)
for dataaxis in data.axis:
output.axis[dataaxis] = empty(1, dtype='O')
if dataaxis == axis:
output.axis[dataaxis][0] = cat(data.axis[dataaxis])
else:
output.axis[dataaxis][0] = data.axis[dataaxis][0]
if len(unique(output.axis[dataaxis][0])) != len(output.axis[dataaxis][0]):
lg.warning('Axis ' + dataaxis + ' does not have unique values')
output.data = empty(1, dtype='O')
if axis == 'trial':
# create new axis
new_axis = empty(1, dtype='O')
n_trial = data.number_of('trial')
trial_name = ['trial{0:06}'.format(x) for x in range(n_trial)]
new_axis[0] = asarray(trial_name, dtype='U')
output.axis['trial_axis'] = new_axis
# concatenate along the extra dimension
all_trial = []
for one_trial in data.data:
all_trial.append(expand_dims(one_trial, -1))
output.data[0] = cat(all_trial, axis=-1)
else:
output.data[0] = cat(data.data, axis=output.index_of(axis))
return output
def compute_target_part_scoremap_slices(self, joint_id, coords, data_item, size, scale):
dist_thresh = self.cfg.pos_dist_thresh * scale
dist_thresh_sq = dist_thresh ** 2
num_joints = self.cfg.num_joints
# TODO: truncate size to agree with resnet depth
size[0] = int(np.sqrt(size[0]))**2 # new
scmap = np.zeros(cat([size, arr([num_joints])]))
locref_size = cat([size, arr([num_joints * 2])])
locref_mask = np.zeros(locref_size)
locref_map = np.zeros(locref_size)
width = size[2] # new
height = size[1] # new
truncated_depth = size[0] - 1 # new
for person_id in range(len(coords)):
for k, j_id in enumerate(joint_id[person_id]):
joint_pt = coords[person_id][k, :]
# print("Joint {} ".format(k), joint_pt)
j_x = np.asscalar(joint_pt[0])
j_y = np.asscalar(joint_pt[1])
# Note: the annotations are XYZ, but the masks are DHW=ZXY
# TODO: for now, truncate to depth of resnet output
j_z = min(int(joint_pt[2])-1, truncated_depth) # new; starts at 1; not affected by stride or distance
# don't loop over entire heatmap, but just relevant locations
j_x_sm = round((j_x - self.half_stride) / self.stride)
j_y_sm = round((j_y - self.half_stride) / self.stride)
min_x = round(max(j_x_sm - dist_thresh - 1, 0))
max_x = round(min(j_x_sm + dist_thresh + 1, width - 1))
min_y = round(max(j_y_sm - dist_thresh - 1, 0))
max_y = round(min(j_y_sm + dist_thresh + 1, height - 1))
for j in range(min_y, max_y + 1): # range(height):
pt_y = j * self.stride + self.half_stride
for i in range(min_x, max_x + 1): # range(width):
# pt = arr([i*stride+half_stride, j*stride+half_stride])
# diff = joint_pt - pt
# The code above is too slow in python
pt_x = i * self.stride + self.half_stride
dx = j_x - pt_x
dy = j_y - pt_y
dist = dx ** 2 + dy ** 2
# print(la.norm(diff))
if dist <= dist_thresh_sq:
# New index: z, which is not iterated
scmap[j_z, j, i, j_id] = 1
locref_mask[j_z, j, i, j_id * 2 + 0] = 1
locref_mask[j_z, j, i, j_id * 2 + 1] = 1
locref_map[j_z, j, i, j_id * 2 + 0] = dx * self.locref_scale
locref_map[j_z, j, i, j_id * 2 + 1] = dy * self.locref_scale
weights = self.compute_scmap_weights(scmap.shape, joint_id, data_item)
return scmap, weights, locref_map, locref_mask
def compute_target_part_scoremap(self, joint_id, coords, data_item, size,
scale):
stride = self.cfg.stride
dist_thresh = self.cfg.pos_dist_thresh * scale
num_joints = self.cfg.num_joints
half_stride = stride / 2
scmap = np.zeros(cat([size, arr([num_joints])]))
locref_size = cat([size, arr([num_joints * 2])])
locref_mask = np.zeros(locref_size)
locref_map = np.zeros(locref_size)
locref_scale = 1.0 / self.cfg.locref_stdev
dist_thresh_sq = dist_thresh**2
width = size[1]
height = size[0]
for person_id in range(len(coords)):
for k, j_id in enumerate(joint_id[person_id]):
joint_pt = coords[person_id][k, :]
j_x = np.asscalar(joint_pt[0])
j_y = np.asscalar(joint_pt[1])
if np.isnan(j_x) or np.isnan(j_y):
continue
# don't loop over entire heatmap, but just relevant locations
j_x_sm = round((j_x - half_stride) / stride)
j_y_sm = round((j_y - half_stride) / stride)
min_x = round(max(j_x_sm - dist_thresh - 1, 0))
max_x = round(min(j_x_sm + dist_thresh + 1, width - 1))
