def write_chip_table(internal_dir,
cx2_cid,
cx2_gid,
cx2_nid,
cx2_roi,
cx2_theta,
prop_dict=None):
helpers.__PRINT_WRITES__ = True
print('Writing Chip Table')
# Make chip_table.csv
header = '# chip table'
column_labels = [
'ChipID', 'ImgID', 'NameID', 'roi[tl_x tl_y w h]', 'theta'
]
column_list = [cx2_cid, cx2_gid, cx2_nid, cx2_roi, cx2_theta]
column_type = [int, int, int, list, float]
if not prop_dict is None:
for key, val in prop_dict.iteritems():
column_labels.append(key)
column_list.append(val)
column_type.append(str)
chip_table = ld2.make_csv_table(column_labels, column_list, header,
column_type)
chip_table_fpath = join(internal_dir, ld2.CHIP_TABLE_FNAME)
write_to_wrapper(chip_table_fpath, chip_table)
def write_name_table(internal_dir, nx2_nid, nx2_name):
helpers.__PRINT_WRITES__ = True
# Make name_table.csv
column_labels = ['nid', 'name']
column_list = [nx2_nid[2:], nx2_name[2:]] # dont write ____ for backcomp
header = '# name table'
name_table = ld2.make_csv_table(column_labels, column_list, header)
name_table_fpath = join(internal_dir, ld2.NAME_TABLE_FNAME)
write_to_wrapper(name_table_fpath, name_table)
def write_name_table(internal_dir, nx2_nid, nx2_name):
helpers.__PRINT_WRITES__ = True
# Make name_table.csv
column_labels = ['nid', 'name']
column_list = [nx2_nid[2:], nx2_name[2:]] # dont write ____ for backcomp
header = '# name table'
name_table = ld2.make_csv_table(column_labels, column_list, header)
name_table_fpath = join(internal_dir, ld2.NAME_TABLE_FNAME)
write_to_wrapper(name_table_fpath, name_table)
def write_image_table(internal_dir, gx2_gid, gx2_gname):
helpers.__PRINT_WRITES__ = True
# Make image_table.csv
column_labels = ['gid', 'gname', 'aif'] # do aif for backwards compatibility
gx2_aif = np.ones(len(gx2_gid), dtype=np.uint32)
column_list = [gx2_gid, gx2_gname, gx2_aif]
header = '# image table'
image_table = ld2.make_csv_table(column_labels, column_list, header)
image_table_fpath = join(internal_dir, ld2.IMAGE_TABLE_FNAME)
write_to_wrapper(image_table_fpath, image_table)
def write_image_table(internal_dir, gx2_gid, gx2_gname):
helpers.__PRINT_WRITES__ = True
# Make image_table.csv
column_labels = ['gid', 'gname',
'aif'] # do aif for backwards compatibility
gx2_aif = np.ones(len(gx2_gid), dtype=np.uint32)
column_list = [gx2_gid, gx2_gname, gx2_aif]
header = '# image table'
image_table = ld2.make_csv_table(column_labels, column_list, header)
image_table_fpath = join(internal_dir, ld2.IMAGE_TABLE_FNAME)
write_to_wrapper(image_table_fpath, image_table)
def write_chip_table(internal_dir, cx2_cid, cx2_gid, cx2_nid,
cx2_roi, cx2_theta, prop_dict=None):
helpers.__PRINT_WRITES__ = True
print('Writing Chip Table')
# Make chip_table.csv
header = '# chip table'
column_labels = ['ChipID', 'ImgID', 'NameID', 'roi[tl_x tl_y w h]', 'theta']
column_list = [cx2_cid, cx2_gid, cx2_nid, cx2_roi, cx2_theta]
column_type = [int, int, int, list, float]
if not prop_dict is None:
for key, val in prop_dict.iteritems():
column_labels.append(key)
column_list.append(val)
column_type.append(str)
chip_table = ld2.make_csv_table(column_labels, column_list, header, column_type)
chip_table_fpath = join(internal_dir, ld2.CHIP_TABLE_FNAME)
write_to_wrapper(chip_table_fpath, chip_table)
def oxsty_mAP_results(allres):
print('oxsty_results> Building oxsty results')
hs = allres.hs
qcx2_res = allres.qcx2_res
SV = allres.SV
# Check directorys where ranked lists of images names will be put
oxsty_qres_dname = 'oxsty_ranked_lists' +allres.title_suffix
oxsty_qres_dpath = join(hs.dirs.qres_dir, oxsty_qres_dname)
helpers.ensure_path(oxsty_qres_dpath)
oxford_gt_dir = join(hs.dirs.db_dir, 'oxford_style_gt')
helpers.assertpath(oxford_gt_dir)
compute_ap_exe = normpath(join(oxford_gt_dir, '../compute_ap'))
if not helpers.checkpath(compute_ap_exe):
compute_ap_exe = normpath(join(oxford_gt_dir, '/compute_ap'))
helpers.assertpath(compute_ap_exe)
# Get the mAP scores using philbins program
query_mAP_list = []
query_mAP_cx = []
for qcx in iter(hs.test_sample_cx):
