def show_density_plot(self, cfig):
cdiag = self.diag
cparams = cdiag['params']
name = cparams['name']
maxfreq = cparams['max_frequency']
resolution = cparams['voxel_size']
prior = eval_objective(cparams['prior'])
prior.set_params(cparams)
N = self.M.shape[0]
rad_cutoff = cparams.get('rad_cutoff', 1.0)
rad = min(rad_cutoff, maxfreq * 2.0 * resolution)
# Statistics of M
plt.figure(cfig.number)
plt.clf()
plt.suptitle(name + ' Density Statistics')
nHistBins = 0.5 * self.M.shape[0]
logprobScale = np.log(self.M.size / nHistBins)
plt.subplot(2, 1, 1)
plt.hist(self.M.reshape((-1, )), bins=nHistBins, log=True)
histxLims = plt.xlim()
histyLims = plt.ylim()
vals = np.linspace(histxLims[0], histxLims[1], 1000)
plt.plot(vals, np.exp(logprobScale - prior.scalar_eval(vals)))
plt.xlim(histxLims)
plt.ylim(histyLims)
plt.title('Voxel Histogram + Prior')
plt.subplot(2, 2, 3)
plt.hist(np.absolute(self.fM).reshape((-1, )),
bins=nHistBins,
log=True)
plt.title('Power Histogram')
(fs, raps) = rot_power_spectra(self.fM, resolution=resolution)
plt.subplot(2, 2, 4)
plt.plot(fs / (N / 2.0) / (2.0 * resolution), raps, label='RAPS')
plt.plot((rad / (2.0 * resolution)) * np.ones((2, )),
np.array([raps[raps > 0].min(),
raps.max()]))
plt.yscale('log')
plt.title('Rotationally Averaged Power Spectra')
def show_density_plot(self,cfig):
cdiag = self.diag
cparams = cdiag['params']
name = cparams['name']
maxfreq = cparams['max_frequency']
resolution = cparams['voxel_size']
prior = eval_objective(cparams['prior'])
prior.set_params(cparams)
N = self.M.shape[0]
rad_cutoff = cparams.get('rad_cutoff', 1.0)
rad = min(rad_cutoff,maxfreq*2.0*resolution)
# Statistics of M
plt.figure(cfig.number)
plt.clf()
plt.suptitle(name + ' Density Statistics')
nHistBins = 0.5*self.M.shape[0]
logprobScale = n.log(self.M.size/nHistBins)
plt.subplot(2,1,1)
plt.hist(self.M.reshape((-1,)),bins=nHistBins,log=True)
histxLims = plt.xlim()
histyLims = plt.ylim()
vals = n.linspace(histxLims[0],histxLims[1],1000)
plt.plot(vals,n.exp(logprobScale-prior.scalar_eval(vals)))
plt.xlim(histxLims)
plt.ylim(histyLims)
plt.title('Voxel Histogram + Prior')
plt.subplot(2,2,3)
plt.hist(n.absolute(self.fM).reshape((-1,)),bins=nHistBins,log=True)
plt.title('Power Histogram')
(fs,raps) = rot_power_spectra(self.fM,resolution=resolution)
plt.subplot(2,2,4)
plt.plot(fs/(N/2.0)/(2.0*resolution),raps,label='RAPS')
plt.plot((rad/(2.0*resolution))*n.ones((2,)),
n.array([raps[raps>0].min(),raps.max()]))
plt.yscale('log')
plt.title('Rotationally Averaged Power Spectra')
def __init__(self, expbase, cmdparams=None):
"""cryodata is a CryoData instance.
expbase is a path to the base of folder where this experiment's files
will be stored. The folder above expbase will also be searched
for .params files. These will be loaded first."""