min_y = round(max(j_y_sm - dist_thresh - 1, 0))
max_y = round(min(j_y_sm + dist_thresh + 1, height - 1))
for j in range(int(min_y), int(max_y) + 1): # range(height):
pt_y = j * stride + half_stride
for i in range(int(min_x),
int(max_x) + 1): # range(width):
# pt = arr([i*stride+half_stride, j*stride+half_stride])
# diff = joint_pt - pt
# The code above is too slow in python
pt_x = i * stride + half_stride
dx = j_x - pt_x
dy = j_y - pt_y
dist = dx**2 + dy**2
# print(la.norm(diff))
if dist <= dist_thresh_sq:
scmap[j, i, j_id] = 1
locref_mask[j, i, j_id * 2 + 0] = 1
locref_mask[j, i, j_id * 2 + 1] = 1
locref_map[j, i, j_id * 2 + 0] = dx * locref_scale
locref_map[j, i, j_id * 2 + 1] = dy * locref_scale
weights = self.compute_scmap_weights(scmap.shape, joint_id, data_item)
return scmap, weights, locref_map, locref_mask
def __generate_linear_row(T, depth, zeros, dim):
"""Generates linear state transition matrix rows"""
core = np.eye(dim)
if depth == 1:
return cat((np.zeros((dim, dim * zeros)), core), axis=1)
depth = depth - 1
return cat((
__generate_linear_row(T, depth, zeros, dim),
T**depth / math.factorial(depth) * core,
),
axis=1)
def cgpstudy():
"""
Connecting the building blocks of a cGP study.
This implements the simulation pipeline in Gjuvsland et al. (2011).
"""
from numpy import concatenate as cat
gt = np.array(genotypes)
par = cat([gt2par(list(g), hetpar, loc2par) for g in gt])
ph = cat([par2ph(p) for p in par])
return gt, par, ph
def subband_group_sum(self,s,group_type,average=True, energy=True):
"""
Computes the sum (or average) (energy) of one of two group types for subband s
Group_type can be either 'parent_children' or 'children' or 'parent_child'
"""
if group_type != 'children':
w_parent = self.get_upsampled_parent(s) #THIS BAD, we have two functions calling each other (potentially)
if energy:
w_parent = np.abs(w_parent)**2
w_child = self.get_subband(s)
if energy:
w_child = np.abs(w_child)**2
if group_type == 'parent_children' or group_type == 'children':
w_child = np.sum(cat([w_child[self.ds_slices[i]][...,np.newaxis]
for i in xrange(len(self.ds_slices))],
axis=self.int_dimension),axis=self.int_dimension)
w_child_us = np.zeros(2*np.asarray(w_child.shape))
for j in xrange(len(self.ds_slices)):
w_child_us[self.ds_slices[j]] = w_child
del w_child
if group_type == 'parent_children':
divisor = 2.0**self.int_dimension + 1
else: #children only
w_parent = 0
divisor = 2.0**self.int_dimension
elif group_type == 'parent_child':
w_child_us = w_child
divisor = 2.0
w_parent += w_child_us
if average:
w_parent /= divisor
return w_parent
def __generate_linear_cols(T, depth, target, dim):
"""Carefully concatenates linear state transition matrix rows"""
if depth == 1:
return __generate_linear_row(T, depth, target, dim)
return cat((__generate_linear_row(T, depth, target, dim),
__generate_linear_cols(T, depth - 1, target + 1, dim)),
axis=0)
def add_input_key(self, cfg: GenomeConfig, k: int):
"""Extend the input-key list with the given key, and expand the corresponding weights."""
# Update self.weight_xh_full if key never seen before
if k not in self.input_keys_full:
# Find the index to insert the key
lst = [
i + 1 for i in range(len(self.input_keys_full))
if self.input_keys_full[i] < k
] # List of indices
i = lst[-1] if lst else 0 # Index to insert key in
# Save key to list
self.input_keys_full.insert(i, k)
# Update weight_xh_full correspondingly by inserting random initialized tensor in correct position
new_tensor = rnn.init(cfg, hid_dim=self.hid_dim, input_size=1)
assert new_tensor.shape == (self.hid_dim, 1)
self.weight_xh_full = cat((self.weight_xh_full[:, :i], new_tensor,
self.weight_xh_full[:, i:]),
axis=1)
# Update input_keys (current key-set) analogously
if k not in self.input_keys:
lst = [
i + 1 for i in range(len(self.input_keys))
if self.input_keys[i] < k
] # List of indices
i = lst[-1] if lst else 0
self.input_keys.insert(i, k)
def cgpstudy():
"""
Basic example of connecting the building blocks of a cGP study.
