def inference_pass(query, query_id, image_ix, if_data, gm_method='original'):
""" Run the full inference pass using a query and image
Attributes:
query (obj): the query to execute
query_id (int): a query identification number
image_ix (int): the image to run against the image
if_data (obj): an ImageFetchData object
Returns:
(object) opengm graphical model
(object) DetectionTracker
(float) energy value
(numpy array) array of best-match indices
(numpy array) post-inference marginals
(float) total pass time
"""
import time
gm = None
tracker = None
start = time.time()
if_data.configure(image_ix, query)
if gm_method == 'original':
gm, tracker = ifc.generate_pgm(if_data, verbose=False)
elif gm_method == 'uniform' or gm_method == 'empirical':
gm, tracker = ifc.generate_pgm_all_objects(if_data, verbose=False, method=gm_method)
energy, indices, marginals = ifc.do_inference(gm)
duration = time.time() - start
print "query {0} on image {1}: {2} sec".format(query_id, image_ix, duration)
return gm, tracker, energy, indices, marginals, duration
def ex4(srao_query_string, image_index, gm_method='original', verbose=True):
""" querygen
"""
import image_fetch_core as ifc
import image_fetch_utils as ifu; reload(ifu)
import image_fetch_plot as ifp; reload(ifp)
import image_fetch_querygen as ifq; reload(ifq)
vgd, potentials, platt_mod, bin_mod, queries, ifdata = dp.get_all_data()
query = ifq.gen_srao(srao_query_string)
ifdata.configure(image_index, query)
gm = None
tracker = None
if gm_method == 'original':
gm, tracker = ifc.generate_pgm(ifdata, verbose)
elif gm_method == 'uniform' or gm_method == 'empirical':
gm, tracker = ifc.generate_pgm_all_objects(ifdata, verbose=verbose, method=gm_method)
energy, best_match_ix, marginals = ifc.do_inference(gm)
#ifp.draw_gm(gm)
ifp.draw_best_objects(tracker, best_match_ix, energy, filename = out_path + "mq_i{0}.png".format(image_index))
file_prefix = "mq_i{}_".format(image_index)
ifp.draw_all_heatmaps(tracker, ifdata.object_detections, marginals, gm_method, out_path, file_prefix)
def alternate_inference_test(query_ix, image_ix, gm_method='original', verbose=True):
import image_fetch_core as ifc
import opengm as ogm
vgd, potentials, platt_mod, bin_mod, queries, ifdata = dp.get_all_data()
query = vgd['vg_data_test'][query_ix].annotations
ifdata.configure(image_ix, query)
gm = None
tracker = None
if gm_method == 'original':
gm, tracker = ifc.generate_pgm(ifdata, verbose)
elif gm_method == 'uniform' or gm_method == 'empirical':
gm, tracker = ifc.generate_pgm_all_objects(ifdata, verbose, method=gm_method)
#inf_param = ogm.InfParam(steps=120, damping=0., convergenceBound=0.001)
#infr = ogm.inference.BeliefPropagation(gm, parameter=inf_param)
#infr.infer(infr.verboseVisitor())
return gm, tracker#, infr
def ex1(query_index,
image_index,
inf_alg='bp',
gm_method='original',
do_suppl_plots=True,
save_gm=False,
verbose=True):
""" generate plots for a query/image pair
"""
import image_fetch_core as ifc
reload(ifc)
import image_fetch_plot as ifp
reload(ifp)
vgd, potentials, platt_mod, bin_mod, queries, ifdata = dp.get_all_data()
query = vgd['vg_data_test'][query_index].annotations
ifdata.configure(image_index, query)
gm = None
tracker = None
if gm_method == 'original':
gm, tracker = ifc.generate_pgm(ifdata, verbose)
elif gm_method == 'uniform' or gm_method == 'empirical':
gm, tracker = ifc.generate_pgm_all_objects(ifdata,
verbose=verbose,
method=gm_method)
file_prefix = "q{0}_i{1}_".format(query_index, image_index)
energy = None
best_match_ix = None
marginals = None
obj_file = None
if inf_alg == 'bp':
obj_file = out_path + file_prefix + gm_method + '_bp_objects.png'
energy, best_match_ix, marginals = ifc.do_inference(gm)
if do_suppl_plots:
ifp.draw_all_heatmaps(tracker, ifdata.object_detections, marginals,
gm_method, out_path, file_prefix)
#ifp.p_compare(tracker, ifdata.object_detections, marginals, out_path+file_prefix+'sctr.png')
elif inf_alg == 'astar':
obj_file = out_path + file_prefix + 'as_objects.png'
energy, best_match_ix = ifc.do_inference_astar(gm)
ifp.draw_best_objects(tracker, best_match_ix, energy, filename=obj_file)
if save_gm: ifp.draw_gm(gm)
def gen_viz_file(query_ixs, image_ixs, output_path):
import image_fetch_core as ifc; reload(ifc)
import image_fetch_utils as ifu; reload(ifu)
vgd, potentials, platt_mod, bin_mod, queries, ifdata = dp.get_all_data()
for query_ix in query_ixs:
filename = 'q_' + str(query_ix) + '.csv'
f = open(output_path+filename, 'w')
query = queries['simple_graphs'][query_ix]
query_str = ifu.sg_to_str(query.annotations)
tp_simple = ifu.get_partial_scene_matches(vgd['vg_data_test'], queries['simple_graphs'])
tp_indices = tp_simple[query_ix]
for i in image_ixs: #range(0, len(image_ixs)):
ifdata.configure(i, query.annotations)
gm, tracker = ifc.generate_pgm(ifdata, verbose=False)
energy, best_match_ix, marginals = ifc.do_inference(gm)
line = '{:03d}, 0, "{}"\n'.format(i, query_str)
f.write(line)
line = '{:03d}, 2, {:0.4f}\n'.format(i, energy)
f.write(line)
if i in tp_indices:
line = '{:03d}, 3, "match"\n'.format(i)
else:
line = '{:03d}, 3, "no match"\n'.format(i)
f.write(line)
for obj_ix in range(0, len(best_match_ix)):
obj_name = tracker.object_names[obj_ix]
box_ix = best_match_ix[obj_ix]
bc = tracker.box_coords[box_ix]
line = '{:03d}, 1, {}, "{}", {}, {}, {}, {}\n'.format(i, obj_ix, obj_name, int(bc[0]), int(bc[1]), int(bc[2]), int(bc[3]))
f.write(line)
f.close()