BackgroundWorker.__init__(self)
# Create a background thread which handles IO
self.io_queue = Queue()
self.io_thread = Thread(target=self.ioworker)
self.io_thread.daemon = True
self.io_thread.start()
# General setup ----------------------------------------------------
self.expbase = expbase
self.outbase = None
# Paramter setup ---------------------------------------------------
# search above expbase for params files
_,_,filenames = os.walk(opj(expbase,'../')).next()
self.paramfiles = [opj(opj(expbase,'../'), fname) \
for fname in filenames if fname.endswith('.params')]
# search expbase for params files
_,_,filenames = os.walk(opj(expbase)).next()
self.paramfiles += [opj(expbase,fname) \
for fname in filenames if fname.endswith('.params')]
if 'local.params' in filenames:
self.paramfiles += [opj(expbase,'local.params')]
# load parameter files
self.params = Params(self.paramfiles)
self.cparams = None
if cmdparams is not None:
# Set parameter specified on the command line
for k,v in cmdparams.iteritems():
self.params[k] = v
# Dataset setup -------------------------------------------------------
self.imgpath = self.params['inpath']
psize = self.params['resolution']
if not isinstance(self.imgpath,list):
imgstk = MRCImageStack(self.imgpath,psize)
else:
imgstk = CombinedImageStack([MRCImageStack(cimgpath,psize) for cimgpath in self.imgpath])
if self.params.get('float_images',True):
imgstk.float_images()
self.ctfpath = self.params['ctfpath']
mscope_params = self.params['microscope_params']
if not isinstance(self.ctfpath,list):
ctfstk = CTFStack(self.ctfpath,mscope_params)
else:
ctfstk = CombinedCTFStack([CTFStack(cctfpath,mscope_params) for cctfpath in self.ctfpath])
self.cryodata = CryoDataset(imgstk,ctfstk)
self.cryodata.compute_noise_statistics()
if self.params.get('window_images',True):
imgstk.window_images()
minibatch_size = self.params['minisize']
testset_size = self.params['test_imgs']
partition = self.params.get('partition',0)
num_partitions = self.params.get('num_partitions',1)
seed = self.params['random_seed']
if isinstance(partition,str):
partition = eval(partition)
if isinstance(num_partitions,str):
num_partitions = eval(num_partitions)
if isinstance(seed,str):
seed = eval(seed)
self.cryodata.divide_dataset(minibatch_size,testset_size,partition,num_partitions,seed)
self.cryodata.set_datasign(self.params.get('datasign','auto'))
if self.params.get('normalize_data',True):
self.cryodata.normalize_dataset()
self.voxel_size = self.cryodata.pixel_size
# Iterations setup -------------------------------------------------
self.iteration = 0
self.tic_epoch = None
self.num_data_evals = 0
self.eval_params()
outdir = self.cparams.get('outdir',None)
if outdir is None:
if self.cparams.get('num_partitions',1) > 1:
outdir = 'partition{0}'.format(self.cparams['partition'])
else:
outdir = ''
self.outbase = opj(self.expbase,outdir)
if not os.path.isdir(self.outbase):
os.makedirs(self.outbase)
# Output setup -----------------------------------------------------
self.ostream = OutputStream(opj(self.outbase,'stdout'))
self.ostream(80*"=")
self.ostream("Experiment: " + expbase + \
" Kernel: " + self.params['kernel'])
self.ostream("Started on " + socket.gethostname() + \
" At: " + time.strftime('%B %d %Y: %I:%M:%S %p'))
self.ostream("Git SHA1: " + gitutil.git_get_SHA1())
self.ostream(80*"=")
gitutil.git_info_dump(opj(self.outbase, 'gitinfo'))
self.startdatetime = datetime.now()
# for diagnostics and parameters
self.diagout = Output(opj(self.outbase, 'diag'),runningout=False)
# for stats (per image etc)
self.statout = Output(opj(self.outbase, 'stat'),runningout=True)
# for likelihoods of individual images
self.likeout = Output(opj(self.outbase, 'like'),runningout=False)
self.img_likes = n.empty(self.cryodata.N_D)
self.img_likes[:] = n.inf
