- html - 出于某种原因,IE8 对我的 Sass 文件中继承的 html5 CSS 不友好?
- JMeter 在响应断言中使用 span 标签的问题
- html - 在 :hover and :active? 上具有不同效果的 CSS 动画
- html - 相对于居中的 html 内容固定的 CSS 重复背景?
我正在将 IDL 代码(由 Oleg Kochukhov 编写)转换为 Python。该代码使用吉洪诺夫或最大熵方法在谱线剖面上生成星体表面图。
我使用 scipy.optimize.minimize 生成线条轮廓图。但过程太慢且结果不兼容。我在互联网上搜索解决方案,但没有找到任何有用的解决方案。
我在下面添加了可运行的代码:
import numpy as np
from scipy.optimize import minimize
import matplotlib.pyplot as plt
from mpl_toolkits.basemap import Basemap
import matplotlib.gridspec as gridspec
#syc = 0
def DI_GridInit(ntot):
# generate stellar surface grid
nlat = int(round(0.5 * (1.0 + np.sqrt(1.0 + np.pi * ntot))) - 1)
nlon = np.zeros(nlat, dtype=int)
xlat = np.pi * (np.arange(nlat, dtype=float) + 0.5) / nlat - np.pi / 2.0
xcirc = 2.0 * np.cos(xlat[1:])
nlon[1:] = np.around(xcirc * nlat) + 1
nlon[0] = ntot - sum(nlon[1:])
if abs(nlon[0] - nlon[nlat - 1]) > nlat:
nlon[1:] = nlon[1:] + (nlon[0] - nlon[nlat - 1]) / nlat
nlon[0] = ntot - sum(nlon[1:])
if nlon[0] < nlon[nlat - 1]:
nlon[1:] = nlon[1:] - 1
nlon[0] = ntot - sum(nlon[1:])
# generate Descartes coordinates for the surface grid in
# stellar coordinates, areas of surface elements and
# regularization indices: (lower, upper, right, left)
x0, j = np.zeros((ntot, 3), dtype=float), 0
latitude, longitude = np.zeros(ntot, dtype=float), np.zeros(ntot, dtype=float)
sa, ireg = np.zeros(ntot, dtype=float), np.zeros((ntot, 4), dtype=int)
slt = np.hstack((0., (xlat[1:nlat] + xlat[0:nlat - 1]) / 2. + np.pi / 2., np.pi))
for i in range(nlat):
coslat = np.cos(xlat[i])
sinlat = np.sin(xlat[i])
xlon = 2 * np.pi * (np.arange(nlon[i]) + 0.5) / nlon[i]
sinlon = np.sin(xlon)
coslon = np.cos(xlon)
x0[:, 0][j:j + nlon[i]] = coslat * sinlon
x0[:, 1][j:j + nlon[i]] = -coslat * coslon
x0[:, 2][j:j + nlon[i]] = sinlat
latitude[j:j + nlon[i]] = xlat[i]
longitude[j:j + nlon[i]] = xlon
sa[j:j + nlon[i]] = 2. * np.pi * (np.cos(slt[i]) - np.cos(slt[i + 1])) / nlon[i]
ireg[:, 2][j:j + nlon[i]] = np.roll(j + np.arange(nlon[i], dtype=int), -1)
ireg[:, 3][j:j + nlon[i]] = np.roll(j + np.arange(nlon[i], dtype=int), 1)
if (i > 0):
il_lo = j - nlon[i - 1] + np.arange(nlon[i - 1], dtype=int)
else:
il_lo = j + nlon[i] + np.arange(nlon[i + 1], dtype=int)
if (i < nlat - 1):
il_up = j + nlon[i] + np.arange(nlon[i + 1], dtype=int)
else:
il_up = il_lo
for k in range(j, j + nlon[i]):
dlat_lo = longitude[k] - longitude[il_lo]
ll = np.argmin(abs(dlat_lo))
ireg[k][0] = il_lo[ll]
dlat_up = longitude[k] - longitude[il_up]
ll = np.argmin(abs(dlat_up))
ireg[k][1] = il_up[ll]
j += nlon[i]
theta = np.arccos(x0[:, 2])
phi = np.arctan2(x0[:, 0], -x0[:, 1])
ii = np.argwhere(phi < 0).T[0]
nii = len(ii)
phi[ii] = 2.0 * np.pi - abs(phi[ii]) if nii else None