This implements the simulation pipeline used in Gjuvsland et al. (2007) and
Gjuvsland et al. (2011)
:TODO: save and read into R
:TODO: in R create classic lineplots for 3-locus GP mapping
:TODO: in R use noia package to partition variance
"""
from numpy import concatenate as cat
gt = np.array(genotypes)
par = cat([gt2par(list(g), hetpar, absvar) for g in gt])
ph = cat([par2ph(p) for p in par])
def generate_scmap(cfg, joint_id, coords, size):
dist_thresh = cfg.pos_dist_thresh * cfg.global_scale
num_joints = cfg.num_joints
scmap = np.zeros(cat([size, arr([num_joints])]))
dist_thresh_sq = dist_thresh ** 2
width = size[1]
height = size[0]
for person_id in range(len(coords)):
for k, j_id in enumerate(joint_id[person_id]):
joint_pt = coords[person_id][k, :]
j_x = np.asscalar(joint_pt[0])
j_y = np.asscalar(joint_pt[1])
# don't loop over entire heatmap, but just relevant locations
min_x = int(round(max(j_x - dist_thresh - 1, 0)))
max_x = int(round(min(j_x + dist_thresh + 1, width - 1)))
min_y = int(round(max(j_y - dist_thresh - 1, 0)))
max_y = int(round(min(j_y + dist_thresh + 1, height - 1)))
for j in range(min_y, max_y + 1): # range(height):
pt_y = j
for i in range(min_x, max_x + 1): # range(width):
pt_x = i
dx = j_x - pt_x
dy = j_y - pt_y
dist = dx ** 2 + dy ** 2
if dist <= dist_thresh_sq:
scmap[j, i, j_id] = 1
return scmap * 255
def update(self, fore, back):
template = self.template[:back.shape[0], :back.shape[1], :]
if self.reset:
id=torch.reshape(torch.arange(1,template.shape[0]*template.shape[1]+1), \
(template.shape[0],template.shape[1]))
self.slice.append(id)
self.slice_volume = torch.stack(self.slice)
self.max_id = torch.max(id)
self.id_aligned = torch.cat([self.one.long(), id.view(-1)])
self.id_len = torch.ones_like(self.id_aligned)
self.reset = False
else:
flow_check, fore_warp = self.check(fore, back)
id = torch.zeros_like(self.slice[-1])
id[[fore_warp[:, :, 0],
fore_warp[:, :, 1]]] = self.slice[-1] * flow_check.long()
# self.id_len[id]+=1
new = torch.where(id > 0, self.zero,
self.one).nonzero(as_tuple=True)
id[new] = torch.arange(len(new[0])) + self.max_id + 1
# self.cut(flow_check)
# self.id_aligned=torch.cat([self.id_aligned,id[new]])
# self.id_len=torch.cat([self.id_len,torch.ones_like(id[new])])
self.max_id = torch.max(id)
self.slice.append(id)
def __init__(self, data, embedding, layers):
self.data = data
self.embedding = embedding
self.layers = []
curr_length = data.shape[0]
#Generate latent variables
for i, (downsampling, latent_size, window, model,
opt) in enumerate(layers):
curr_length = curr_length // downsampling
if (i < len(layers) - 1):
next_window = layers[i + 1][2]
offset = max(window // downsampling, next_window)
padding = max(window // downsampling, next_window)
latent_length = offset + curr_length + padding
else:
offset = window // downsampling
padding = window // downsampling
latent_length = offset + curr_length + padding
if i == 0:
pad = np.zeros((window, ), dtype=np.int32)
self.data = np.cat((pad, data, pad))
latent_var = LatentVar((latent_length, latent_size), offset=window)
self.layers.append((offset, curr_length, latent_var, downsampling,
window, model, opt))
def batch_seq_pool(cls, seq_emb:Tensor, lengths=List[int]):
"""
Concatenate the mean, max and last hidden representations of a batch of sequences.
