def __init__(self,dataset,model_name,num_obs_vars,m='0',r1='1', lbp_iters=3000):
# TODO: experiment with different values of fastinf
self.dataset = dataset
self.fd = FastinfDiscretizer(self.dataset, model_name)
self.res_fname = config.get_fastinf_res_file(dataset,model_name,m,r1)
# TODO: experiment with different amounts of smoothing
# amount of smoothing is correlated with fastinf slowness, values [0,1)
self.smoothing = 0.5
self.cache_fname = config.get_fastinf_cache_file(dataset,model_name,m,r1,self.smoothing)
if opexists(self.cache_fname):
with open(self.cache_fname) as f:
print("Loading fastinf cache from file")
self.cache = cPickle.load(f)
else:
self.cache = {}
self.cmd = config.fastinf_bin+" -i %s -m 0 -Is %f -Imm %d"%(self.res_fname, self.smoothing, lbp_iters)
self.num_obs_vars = num_obs_vars
self.tt = ut.TicToc().tic()
self.process = pexpect.spawn(self.cmd)
self.blacklist = []
marginals = self.get_marginals()
print("FastinfModel: Computed initial marginals in %.3f sec"%self.tt.qtoc())
def run_fastinf_different_settings(d, ms, rs, suffixs, num_bins = 5):
settings = list(itertools.product(suffixs, ms, rs))
table_gt = d.get_cls_ground_truth().arr.astype(int)
print 'run with a total of %d settings'%len(settings)
for setindx in range(comm_rank, len(settings), comm_size):
second_table = None
setin = settings[setindx]
suffix = setin[0]
m = str(setin[1])
r1 = str(setin[2])
print 'node %d runs %s, m=%s, r1=%s'%(comm_rank, suffix, m, r1)
filename = config.get_fastinf_mrf_file(d, suffix)
data_filename = config.get_fastinf_data_file(d, suffix)
if suffix == 'perfect':
table = np.hstack((table_gt, table_gt))
bounds = np.tile(np.linspace(0, 1, num_bins+1),(table_gt.shape[1],1))
print bounds
elif suffix == 'GIST':
table = gist_classify_dataset(d)
bounds, discr_table = discretize_table(table, num_bins)
table = np.hstack((table_gt, discr_table))
elif suffix == 'CSC':
filename_csc = os.path.join(config.get_ext_dets_foldname(d),'table')
print filename_csc
if not os.path.exists(filename_csc):
raise RuntimeWarning('The csc classification could not be loaded from %s'%filename_csc)
orig_table = cPickle.load(open(filename_csc,'r'))
if isinstance(orig_table, Table):
orig_table = orig_table.arr[:,:-1]
bounds, discr_table = discretize_table(orig_table, num_bins)
table = np.hstack((table_gt, discr_table))
elif suffix == 'CSC_regions':
rm = RegionModel("1big_2small", 0.5)
detector = 'csc_default'
from synthetic.dataset_policy import DatasetPolicy
orig_table = DatasetPolicy.load_ext_detections(d, detector)
gt = d.get_det_gt().copy()
# we need to spice up the gt by a score of 1 for each class (results in less code)
gt.cols.append('score')
gt.arr = np.hstack((gt.arr, np.ones((gt.shape[0], 1))))
table_gt_region = create_regioned_table(rm, gt, d.images, len(d.classes))
# At this point we need to split them for the different regions
orig_table_region = create_regioned_table(rm, orig_table, d.images, len(d.classes))
bounds, discr_table_region = discretize_table(orig_table_region, num_bins)
table = np.hstack((table_gt_region, discr_table_region))
elif suffix == 'GIST_CSC':
filename_csc = os.path.join(config.get_ext_dets_foldname(d),'table')
if not os.path.exists(filename_csc):
raise RuntimeWarning('The csc classification could not be loaded from %s'%filename_csc)
orig_table = cPickle.load(open(filename_csc,'r'))
if isinstance(orig_table, Table):
orig_table = orig_table.arr[:,:-1]
bounds, discr_table = discretize_table(orig_table, num_bins)
table = np.hstack((table_gt, discr_table))
store_bound(d, 'CSC', bounds)
second_table = gist_classify_dataset(d)
sec_bounds, second_table = discretize_table(second_table, num_bins)
store_bound(d, 'GIST', sec_bounds)
full_bound = np.hstack((sec_bounds, bounds))
store_bound(d, 'GIST_CSC', full_bound)
if not suffix == 'GIST_CSC':
store_bound(d, suffix, bounds)
print 'set up table on %d, write out mrf for %s, m=%s, r1=%s'%(comm_rank, suffix, m, r1)