grid = {'ntot': ntot, 'nlat': nlat, 'nlon': nlon, 'xyz': x0, 'lat': latitude,
'lon': longitude, 'area': sa, 'ireg': ireg, 'phi': phi, 'theta': theta}
return grid
def DI_Map(grid, spots):
map = np.ones(grid['ntot'], dtype=float)
for i in range(spots['n']):
dlon = grid['lon'] - np.deg2rad(spots['tbl'][i, 0])
dlat = grid['lat'] - np.deg2rad(spots['tbl'][i, 1])
da = (2.0 * np.arcsin(np.sqrt(np.sin(0.5 * dlat) ** 2 +
np.cos(np.deg2rad(spots['tbl'][i, 1])) *
np.cos(grid['lat']) * np.sin(0.5 * dlon) ** 2)))
ii = np.argwhere(da <= np.deg2rad(spots['tbl'][i, 2])).T[0]
ni = len(ii)
map[ii] = spots['tbl'][i, 3] if ni > 0 else None
return map
def DI_Prf(grid, star, map, phase=None, vv=None, vr=None, nonoise=None):
# velocity array
if vv is not None:
nv = len(vv)
else:
nv = int(np.ceil(2.0 * star['vrange'] / star['vstep']))
vv = -star['vrange'] + np.arange(nv, dtype=float) * star['vstep']
# phase array
if phase is None:
phase = np.arange(star['nphases'], dtype=float) / star['nphases']
# velocity correction for each phase
vr = np.zeros(star['nphases'], dtype=float) if vr == None else None
# fixed trigonometric quantities
cosi = np.cos(np.deg2rad(star['incl'])); sini = np.sin(np.deg2rad(star['incl']))
coslat = np.cos(grid['lat']); sinlat = np.sin(grid['lat'])
# FWHM to Gaussian sigma
sigm = star['fwhm'] / np.sqrt(8.0 * np.log(2.0))
isig = (-0.5 / sigm ** 2)
# initialize line profile and integrated field arrays
prf = np.zeros((nv, len(phase)), dtype=float)
# gradient if called with 5 - variable input
grad = np.zeros((nv, len(phase), grid['ntot']), dtype=float)
# phase loop
for i in range(len(phase)):
coslon = np.cos(grid['lon'] + 2.0 * np.pi * phase[i])
sinlon = np.sin(grid['lon'] + 2.0 * np.pi * phase[i])
mu = sinlat * cosi + coslat * sini * coslon
ivis = np.argwhere(mu > 0.).T[0]
dv = -sinlon[ivis] * coslat[ivis] * star['vsini']
avis = grid['area'][ivis] * mu[ivis] * (1.0 - star['limbd'] + star['limbd'] * mu[ivis])
if star['type'] == 0:
wgt = avis * map[ivis]
wgtn = sum(wgt)
for j in range(nv):
plc = 1.0 - star['d'] * np.exp(isig * (vv[j] + dv - vr[i]) ** 2)
prf[j][i] = sum(wgt * plc) / wgtn
grad[j][i][ivis] = avis * plc / wgtn - avis * prf[j][i] / wgtn
elif star['type'] == 1:
wgt = avis
wgtn = sum(wgt)
for j in range(nv):
plc = 1.0 - map[ivis] * star['d'] * np.exp(isig * (vv[j] + dv - vr[i]) ** 2)
prf[j][i] = sum(wgt * plc) / wgtn
grad[j][i][ivis] = -wgt / wgtn * star['d'] * np.exp(isig * (vv[j] + dv - vr[i]) ** 2)
# output structure
syn = {'v': vv, 'phase': phase, 'prf': prf}
# add noise
if star['snr'] != -1 and nonoise != None:
obs = syn['prf'] * 0.0
for i in range(star['nphases']):
obs[:, i] = syn['prf'][:, i] + np.random.standard_normal((len(syn['v']),)) / star['snr']
syn['obs'] = obs