Parameters
----------
seq_emb: Tensor
Tensor of shape (bs, sequence length, dimension)
This tensor reprsents the hidden states of final layer of the encoder from a language model.
lengths: List
list of integers indicating the sequence lengths
Returns
-------
Tensor
Tensor of size (bs, 2400)
"""
assert seq_emb.shape[0] == len(lengths), 'Number of elements in lengths should match the first dimension of seq_emb'
# ignore information beyond the sequence length, which is padding
embs = [seq_emb[i, :x, :] for i, x in enumerate(lengths)]
# calculate the pooled features ignoring the padding
features = [cat([emb.mean(axis=0), emb.max(axis=0), emb[-1,:]], axis=-1) for emb in embs]
combined_features = stack(features)
# check that the dimensionality of the document embedding is 3x the dimensionality of the
# final hidden states of the encoder.
assert combined_features.shape[-1] == (seq_emb.shape[-1] * 3)
return combined_features
def DWResidual(self):
if 'matResidual' in locals():
matResidual.mesh = self.mesh
else:
matResidual = discretization(self.coeff, self.mesh,
self.enrichOrder)
matGroup = matResidual.matGroup()
A, B, BonQ, C, D, E, F, G, H, L, R = matGroup
LHS = np.bmat([[A, -B, C], [BonQ, D, E]])
RHS = cat((R, F))
residual = np.zeros(self.numEle)
numEnrich = self.numBasisFuncs + self.enrichOrder
adjointGradState, adjointStateFace = self.separateSoln(
self.adjointSoln)
for i in np.arange(self.numEle):
primalResidual = (LHS.dot(self.primalSoln) - RHS).A1
uLength = self.numEle * numEnrich
stepLength = i * numEnrich
uDWR = primalResidual[stepLength:stepLength + numEnrich].dot(
(1 - adjointGradState)[stepLength:stepLength + numEnrich])
qDWR = primalResidual[uLength + stepLength:uLength +
stepLength + numEnrich]\
.dot((1 - adjointGradState)[uLength + stepLength:uLength +
stepLength + numEnrich])
residual[i] = uDWR + qDWR
# sort residual index
residualIndex = np.argsort(np.abs(residual))
# select top \theta% elements with the largest error
theta = 0.15
refineIndex = residualIndex[int(self.numEle *
(1 - theta)):len(residual)] + 1
return np.abs(np.sum(residual)), refineIndex
def solveAdjoint(self):
"""Solve the adjoint problem"""
# solve in the enriched space
_coeff = copy(self.coeff)
_coeff.pOrder = _coeff.pOrder + 1
if 'matAdjoint' in locals():
matAdjoint.mesh = self.mesh
else:
matAdjoint = discretization(_coeff, self.mesh)
matGroup = matAdjoint.matGroup()
A, B, _, C, D, E, F, G, H, L, R = matGroup
# add adjoint LHS conditions
F = np.zeros(len(F))
R[-1] = -boundaryCondition('adjoint')[1]
# assemble global matrix LHS
LHS = np.bmat([[A, -B, C], [B.T, D, E], [C.T, G, H]])
sLHS = csr_matrix(LHS)
RHS = cat((R, F, L))
# solve in one shoot using GMRES
def invRHS(vec):
"""Construct preconditioner"""
matVec = spla.spsolve(sLHS, vec)
return matVec
n = len(RHS)
preconditioner = spla.LinearOperator((n, n), invRHS)
soln = spla.gmres(sLHS, RHS, M=preconditioner)[0]
# soln = np.linalg.solve(LHS.T, RHS)
self.adjointSoln = soln
def xy_to_cxcy(boxes):
'''
将(x_min, y_min, x_max, y_max)形式的anchor转换成(center_x, center_y, w, h)形式的
:param boxes: bounding boxes in boundary coordinates, a array of size (n_boxes, 4)
:return: bounding boxes in center-size coordinates, a array of size (n_boxes, 4)
'''
return np.cat([(boxes[:,2:]+boxes[:,:2])/2,boxes[:,2:]-boxes[:,:2]],axis=1)
def fit(self,train_x,train_y):
self.train_y = train_y
n,dim = train_x.shape
self.w = np.zeros(dim+1)
ones = np.ones(n,1)
self.train_x = np.cat((train_x,ones),dim=-1)
for i in range(self.max_iter):
self._update_para()
def compute_target_part_scoremap_numpy(
self, joint_id, coords, data_item, size, scale
