hmesh['pnn'] = tools.paintnn(hposd, bs, ncp)
hmesh['target'] = hmesh['pnn'].copy()
print('All the mesh have been generated for seed = %d'%seed)
#Create training voxels
ftlist = [mesh[i].copy() for i in ftname]
ftlistpad = [np.pad(i, pad, 'wrap') for i in ftlist]
targetmesh = hmesh['target']
targetmesh[targetmesh > 1] = 1
ncube = int(ncp/cube_size)
inp = dtools.splitvoxels(ftlistpad, cube_size=cube_sizeft, shift=cube_size, ncube=ncube)
yinp = dtools.splitvoxels(targetmesh, cube_size=cube_size, shift=cube_size, ncube=ncube)
recp = sess.run(output, feed_dict={xx:inp, yy:yinp})
mesh['predict'] = dtools.uncubify(recp[:,:,:,:,0], shape)
meshes[seed] = [mesh, hmesh]
print('All the predictions have been generated for seed = %d'%seed)
##############################
##Power spectrum
kk = tools.fftk(shape, bs)
kmesh = sum(i**2 for i in kk)**0.5
fig, ax = plt.subplots(1, 2, figsize = (10, 4))
for seed in seeds:
predict, hpmeshd = meshes[seed][0]['predict'], meshes[seed][1]['target'],
k, pkpred = tools.power(predict/predict.mean(), boxsize=bs, k=kmesh)
hmesh['pnnsat'] = tools.paintnn(hposd[galtype], bs, ncp)
hmesh['pnncen'] = tools.paintnn(hposd[~galtype], bs, ncp)
print('All the mesh have been generated for seed = %d'%seed)
#Create training voxels
ftlist = [mesh[i].copy() for i in ftname]
ftlistpad = [np.pad(i, pad, 'wrap') for i in ftlist]
targetmesh = [hmesh['pnncen'], hmesh['pnnsat']]
ntarget = len(targetmesh)
ncube = int(ncp/cube_size)
inp = dtools.splitvoxels(ftlistpad, cube_size=cube_sizeft, shift=cube_size, ncube=ncube)
satmesh,cenmesh, rates = sess.run([samplesat, prediction, rate], feed_dict={input:inp, keepprob:1})
mesh['predictsat'] = dtools.uncubify(satmesh[:,:,:,:,0], [nc,nc,nc])
mesh['predictcen'] = dtools.uncubify(cenmesh[:,:,:,:,0], [nc,nc,nc])
mesh['predict'] = mesh['predictcen'] + mesh['predictsat']
mesh['rates'] = dtools.uncubify(rates[:,:,:,:,0], [nc,nc,nc])
meshes[seed] = [mesh, hmesh]
print('Number of predicted & true satellites = ', satmesh.sum(), hmesh['pnnsat'].sum())
##############################
##Power spectrum
kk = tools.fftk(shape, bs)
kmesh = sum(i**2 for i in kk)**0.5
fig, ax = plt.subplots(2, 3, figsize = (12, 8))
for seed in tseeds:
meshs, meshf, meshd = session.run([linmesh, final, samples])
title = session.run([loss, chisq, prior, grad])
np.save(optfolder + '/init.f4', meshs)
dg.makefig(literals['truemeshes'], [meshs, meshf, meshd], optfolder+'%s.png'%('init'), boxsize=bs, title='%s'%title)
optimizer.minimize(session, loss_callback=lcallback, fetches=[[[loss, chisq, prior, grad],
[linmesh, final, samples]]])
meshs, meshf, meshd = session.run([linmesh, final, samples])
title = session.run([loss, chisq, prior, grad])
reconsplit[ii] = meshs
np.save(optfolder + '/recon.f4', meshs)
dg.makefig(literals['truemeshes'], [meshs, meshf, meshd], optfolder+'%s.png'%('recon'), boxsize=bs, title='%s'%title)
initval = dtools.uncubify(np.stack(reconsplit, axis=0), [nc,nc,nc])
recong = rmods.graphhposft1(config, modpath, data, pad, maxiter=maxiter, gtol=gtol, anneal=anneal, resnorm=resnorm)
losses = []
literals = {'losses':losses, 'truemeshes':truemeshes, 'bs':bs, 'nc':nc}
tstart = time()
lcallback = lambda x: loss_callback(x, literals=literals, nprint=nprint, nsave=nsave, maxiter=maxiter, t0=tstart)
#
with tf.Session(graph=recong) as session:
g = session.graph
session.run(tf.global_variables_initializer())
linmesh = g.get_tensor_by_name("linmesh:0")
samples = tf.squeeze(g.get_tensor_by_name("samples:0"))
final = g.get_tensor_by_name("final:0")
optimizer = g.get_collection_ref('opt')[0]
loss = g.get_tensor_by_name('loss:0')
plt.close()
losses.append(l)
shape = (nc, nc, nc)
recon = sess.run(xopt)
print(recon)
plt.figure()
im = plt.imshow(recon[0, :, :, :, 0].sum(axis=0))
plt.colorbar(im)
plt.savefig('./figs/n%02d/recon2single%s.png' % (numd * 1e4, suff))
print(recon.mean(), recon.std())
outtruth = sess.run(output, feed_dict={xx: cube_features, yy: cube_target})
outrecon = sess.run(output, feed_dict={xx: recon, yy: cube_target})
recon = dtools.uncubify(recon[:, 2:34, 2:34, 2:34, 0], shape)
outtruth = dtools.uncubify(outtruth[:, :, :, :, 0], shape)
outrecon = dtools.uncubify(outrecon[:, :, :, :, 0], shape)
reconmapp = dtools.uncubify(cube_target[:, :, :, :, 0], shape)
recontruth = dtools.uncubify(cube_features[:, 2:34, 2:34, 2:34, 0], shape)
def getim(ar, axis=0):
#return ar[0, :, :, :, 0].sum(axis=axis)
return ar[:, :, :].sum(axis=axis)
fig, axar = plt.subplots(3, 3, figsize=(18, 18))
fsize = 16
#Create training voxels
ftlist = [mesh[i].copy() for i in ftname]
ftlistpad = [np.pad(i, pad, 'wrap') for i in ftlist]
targetmesh = hmesh['target']
targetmesh[targetmesh > 1] = 1
for size in test_sizes:
ncube = int(ncp / size)
print(size, ncube)
inp = dtools.splitvoxels(ftlistpad,
cube_size=size + 2 * pad,
shift=size,
ncube=ncube)
recp = sess.run(prediction, feed_dict={input: inp, keepprob: 1})
mesh['predict%03d' % size] = dtools.uncubify(recp[:, :, :, :, 0],
shape)
meshes[seed] = [mesh, hmesh]
print('Prediction done for seed = %d' % seed)
##############################
##Power spectrum
kk = tools.fftk(shape, bs)
kmesh = sum(i**2 for i in kk)**0.5
lss = ['-', '--', ':', '-.']
fig, ax = plt.subplots(1, 2, figsize=(10, 4))
for ss, seed in enumerate(seeds):
hpmeshd = meshes[seed][1]['target']
k, pkhd = tools.power(hpmeshd / hpmeshd.mean(), boxsize=bs, k=kmesh)