return syn, grad
def DI_func(cmap, functargs):
# global syc
star = functargs['star']
grid = functargs['grid']
obs = functargs['obs']
invp = functargs['invp']
nv = len(obs['v'])
er = 1.0 / abs(star['snr'])
if 'vr' in obs.keys():
syn, grad = DI_Prf(grid, star, cmap, phase=obs['phase'], vv=obs['v'], vr=obs['vr'])
else:
syn, grad = DI_Prf(grid, star, cmap, phase=obs['phase'], vv=obs['v'])
# shf = 0
# for i in range(len(obs['phase'])):
# plt.plot(obs['v'], obs['obs'][:, i] + shf, 'bo')
# plt.plot(obs['v'], syn['prf'][:, i] + shf, 'r')
# plt.plot(obs['v'], obs['obs'][:, i] - syn['prf'][:, i] + shf, 'k')
# shf += 0.1
# plt.show()
fchi = 0.0
sign = (-1) ** invp['regtype']
for i in range(star['nphases']):
fchi = fchi + sign * sum((syn['prf'][:, i] - obs['obs'][:, i]) ** 2 / er ** 2) / nv
freg = 0
if invp['lambda'] > 0:
if invp['regtype'] == 0:
ir = grid['ireg']
for k in range(len(ir[0, :])):
freg = freg + invp['lambda'] / grid['ntot'] * sum((cmap - cmap[ir[:, k]]) ** 2)
elif invp['regtype'] == 1:
mmap = sum(cmap) / grid['ntot']
nmap = cmap / mmap
freg = freg - invp['lambda'] / grid['ntot'] * sum(nmap * np.log(nmap))
ftot = fchi + freg
syn['obs'] = obs['obs']
# syc += 1
# if syc % 1000 == 0:
# plotting(grid, cmap, syn, star['incl'], typ=star['type'])
#
# print(syc, ftot, sum(cmap))
return ftot
def plotting(grid, map, syn, incl, typ):
nlon = grid['nlon']
nln = max(nlon)
nlt = len(nlon)
ll = np.zeros(nlt + 1, dtype=int)
ll[0] = 0
for i in range(nlt):
ll[i + 1] = ll[i] + nlon[i]
map1 = np.zeros((nlt, nln), dtype=float)
x = np.arange(nln, dtype=float) + 0.5
for i in range(nlt):
lll = ((np.arange(nlon[i] + 2, dtype=float) - 0.5) * nln) / nlon[i]
y = np.hstack((map[ll[i + 1] - 1], map[ll[i]:ll[i+1]-1], map[ll[i]]))
for j in range(nln):
imin = np.argmin(abs(x[j] - lll))
map1[i, j] = y[imin]
light = (190 * (map1 - np.min(map1)) / (np.max(map1) - np.min(map1))) + 50
light_rect = np.flipud(light)
if typ == 0:
cmap = 'gray'
else:
cmap = 'gray_r'
fig = plt.figure()
fig.clear()
spec = gridspec.GridSpec(ncols=3, nrows=3, left=0.10, right=0.98,
top=0.97, bottom=0.07, hspace=0.2, wspace=0.36)
# naive IDW-like interpolation on regular grid
shape = light.shape
nrows, ncols = (shape[0], shape[1])
lon, lat = np.meshgrid(np.linspace(0, 360, ncols), np.linspace(-90, 90, nrows))
for i, item in enumerate([[(0, 0), -0], [(0, 1), -90], [(1, 0,), -180], [(1, 1), -270]]):
ax = fig.add_subplot(spec[item[0]])
# set up map projection
m = Basemap(projection='ortho', lat_0=90 - incl, lon_0=item[1], ax=ax)
# draw lat/lon grid lines every 30 degrees.
m.drawmeridians(np.arange(0, 360, 30))
m.drawparallels(np.arange(-90, 90, 30))
# compute native map projection coordinates of lat/lon grid.
x, y = m(lon, lat)