# optimization state vars ------------------------------------------
init_model = self.cparams.get('init_model',None)
if init_model is not None:
filename = init_model
if filename.upper().endswith('.MRC'):
M = readMRC(filename)
else:
with open(filename) as fp:
M = cPickle.load(fp)
if type(M)==list:
M = M[-1]['M']
if M.shape != 3*(self.cryodata.N,):
M = cryoem.resize_ndarray(M,3*(self.cryodata.N,),axes=(0,1,2))
else:
init_seed = self.cparams.get('init_random_seed',0) + self.cparams.get('partition',0)
print "Randomly generating initial density (init_random_seed = {0})...".format(init_seed), ; sys.stdout.flush()
tic = time.time()
M = cryoem.generate_phantom_density(self.cryodata.N, 0.95*self.cryodata.N/2.0, \
5*self.cryodata.N/128.0, 30, seed=init_seed)
print "done in {0}s".format(time.time() - tic)
tic = time.time()
print "Windowing and aligning initial density...", ; sys.stdout.flush()
# window the initial density
wfunc = self.cparams.get('init_window','circle')
cryoem.window(M,wfunc)
# Center and orient the initial density
cryoem.align_density(M)
print "done in {0:.2f}s".format(time.time() - tic)
# apply the symmetry operator
init_sym = get_symmetryop(self.cparams.get('init_symmetry',self.cparams.get('symmetry',None)))
if init_sym is not None:
tic = time.time()
print "Applying symmetry operator...", ; sys.stdout.flush()
M = init_sym.apply(M)
print "done in {0:.2f}s".format(time.time() - tic)
tic = time.time()
print "Scaling initial model...", ; sys.stdout.flush()
modelscale = self.cparams.get('modelscale','auto')
mleDC, _, mleDC_est_std = self.cryodata.get_dc_estimate()
if modelscale == 'auto':
# Err on the side of a weaker prior by using a larger value for modelscale
modelscale = (n.abs(mleDC) + 2*mleDC_est_std)/self.cryodata.N
print "estimated modelscale = {0:.3g}...".format(modelscale), ; sys.stdout.flush()
self.params['modelscale'] = modelscale
self.cparams['modelscale'] = modelscale
M *= modelscale/M.sum()
print "done in {0:.2f}s".format(time.time() - tic)
if mleDC_est_std/n.abs(mleDC) > 0.05:
print " WARNING: the DC component estimate has a high relative variance, it may be inaccurate!"
if ((modelscale*self.cryodata.N - n.abs(mleDC)) / mleDC_est_std) > 3:
print " WARNING: the selected modelscale value is more than 3 std devs different than the estimated one. Be sure this is correct."
self.M = n.require(M,dtype=density.real_t)
self.fM = density.real_to_fspace(M)
self.dM = density.zeros_like(self.M)
self.step = eval(self.cparams['optim_algo'])
self.step.setup(self.cparams, self.diagout, self.statout, self.ostream)
# Objective function setup --------------------------------------------
param_type = self.cparams.get('parameterization','real')
cplx_param = param_type in ['complex','complex_coeff','complex_herm_coeff']
self.like_func = eval_objective(self.cparams['likelihood'])
self.prior_func = eval_objective(self.cparams['prior'])
if self.cparams.get('penalty',None) is not None:
self.penalty_func = eval_objective(self.cparams['penalty'])
prior_func = SumObjectives(self.prior_func.fspace, \
[self.penalty_func,self.prior_func], None)
else:
prior_func = self.prior_func
self.obj = SumObjectives(cplx_param,
[self.like_func,prior_func], [None,None])
self.obj.setup(self.cparams, self.diagout, self.statout, self.ostream)
self.obj.set_dataset(self.cryodata)
self.obj_wrapper = ObjectiveWrapper(param_type)
self.last_save = time.time()
self.logpost_history = FiniteRunningSum()
self.like_history = FiniteRunningSum()
# Importance Samplers -------------------------------------------------
self.is_sym = get_symmetryop(self.cparams.get('is_symmetry',self.cparams.get('symmetry',None)))
self.sampler_R = FixedFisherImportanceSampler('_R',self.is_sym)
self.sampler_I = FixedFisherImportanceSampler('_I')
self.sampler_S = FixedGaussianImportanceSampler('_S')