):
dist_thresh = float(self.cfg.pos_dist_thresh * scale)
dist_thresh_sq = dist_thresh ** 2
num_joints = self.cfg.num_joints
scmap = np.zeros(cat([size, arr([num_joints])]))
locref_size = cat([size, arr([num_joints * 2])])
locref_mask = np.zeros(locref_size)
locref_map = np.zeros(locref_size)
width = size[1]
height = size[0]
grid = np.mgrid[:height, :width].transpose((1, 2, 0))
i=0
for person_id in range(len(coords)):
for k, j_id in enumerate(joint_id[person_id]):
joint_pt = coords[person_id][k, :]
j_x = np.asscalar(joint_pt[0])
j_x_sm = round((j_x - self.half_stride) / self.stride)
j_y = np.asscalar(joint_pt[1])
j_y_sm = round((j_y - self.half_stride) / self.stride)
min_x = round(max(j_x_sm - dist_thresh - 1, 0))
max_x = round(min(j_x_sm + dist_thresh + 1, width - 1))
min_y = round(max(j_y_sm - dist_thresh - 1, 0))
max_y = round(min(j_y_sm + dist_thresh + 1, height - 1))
x = grid.copy()[:, :, 1]
y = grid.copy()[:, :, 0]
dx = j_x - x * self.stride - self.half_stride
dy = j_y - y * self.stride - self.half_stride
dist = dx ** 2 + dy ** 2
mask1 = dist <= dist_thresh_sq
mask2 = (x >= min_x) & (x <= max_x)
mask3 = (y >= min_y) & (y <= max_y)
mask = mask1 & mask2 & mask3
scmap[mask, j_id] = 1
locref_mask[mask, j_id * 2 + 0] = 1
locref_mask[mask, j_id * 2 + 1] = 1
locref_map[mask, j_id * 2 + 0] = (dx * self.locref_scale)[mask]
locref_map[mask, j_id * 2 + 1] = (dy * self.locref_scale)[mask]
weights = scmap
#weights = self.compute_scmap_weights(scmap.shape, joint_id, data_item)
return scmap, weights, locref_map, locref_mask
def solvePrimal(self):
"""Solve the primal problem"""
if 'matLocal' in locals():
# if matLocal exists,
# only change the mesh instead of initializing again
matLocal.mesh = self.mesh
else:
matLocal = discretization(self.coeff, self.mesh)
matGroup = matLocal.matGroup()
A, B, _, C, D, E, F, G, H, L, R = matGroup
# solve by exploiting the local global separation
K = -cat((C.T, G), axis=1)\
.dot(np.linalg.inv(np.bmat([[A, -B], [B.T, D]]))
.dot(cat((C, E)))) + H
sK = csr_matrix(K)
F_hat = np.array([L]).T - cat((C.T, G), axis=1)\
.dot(np.linalg.inv(np.bmat([[A, -B], [B.T, D]])))\
.dot(np.array([cat((R, F))]).T)
def invRHS(vec):
"""Construct preconditioner"""
matVec = spla.spsolve(sK, vec)
return matVec
n = len(F_hat)
preconditioner = spla.LinearOperator((n, n), invRHS)
stateFace = spla.gmres(sK, F_hat, M=preconditioner)[0]
# stateFace = np.linalg.solve(K, F_hat)
gradState = np.linalg.inv(np.asarray(np.bmat(
[[A, -B], [B.T,
D]]))).dot(cat((R, F)) - cat((C, E)).dot(stateFace))
self.primalSoln = cat((gradState, stateFace))
def cgpstudy():
"""
Basic example of connecting the building blocks of a cGP study.
This top-level orchestration can be done in many different ways, depending
on personal preference and on the need for features such as storage,
caching, memory management or parallelization. Some examples are given in
:mod:`cgp.examples.hpc`.
"""
from numpy import concatenate as cat
gt = np.array(genotypes)
par = cat([monogenicpar(g, hetpar=par0, absvar=absvar) for g in gt])
ph = cat([par2ph(p) for p in par])
agg = cat([ph2agg(p) for p in ph])
summarize(gt, agg)
plt.show()
def __call__(self, other):
assert isinstance(other, CSSCode)
assert other.n * 2 == self.n
Lx, Lz, Hx, Tz, Hz, Tx, Gx, Gz = (other.Lx, other.Lz, other.Hx,
other.Tz, other.Hz, other.Tx,
other.Gx, other.Gz)
assert Gx is None
assert Gz is None
A = self.A.transpose()
LxLz = dot2(cat((Lx, Lz), axis=1), A)
HxTz = dot2(cat((Hx, Tz), axis=1), A)
TxHz = dot2(cat((Tx, Hz), axis=1), A)
n = self.n // 2
Lx, Lz = LxLz[:, :n], LxLz[:, n:]
Hx, Tz = HxTz[:, :n], HxTz[:, n:]