write_out_mrf(table, num_bins, filename, data_filename, second_table=second_table)
add_sets = ['-m',m]
if not r1 == '':
add_sets += ['-r1', r1]
if not second_table == None:
sec_bound_file = '%s_secbounds'%filename
for s in add_sets:
sec_bound_file += '_'+s
np.savetxt(sec_bound_file, sec_bounds)
print '%d start running lbp for %s, m=%s, r1=%s'%(comm_rank, suffix, m, r1)
filename_out = config.get_fastinf_res_file(d, suffix, m, r1)
execute_lbp(filename, data_filename, filename_out, add_settings=add_sets)
def run_fastinf_different_settings(d, ms, rs, suffixs, num_bins = 5):
settings = list(itertools.product(suffixs, ms, rs))
table_gt = d.get_cls_ground_truth().arr.astype(int)
print 'run with a total of %d settings'%len(settings)
for setindx in range(mpi.comm_rank, len(settings), mpi.comm_size):
second_table = None
setin = settings[setindx]
suffix = setin[0]
m = str(setin[1])
r1 = str(setin[2])
print 'node %d runs %s, m=%s, r1=%s'%(mpi.comm_rank, suffix, m, r1)
filename = config.get_fastinf_mrf_file(d, suffix)
data_filename = config.get_fastinf_data_file(d, suffix)
if suffix == 'perfect':
table = np.hstack((table_gt, table_gt))
bounds = np.tile(np.linspace(0, 1, num_bins+1),(table_gt.shape[1],1))
print bounds
elif suffix == 'GIST':
table = gist_classify_dataset(d)
bounds, discr_table = discretize_table(table, num_bins)
table = np.hstack((table_gt, discr_table))
elif suffix == 'CSC':
filename_csc = os.path.join(config.get_ext_dets_foldname(d),'table')
print filename_csc
if not os.path.exists(filename_csc):
raise RuntimeWarning('The csc classification could not be loaded from %s'%filename_csc)
orig_table = cPickle.load(open(filename_csc,'r'))
if isinstance(orig_table, Table):
orig_table = orig_table.arr[:,:-1]
bounds, discr_table = discretize_table(orig_table, num_bins)
table = np.hstack((table_gt, discr_table))
elif suffix == 'CSC_regions':
rm = RegionModel("1big_2small", 0.5)
detector = 'csc_default'
from synthetic.dataset_policy import DatasetPolicy
orig_table = DatasetPolicy.load_ext_detections(d, detector)
gt = d.get_det_gt().copy()
# we need to spice up the gt by a score of 1 for each class (results in less code)
gt.cols.append('score')
gt.arr = np.hstack((gt.arr, np.ones((gt.shape[0], 1))))
table_gt_region = create_regioned_table(rm, gt, d.images, len(d.classes))
# At this point we need to split them for the different regions
orig_table_region = create_regioned_table(rm, orig_table, d.images, len(d.classes))
bounds, discr_table_region = discretize_table(orig_table_region, num_bins)
table = np.hstack((table_gt_region, discr_table_region))
elif suffix == 'GIST_CSC':
filename_csc = os.path.join(config.get_ext_dets_foldname(d),'table')
if not os.path.exists(filename_csc):
raise RuntimeWarning('The csc classification could not be loaded from %s'%filename_csc)
orig_table = cPickle.load(open(filename_csc,'r'))
if isinstance(orig_table, Table):
orig_table = orig_table.arr[:,:-1]
bounds, discr_table = discretize_table(orig_table, num_bins)
table = np.hstack((table_gt, discr_table))
store_bound(d, 'CSC', bounds)
second_table = gist_classify_dataset(d)
sec_bounds, second_table = discretize_table(second_table, num_bins)
store_bound(d, 'GIST', sec_bounds)
full_bound = np.hstack((sec_bounds, bounds))
store_bound(d, 'GIST_CSC', full_bound)
if not suffix == 'GIST_CSC':
store_bound(d, suffix, bounds)
print 'set up table on %d, write out mrf for %s, m=%s, r1=%s'%(mpi.comm_rank, suffix, m, r1)
write_out_mrf(table, num_bins, filename, data_filename, second_table=second_table)
add_sets = ['-m',m]
if not r1 == '':
add_sets += ['-r1', r1]
if not second_table == None:
sec_bound_file = '%s_secbounds'%filename
for s in add_sets:
sec_bound_file += '_'+s
np.savetxt(sec_bound_file, sec_bounds)
print '%d start running lbp for %s, m=%s, r1=%s'%(mpi.comm_rank, suffix, m, r1)
filename_out = config.get_fastinf_res_file(d, suffix, m, r1)
execute_lbp(filename, data_filename, filename_out, add_settings=add_sets)