self.like_func.set_samplers(sampler_R=self.sampler_R,sampler_I=self.sampler_I,sampler_S=self.sampler_S)
def sagd_dostep(data_dir, model_file, use_angular_correlation=True):
cryodata, (M, fM), cparams = sagd_init(data_dir, model_file,
use_angular_correlation)
iteration = cparams['iteration']
sagd_params = {
'iteration': iteration,
'exp_path': 'exp/',
'num_batches': cryodata.N_batches,
'sagd_linesearch': 'max_frequency_changed or ((iteration%5 == 0) if iteration < 2500 else \
(iteration%3 == 0) if iteration < 5000 else \
True)' ,
'sagd_linesearch_accuracy': 1.01 if iteration < 10 else \
1.10 if iteration < 2500 else \
1.25 if iteration < 5000 else \
1.50,
'sagd_linesearch_maxits': 5 if iteration < 2500 else 3,
'sagd_incL': 1.0,
'sagd_momentum': 1 - 1.0/(1.0 + 0.1 * iteration),
# 'sagd_learnrate': '1.0/min(16.0,2**(num_max_frequency_changes))',
'shuffle_minibatches': 'iteration >= 1000',
}
# initial logging
if os.path.exists('exp/sagd_L0.pkl'):
os.remove('exp/sagd_L0.pkl')
# for diagnostics and parameters, # for stats (per image etc), # for likelihoods of individual images
diagout = Output(os.path.join('exp', 'diag'), runningout=False)
statout = Output(os.path.join('exp', 'stat'), runningout=True)
likeout = Output(os.path.join('exp', 'like'), runningout=False)
ostream = OutputStream(None)
# Setup SAGD optimizer
step = SAGDStep()
step.setup(cparams, diagout, statout, ostream)
# Objective function setup
param_type = cparams.get('parameterization', 'real')
cplx_param = param_type in [
'complex', 'complex_coeff', 'complex_herm_coeff'
]
like_func = eval_objective(cparams['likelihood'])
prior_func = eval_objective(cparams['prior'])
obj = SumObjectives(cplx_param, [like_func, prior_func], [None, None])
obj.setup(cparams, diagout, statout, ostream)
obj.set_dataset(cryodata)
obj_wrapper = ObjectiveWrapper(param_type)
is_sym = get_symmetryop(
cparams.get('is_symmetry', cparams.get('symmetry', None)))
sampler_R = FixedFisherImportanceSampler('_R', is_sym)
sampler_I = FixedFisherImportanceSampler('_I')
# sampler_S = FixedGaussianImportanceSampler('_S')
sampler_S = None
like_func.set_samplers(sampler_R=sampler_R,
sampler_I=sampler_I,
sampler_S=sampler_S)
# Start iteration
num_data_evals = 0
num_iterations = 10
for i in range(num_iterations):
cparams['iteration'] = i
sagd_params['iteration'] = i
print('Iteration #:', i)
minibatch = cryodata.get_next_minibatch(True)
num_data_evals += minibatch['N_M']
# setup the wrapper for the objective function
obj.set_data(cparams, minibatch)
obj_wrapper.set_objective(obj)
x0 = obj_wrapper.set_density(M, fM)
evalobj = obj_wrapper.eval_obj
# Get step size
trainLogP, dlogP, v, res_train, extra_num_data = \
step.do_step(x0, sagd_params, cryodata, evalobj, batch=minibatch)
# print('trainLogP:', trainLogP)
# print('dlogP:', dlogP.shape)
# print('v:', v.shape)
# print('res_train:', res_train.keys()) # dict_keys(['CV2_R', 'CV2_I', 'Evar_like', 'Evar_prior', 'sigma2_est', 'correlation', 'power', 'like', 'N_R_sampled', 'N_I_sampled', 'N_Total_sampled', 'totallike_logscale', 'kern_timing', 'angular_correlation_timing', 'like_timing', 'N_R', 'N_I', 'N_Total', 'N_R_sampled_total', 'N_I_sampled_total', 'N_Total_sampled_total', 'L', 'all_logPs', 'all_dlogPs'])
# print('res_train L:', res_train['L'])
# print('res_train like:', res_train['like'])
# print('extra_num_data:', extra_num_data)
# Apply the step
x = x0 + v
# Convert from parameters to value
prevM = np.copy(M)
M, fM = obj_wrapper.convert_parameter(x, comp_real=True)
# Compute net change
dM = prevM - M
# Compute step statistics
step_size = np.linalg.norm(dM)
grad_size = np.linalg.norm(dlogP)
M_norm = np.linalg.norm(M)
num_data_evals += extra_num_data
inc_ratio = step_size / M_norm
# Update import sampling distributions
sampler_R.perform_update()
sampler_I.perform_update()