Tx, Hz = TxHz[:, :n], TxHz[:, n:]
code = CSSCode(Lx=Lx, Lz=Lz, Hx=Hx, Tz=Tz, Hz=Hz, Tx=Tx)
return code
def glue_logops():
m = argv.get("m", 4)
n = argv.get("n", m + m + 1)
dist = argv.get("dist", 3)
N = argv.get("N", 2)
M = argv.get("M", N)
p = argv.get("p", 0.03)
weight = argv.weight
codes = []
code = None
for i in range(N):
Hx, Hz = make_quantum(n, m, dist, weight)
#Hx, Hz = make_surface()
print("Hx, Hz:")
print(shortstrx(Hx, Hz))
c = CSSCode(Hx=Hx, Hz=Hz)
codes.append(c)
code = c if code is None else code + c
A, B = codes
code = A + B
print(code)
print("Lx, Lz:")
print(shortstrx(code.Lx, code.Lz))
print("Hx, Hz:")
print(shortstrx(code.Hx, code.Hz))
print()
#Hx = cat((code.Lx, code.Hx))
Hx = code.Hx
Hz = cat((code.Lz, code.Hz))
idxs = list(range(2 * n))
idxs.sort(key=lambda i: (-code.Lz[:, i].sum(), ))
Hx = Hx[:, idxs]
Hz = Hz[:, idxs]
print(shortstrx(Hx, Hz))
i0 = argv.get("i0")
i1 = argv.get("i1")
if i0 is None:
return
# code = code.glue(0, n)
# print(code)
# print(shortstrx(code.Hx, code.Hz))
Hx, Hz = glue1_quantum(Hx, Hz, i0, i1)
print("Hx, Hz:")
print(shortstrx(Hx, Hz))
def fig3():
x = np.linspace(-4, 4, 50)
y = np.linspace(-4, 4, 50)
x, y = np.meshgrid(x, y)
z = maximum(2 * x**2 + y**2 - x * y, abs(x) + 2 * abs(y))
levels = cat((arange(0, 4, 1), arange(4, 16, 2)))
plt.contour(x, y, z, levels=levels, cmap='jet')
txt = r'$max(2x^2 + y^2 - xy, |x| + 2|y|)$'
return txt
def fig5():
x = np.linspace(0.001, 2, 50)
y = np.linspace(0.001, 2, 50)
x, y = np.meshgrid(x, y)
z = x * log(x) + y * log(y)
levels = cat((arange(-1, 0.2, 0.1), arange(0.2, 1, 0.2)))
plt.contour(x, y, z, levels=levels, cmap='jet')
txt = r'$xlog(x) + ylog(y)$'
return txt
def fig2():
x = np.linspace(-1, 1, 50)
y = np.linspace(-1, 1, 50)
x, y = np.meshgrid(x, y)
z = x**2 + y**2
levels = cat((arange(0, 0.1, 0.05), arange(0.1, 1, 0.15)))
plt.contour(x, y, z, levels=levels, cmap='jet')
txt = r'$x^2 + y^2$'
return txt
def gaussian_scmap(self, joint_id, coords, data_item, size, scale):
stride = self.cfg.stride
#dist_thresh = float(self.cfg.pos_dist_thresh * scale)
num_joints = self.cfg.num_joints
half_stride = stride / 2
scmap = np.zeros(cat([size, arr([num_joints])]))
locref_size = cat([size, arr([num_joints * 2])])
locref_mask = np.zeros(locref_size)
locref_map = np.zeros(locref_size)
width = size[1]
height = size[0]
dist_thresh = float((width + height) / 6)
locref_scale = 1.0 / self.cfg.locref_stdev
dist_thresh_sq = dist_thresh**2
std = dist_thresh / 4
# Grid of coordinates
grid = np.mgrid[:height, :width].transpose((1, 2, 0))
grid = grid * stride + half_stride
for person_id in range(len(coords)):
for k, j_id in enumerate(joint_id[person_id]):
joint_pt = coords[person_id][k, :]
j_x = np.asscalar(joint_pt[0])
j_x_sm = round((j_x - half_stride) / stride)
j_y = np.asscalar(joint_pt[1])
j_y_sm = round((j_y - half_stride) / stride)
map_j = grid.copy()
# Distance between the joint point and each coordinate
dist = np.linalg.norm(grid - (j_y, j_x), axis=2)**2
scmap_j = np.exp(-dist / (2 * (std**2)))
scmap[..., j_id] = scmap_j
locref_mask[dist <= dist_thresh_sq, j_id * 2 + 0] = 1
locref_mask[dist <= dist_thresh_sq, j_id * 2 + 1] = 1
dx = j_x - grid.copy()[:, :, 1]
dy = j_y - grid.copy()[:, :, 0]
locref_map[..., j_id * 2 + 0] = dx * locref_scale
locref_map[..., j_id * 2 + 1] = dy * locref_scale
weights = self.compute_scmap_weights(scmap.shape, joint_id, data_item)
return scmap, weights, locref_map, locref_mask
def preprocess(
train_dir="data/in-hospital-mortality/train",
test_dir="data/in-hospital-mortality/test",
split=False,
):
train_reader = InHospitalMortalityReader(