# contour data over the map.
m.contourf(x, y, light, 15, vmin=0., vmax=255., cmap=cmap)
if i in [0, 2]:
x2, y2 = m(180 - item[1], incl)
else:
x2, y2 = m(180 + item[1], incl)
x1, y1 = (-10, 5)
ax.annotate(str('%0.2f' % (abs(item[1]) / 360.)), xy=(x2, y2), xycoords='data',
xytext=(x1, y1), textcoords='offset points',
color='r')
ax5 = fig.add_subplot(spec[-1, :2])
ax5.imshow(light_rect, vmin=0., vmax=255., cmap=cmap, interpolation='none', extent=[0, 360, -90, 90])
ax5.set_xticks(np.arange(0, 420, 60))
ax5.set_yticks(np.arange(-90, 120, 30))
ax5.set_xlabel('Longitude ($^\circ$)', fontsize=7)
ax5.set_ylabel('Latitude ($^\circ$)', fontsize=7)
ax5.tick_params(labelsize=7)
ax6 = fig.add_subplot(spec[0:, 2])
shf = 0.0
for i in range(len(syn['phase'])):
ax6.plot(syn['v'], syn['obs'][:, -i - 1] + shf, 'bo', ms=2)
ax6.plot(syn['v'], syn['prf'][:, -i - 1] + shf, 'r', linewidth=1)
ax6.text(min(syn['v']), max(syn['obs'][:, -i - 1] + shf), str('%0.2f' % syn['phase'][-i - 1]),
fontsize=7)
shf += 0.1
p1 = ax6.lines[0]
p2 = ax6.lines[-1]
p1datay = p1.get_ydata()
p1datax = p1.get_xdata()
p2datay = p2.get_ydata()
y1, y2 = min(p1datay) - min(p1datay) / 20.,max(p2datay) + min(p1datay) / 10.
ax6.set_ylim([y1, y2])
ax6.set_xlabel('V ($km s^{-1}$)', fontsize=7)
ax6.set_ylabel('I / Ic', fontsize=7)
ax6.tick_params(labelsize=7)
max_ = int(max(p1datax))
ax6.set_xticks([-max_, np.floor(-max_ / 2.), 0.0, np.ceil(max_ / 2.), max_])
plt.show()
if __name__ == "__main__":
# Star parameters
star = {'ntot': 1876, 'type': 0, 'incl': 70, 'vsini': 50, 'fwhm': 7.0, 'd': 0.6,
'limbd': 0.5, 'nphases': 5, 'vrange': np.sqrt(50 ** 2 + 7.0 ** 2) * 1.4,
'vstep': 1.0, 'snr': 500}
# Spot parameters
lon_spot = [40, 130, 220, 310]
lat_spot = [-30, 0, 60, 30]
r_spot = [20, 20, 20, 20]
c_spot = [0.1, 0.2, 0.25, 0.3]
tbl = np.array([lon_spot, lat_spot, r_spot, c_spot]).T
spots = {'n': len(lon_spot), 'type': star['type'], 'tbl': tbl}
# Generate grid
grid = DI_GridInit(star['ntot'])
# Generate map
cmap = DI_Map(grid, spots)
# Generate spectral line profiles
csyn, grad = DI_Prf(grid, star, cmap, nonoise=True)
# Plotting map and line profiles
plotting(grid, cmap, csyn, star['incl'], star['type'])
# Generate map over the line profiles using scipy.optimize.minimize
invp = {'lambda': 20, 'regtype': 0, 'maxiter': 10}
grid_inv = DI_GridInit(star['ntot'])
functargs = {'star': star, 'grid': grid_inv, 'obs': csyn, 'invp': invp}
cmap = np.ones(star['ntot'])
cmap[0] = 0.99
bnd = list(zip(np.zeros(len(cmap), dtype=float), np.ones(len(cmap), dtype=float)))
minimize(DI_func, cmap, args=functargs, method='TNC', bounds=bnd,
callback=None, options={'eps': 0.1, 'maxiter': 5, 'disp': True})