dataset_dir=train_dir, listfile=f"{train_dir}/listfile.csv")
test_reader = InHospitalMortalityReader(
dataset_dir=test_dir, listfile=f"{test_dir}/listfile.csv")
train_data = []
test_data = []
for i in range(train_reader.get_number_of_examples()):
data = train_reader.read_example(i)
index = np.array([[i] * data["X"].shape[0]]).T
label = np.array([[data["y"]] * data["X"].shape[0]]).T
tmp = np.concatenate((data["X"], label), axis=1)
out = np.concatenate((index, tmp), axis=1)
train_data.append(out)
for j in range(test_reader.get_number_of_examples()):
data = test_reader.read_example(j)
index = np.array([[i + j] * data["X"].shape[0]]).T
label = np.array([[data["y"]] * data["X"].shape[0]]).T
tmp = np.concatenate((data["X"], label), axis=1)
out = np.concatenate((index, tmp), axis=1)
test_data.append(out)
# Stack training data and testing data
train_data = np.vstack(train_data)
test_data = np.vstack(test_data)
if split:
# Create dataframe
train_df = pd.DataFrame(train_data, index=None, columns=HEADERS)
test_df = pd.DataFrame(test_data, index=None, columns=HEADERS)
# Preprocess coma scales
train_df = preprocess_coma_scales(train_df)
test_df = preprocess_coma_scales(test_df)
return train_df, test_df
else:
# Create dataframe
all_data = np.cat(X)
df = pd.DataFrame(all_data, index=None, columns=HEADERS)
# Preprocess coma scales
df = preprocess_coma_scales(df)
return df
def bbox_transform_inv(cls, bbox, deltas):
bbox_xywh = cls.xxyy2xywh(bbox)
pred_ctr_x = deltas[:, 0] * bbox_xywh[:, 2] + bbox_xywh[:, 0]
pred_ctr_y = deltas[:, 1] * bbox_xywh[:, 3] + bbox_xywh[:, 1]
pred_w = np.exp(deltas[:, 2]) * bbox_xywh[:, 2]
pred_h = np.exp(deltas[:, 3]) * bbox_xywh[:, 3]
pred_ctr_x = pred_ctr_x.view(-1, 1)
pred_ctr_y = pred_ctr_y.view(-1, 1)
pred_w = pred_w.view(-1, 1)
pred_h = pred_h.view(-1, 1)
pred_box = np.cat([pred_ctr_x, pred_ctr_y, pred_w, pred_h], dim=1)
return cls.xywh2xxyy(pred_box)
def step(self, agents):
"""
:param agents:
:return:
"""
for device, agent in zip(self.devices, agents):
stage_state = self.state()
device_state, available_action = device.get_state()
state = np.cat([stage_state, device_state])
action = agent.choose_action(state=state,
available_action=available_action)
device.act(action)
reward = device.get_reward() # 可补充其他奖赏逻辑
agent.learn(state=state,
available_action=available_action,
action=action,
reward=reward)
def get_density(mu, var, pi, N=50, X_range=(0, 5), Y_range=(0, 5)):
""" Get the mesh to compute the density on. """
X = np.linspace(*X_range, N)
Y = np.linspace(*Y_range, N)
X, Y = np.meshgrid(X, Y)
# get the design matrix
points = np.cat([X.reshape(-1, 1), Y.reshape(-1, 1)], axis=1)
points = Variable(torch.from_numpy(points).float())
# compute the densities under each mixture
P = get_k_likelihoods(points, mu, var)
# sum the densities to get mixture density
Z = torch.sum(P, dim=0).data.numpy().reshape([N, N])
return X, Y, Z
def run(self):
trajectory = self.trajectory_generator.trajectory
observation = self.sensor.observe(trajectory)
time = self.time
estimation = np.zeros(shape=(4, len(time)))
estimation[:, 0] = cat((observation[:, 0], np.zeros(2)), axis=0)
self.filter.init(estimation[:, 0])
for k in range(1, len(time)):
y = observation[:, k]
self.filter.update(y)
estimation[:, k] = self.filter.x
self.estimation = estimation
self.observation = observation
def feature_extraction(Z, N): # function Descriptors=feature_extraction(Z,N)
#
#% The features are the magnitude of the Zernike moments.
#% Collect the moments into (2*l+1)-dimensional vectors
#% and define the rotationally invariant 3D Zernike descriptors
#% Fnl as norms of vectors Znl.
#
F = zeros((N + 1, N + 1)) + NaN # F=zeros(N+1,N+1)+NaN;
#% Calcuate the magnitude
for n in xrange(N + 1): # Warn: xrange #for n=0:N
for l in xrange(n + 1): # Warn: xrange # for l=0:n
if mod(n - l, 2) == 0: # if mod(n-l,2)==0
aux_1 = Z[
n, l, 0 : l + 1
] # Warn: 0-1 # aux_1=Z(n+1,l+1,1:l+1); <----- need to keep the trailing l+1 for slicing
# aux_1=aux_1(:)';
if l > 0: # if l>0
# % Taking the values for m<0
aux_2 = conj(
aux_1[1 : l + 1]
) # Warn: 0-1 # aux_2=conj(aux_1(2:(l+1))); <---- Trailing +1 again
for m in xrange(l): # for m=1:l
aux_2[m] = aux_2[m] * power(
-1, m + 1
) # Warn: 0-1 # aux_2(m)=aux_2(m)*power(-1,m);
# end
# % Sorting the vector
aux_2 = flipud(aux_2) # aux_2=rot90(aux_2,2);
aux_1 = cat((aux_2, aux_1), axis=1) # Warn: 0-1 axis # aux_1=cat(2,aux_2,aux_1);
# end
F[n, l] = norm(aux_1, ord=2) # Warn: 0-1 axis # F(n+1,l+1)=norm(aux_1,2);
# end
# end
# end
#% Generate vector of Zernike descriptors
#% Column vector that store the magnitude
#% of the Zernike moments for n,l.
F = transpose(F) # Warn: Matlab col-major
idx = F >= 0 # idx=find(F>=0);
return F[idx]
def compute_targets_and_weights(self, joint_id, coords, data_item, size, scale, batch):
stride = self.cfg.stride
dist_thresh = self.cfg.pos_dist_thresh * scale
num_joints = self.cfg.num_joints
half_stride = stride / 2
scmap = np.zeros(cat([size, arr([num_joints])]))
locref_shape = cat([size, arr([num_joints * 2])])
locref_mask = np.zeros(locref_shape)
locref_map = np.zeros(locref_shape)
pairwise_shape = cat([size, arr([num_joints * (num_joints - 1) * 2])])
pairwise_mask = np.zeros(pairwise_shape)
pairwise_map = np.zeros(pairwise_shape)
dist_thresh_sq = dist_thresh ** 2
width = size[1]
height = size[0]
for person_id in range(len(coords)):
for k, j_id in enumerate(joint_id[person_id]):
joint_pt = coords[person_id][k, :]
j_x = np.asscalar(joint_pt[0])
j_y = np.asscalar(joint_pt[1])
# don't loop over entire heatmap, but just relevant locations
j_x_sm = round((j_x - half_stride) / stride)
j_y_sm = round((j_y - half_stride) / stride)
min_x = round(max(j_x_sm - dist_thresh - 1, 0))
max_x = round(min(j_x_sm + dist_thresh + 1, width - 1))
min_y = round(max(j_y_sm - dist_thresh - 1, 0))
max_y = round(min(j_y_sm + dist_thresh + 1, height - 1))
for j in range(min_y, max_y + 1): # range(height):
pt_y = j * stride + half_stride
for i in range(min_x, max_x + 1): # range(width):
# pt = arr([i*stride+half_stride, j*stride+half_stride])
# diff = joint_pt - pt
# The code above is too slow in python
pt_x = i * stride + half_stride
dx = j_x - pt_x
dy = j_y - pt_y
dist = dx ** 2 + dy ** 2
# print(la.norm(diff))
if dist <= dist_thresh_sq:
dist = dx ** 2 + dy ** 2
locref_scale = 1.0 / self.cfg.locref_stdev
current_normalized_dist = dist * locref_scale ** 2
prev_normalized_dist = locref_map[j, i, j_id * 2 + 0] ** 2 + \
locref_map[j, i, j_id * 2 + 1] ** 2
update_scores = (scmap[j, i, j_id] == 0) or prev_normalized_dist > current_normalized_dist
if self.cfg.location_refinement and update_scores:
self.set_locref(locref_map, locref_mask, locref_scale, i, j, j_id, dx, dy)
if self.cfg.pairwise_predict and update_scores:
for k_end, j_id_end in enumerate(joint_id[person_id]):
if k != k_end:
self.set_pairwise_map(pairwise_map, pairwise_mask, i, j, j_id, j_id_end,
coords, pt_x, pt_y, person_id, k_end)
scmap[j, i, j_id] = 1
scmap_weights = self.compute_scmap_weights(scmap.shape, joint_id, data_item)
# Update batch
batch.update({
Batch.part_score_targets: scmap,
Batch.part_score_weights: scmap_weights
})
if self.cfg.location_refinement:
batch.update({
Batch.locref_targets: locref_map,
Batch.locref_mask: locref_mask
})
if self.cfg.pairwise_predict:
batch.update({
Batch.pairwise_targets: pairwise_map,
Batch.pairwise_mask: pairwise_mask
})
return batch