代码包括以下部分。
'DI_GridInit':为 map 生成网格
'DI_Map':根据星点参数(如经度、纬度、半径、对比度)生成星面图
'DI_Prf':根据 map 生成谱线剖面
现在我想获取生成的和加噪的线轮廓上的表面图。我使用 scipy.optimize.minimize (TNC 方法)来获取表面图。我使用“DI_func”作为最小化中的函数。但“最小化”太慢了。问题是什么。我怎样才能加快这个速度。
最佳答案
这是DI_Prf
的修改版本,其中是DI_func
执行期间的主要计算时间:
def DI_Prf(grid, star, map, phase=None, vv=None, vr=None, nonoise=None):
# velocity array
if vv is not None:
nv = len(vv)
else:
nv = int(np.ceil(2.0 * star['vrange'] / star['vstep']))
vv = -star['vrange'] + np.arange(nv, dtype=float) * star['vstep']
# phase array
if phase is None:
phase = np.arange(star['nphases'], dtype=float) / star['nphases']
# velocity correction for each phase
vr = np.zeros(star['nphases'], dtype=float) if vr == None else None
# fixed trigonometric quantities
cosi = np.cos(np.deg2rad(star['incl'])); sini = np.sin(np.deg2rad(star['incl']))
coslat = np.cos(grid['lat']); sinlat = np.sin(grid['lat'])
# FWHM to Gaussian sigma
sigm = star['fwhm'] / np.sqrt(8.0 * np.log(2.0))
isig = (-0.5 / sigm ** 2)
# initialize line profile and integrated field arrays
prf = np.zeros((nv, len(phase)), dtype=float)
# gradient if called with 5 - variable input
grad = np.zeros((nv, len(phase), grid['ntot']), dtype=float)
# phase loop
for i in range(len(phase)):
coslon = np.cos(grid['lon'] + 2.0 * np.pi * phase[i])
sinlon = np.sin(grid['lon'] + 2.0 * np.pi * phase[i])
mu = sinlat * cosi + coslat * sini * coslon
ivis = np.argwhere(mu > 0.).T[0]
dv = -sinlon[ivis] * coslat[ivis] * star['vsini']
avis = grid['area'][ivis] * mu[ivis] * (1.0 - star['limbd'] + star['limbd'] * mu[ivis])
if star['type'] == 0:
wgt = avis * map[ivis]
wgtn = sum(wgt)
#for j in range(nv):
# plc = 1.0 - star['d'] * np.exp(isig * (vv[j] + dv - vr[i]) ** 2)
# prf[j][i] = sum(wgt * plc) / wgtn
# grad[j][i][ivis] = avis * plc / wgtn - avis * prf[j][i] / wgtn
plc = 1.0 - star['d'] * np.exp(isig * (vv[:, np.newaxis] + dv[np.newaxis, :] - vr[i]) ** 2)
prf[:, i] = np.sum(wgt * plc, axis=1) / wgtn
grad[:, i, ivis] = avis * plc / wgtn - (avis[:, np.newaxis]*prf[:, i]).T / wgtn
elif star['type'] == 1:
wgt = avis
wgtn = sum(wgt)
for j in range(nv): # to be modified too
plc = 1.0 - map[ivis] * star['d'] * np.exp(isig * (vv[j] + dv - vr[i]) ** 2)
prf[j][i] = sum(wgt * plc) / wgtn
grad[j][i][ivis] = -wgt / wgtn * star['d'] * np.exp(isig * (vv[j] + dv - vr[i]) ** 2)
# output structure
syn = {'v': vv, 'phase': phase, 'prf': prf}
# add noise
if star['snr'] != -1 and nonoise != None:
#for i in range(star['nphases']):
obs = syn['prf'] + np.random.standard_normal(size=syn['prf'].shape) / star['snr']
syn['obs'] = obs
return syn, grad
时间减少了 3:
%%timeit
syn, grad = DI_Prf(grid, star, cmap, phase=obs['phase'], vv=obs['v'])
# 127 ms ± 2.61 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
# 40.7 ms ± 683 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
Numpy 的主要思想是不使用循环,而是使用多维数组,并使用广播功能。
例如:
fchi = 0.0
for i in range(star['nphases']):
fchi = fchi + sign * sum((syn['prf'][:, i] - obs['obs'][:, i]) ** 2 / er ** 2) / nv
可以替换为:
fchi = sign / nv / er ** 2 * np.sum( np.sum((syn['prf'] - obs['obs']) ** 2, axis=1 ) )
与np.random.standard_normal(size=syn['prf'].shape)相同
这里并不是一个很大的改进,因为star['nphases']
很小,但对于另一个轴来说相对重要。您可以进一步删除 DI_Prf
中各阶段的 for 循环,但这需要一些思考
关于python - scipy.optimize.minimize 太慢了。我怎样才能加快速度,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/51988141/
我在想出一个算法时遇到了麻烦... 我有一系列 GPS 数据,以 1 秒为间隔记录时间、速度、距离。假设距离是米,速度是米/秒。可能有超过 2 小时的数据,或 7200 个点。这里的“时间”字段主要是
使用java排序器,即: Collections.sort(myArrayList, new Comparator() { @Override public int c
有什么区别吗 SELECT * FROM my_table 和 SELECT my_column_id FROM my_table 地点: my_table 有百万行 网站上有大量并发用户进行sql查
有2个样本。 在第一个示例中,使用 orderby 可以更快地获得结果。 (根据 phpmyadmin 速度报告) 在另一个例子中,我没有使用 order by,它给出的结果较慢。 (根据 phpmy
我注意到,如果我将训练数据加载到内存中并将其作为 numpy 数组提供到图中,与使用相同大小的 shuffle 批次相比,速度会有很大差异,我的数据有大约 1000 个实例。 使用内存 1000 次迭
我在 python 中使用破折号。我正在绘制记录到 SQLite 数据库中的实时数据,目前,我正在绘制单个值与时间线图。我计划再添加 20 个图表,但目前,随着时间的增加, plotly 变慢,我认为
我试图调用 hasNext Velocity 模板中的方法,以便根据 foreach 循环中的位置影响行为 - 仅 hasNext没有按照文档工作。 这是 Velocity 用户指南的片段,关于 ha
在我正在制作的游戏中,我有两个点,pt1 和 pt2,我想计算出它们之间的角度。我已经在较早的计算中计算出距离。显而易见的方法是对垂直距离上的水平距离进行反正切 (tan(theta) = opp/a
我经常遇到字符串值不存在和/或为空的情况。这是测试这种情况的最佳方法吗? #if( $incentive.disclaimer && $!incentive.disclaimer != '' )
我想将一个模板nested包含在其他模板cont1,cont2和cont3中。 并且嵌套模板应仅对cont1隐藏一个特定控件。 在包含在cont1中之前,我想为一些标志变量$hideMyControl
是否可以更改从“Windows Azure Media Encoder”输出的音频的播放速度? 我正在使用配置为“WMA High Quality Audio”的“Windows Azure Medi
我使用速度将String(template)与字段合并 hi there I'am ${name}, And I'am ${age} old. velocity将字段${name}和${age}与一种
我使用的是 LockedBitmap 类,它简化了 C# 中位图数据的处理。目前它正在将数据复制到本地 byte[] 数组中,然后通过其类方法访问该数组以获取/设置像素颜色值。 这比直接通过指针访问锁
我尝试在 VM_global_library.vm 文件中添加一堆 #set($x=abc) 语句,但这些变量在我的 VM 模板中不可用。 我想为图像的基本路径等设置一个全局变量。这可能吗? 最佳答案
我的项目结构: -src --main ---java ----makers -----SomeClass ---resources ----htmlPattern.vm 如何告诉 SomeClass
我正在尝试从 Velocity 中的字符串中删除不需要的字符(换行符可以,但不能像 EM 和 CAN ASCII 控制字符那样)。 #set($cleanScreen = $cleanScreen.r
我想在日.月.年之间的点处分割日期。例如:2015 年 1 月 14 日至 {14, 01, 2015}这是我使用的代码:dates3.get(0) 包含我从页面的文本字段获取的字符串“14.01.2
之后,从 1.5 升级到速度引擎 1.7 出现了 1.5 没有的问题。为了解释这个问题,我必须展示一个代码片段: #foreach($someVariable in $someCollection)
我想知道从表中选择所有字段是否更快: SELECT * 或只选择您真正需要的: SELECT field1, field2, field3, field4, field5... 假设表有大约 10 个
我正在尝试模仿照片应用程序的行为,在该应用程序中,用户用手指平移照片并且照片具有一定的速度。由于我不会深入的原因,我不能将 UIScrollView 与它的缩放 UIImageView 一起使用,而是
我是一名优秀的程序员,十分优秀!