res = qcx2_res[qcx]
mAP = get_oxsty_mAP_score_from_res(hs, res, SV, oxsty_qres_dpath,
compute_ap_exe, oxford_gt_dir)
query_mAP_list.append(mAP)
query_mAP_cx.append(qcx)
print('')
# Calculate the scalar mAP score for the experiemnt
scalar_mAP = np.mean(np.array(query_mAP_list))
scalar_mAP_str = '# mAP score = %r\n' % scalar_mAP
# build a CSV file with the results
header = '# Oxford Style Map Scores: title_suffix=%r\n' % allres.title_suffix
header += scalar_mAP_str
header += helpers.get_timestamp(format='comment')+'\n'
header += '# Full Parameters: \n#' + params.param_string().replace('\n','\n#')+'\n\n'
column_labels = ['QCX', 'mAP']
column_list = [query_mAP_cx, query_mAP_list]
oxsty_map_csv = load_data2.make_csv_table(column_labels, column_list, header)
return oxsty_map_csv, scalar_mAP_str
def build_rankres_str(allres):
'Builds csv files showing the cxs/scores/ranks of the query results'
hs = allres.hs
#SV = allres.SV
#qcx2_res = allres.qcx2_res
cx2_cid = hs.tables.cx2_cid
#cx2_nx = hs.tables.cx2_nx
test_samp = hs.test_sample_cx
train_samp = hs.train_sample_cx
indx_samp = hs.indexed_sample_cx
# Get organized data for csv file
(qcx2_top_true_rank,
qcx2_top_true_score,
qcx2_top_true_cx) = allres.top_true_qcx_arrays
(qcx2_bot_true_rank,
qcx2_bot_true_score,
qcx2_bot_true_cx) = allres.bot_true_qcx_arrays
(qcx2_top_false_rank,
qcx2_top_false_score,
qcx2_top_false_cx) = allres.top_false_qcx_arrays
# Number of groundtruth per query
qcx2_numgt = np.zeros(len(cx2_cid)) - 2
for qcx in test_samp:
qcx2_numgt[qcx] = len(hs.get_other_indexed_cxs(qcx))
# Easy to digest results
num_chips = len(test_samp)
num_nonquery = len(np.setdiff1d(indx_samp, test_samp))
# Find the test samples WITH ground truth
test_samp_with_gt = np.array(test_samp)[qcx2_numgt[test_samp] > 0]
if len(test_samp_with_gt) == 0:
warnings.warn('[rr2] there were no queries with ground truth')
#train_nxs_set = set(cx2_nx[train_samp])
flag_cxs_fn = hs.flag_cxs_with_name_in_sample
def ranks_less_than_(thresh, intrain=None):
#Find the number of ranks scoring more than thresh
# Get statistics with respect to the training set
if len(test_samp_with_gt) == 0:
test_cxs_ = np.array([])
elif intrain is None: # report all
test_cxs_ = test_samp_with_gt
else: # report either or
in_train_flag = flag_cxs_fn(test_samp_with_gt, train_samp)
if intrain is False:
in_train_flag = True - in_train_flag
test_cxs_ = test_samp_with_gt[in_train_flag]
# number of test samples with ground truth
num_with_gt = len(test_cxs_)
if num_with_gt == 0:
return [], ('NoGT', 'NoGT', -1, 'NoGT')
# find tests with ranks greater and less than thresh
testcx2_ttr = qcx2_top_true_rank[test_cxs_]
greater_cxs = test_cxs_[np.where(testcx2_ttr > thresh)[0]]
num_greater = len(greater_cxs)
num_less = num_with_gt - num_greater
num_greater = num_with_gt - num_less
frac_less = 100.0 * num_less / num_with_gt
fmt_tup = (num_less, num_with_gt, frac_less, num_greater)
return greater_cxs, fmt_tup
greater5_cxs, fmt5_tup = ranks_less_than_(5)
greater1_cxs, fmt1_tup = ranks_less_than_(1)
#
gt5_intrain_cxs, fmt5_in_tup = ranks_less_than_(5, intrain=True)
gt1_intrain_cxs, fmt1_in_tup = ranks_less_than_(1, intrain=True)
#
gt5_outtrain_cxs, fmt5_out_tup = ranks_less_than_(5, intrain=False)
gt1_outtrain_cxs, fmt1_out_tup = ranks_less_than_(1, intrain=False)
#
allres.greater1_cxs = greater1_cxs
allres.greater5_cxs = greater5_cxs
#print('greater5_cxs = %r ' % (allres.greater5_cxs,))
#print('greater1_cxs = %r ' % (allres.greater1_cxs,))
# CSV Metadata
header = '# Experiment allres.title_suffix = ' + allres.title_suffix + '\n'
header += helpers.get_timestamp(format_='comment') + '\n'
# Scalar summary
scalar_summary = '# Num Query Chips: %d \n' % num_chips
scalar_summary += '# Num Query Chips with at least one match: %d \n' % len(test_samp_with_gt)
scalar_summary += '# Num NonQuery Chips: %d \n' % num_nonquery
scalar_summary += '# Ranks <= 5: %r/%r = %.1f%% (missed %r)\n' % (fmt5_tup)
scalar_summary += '# Ranks <= 1: %r/%r = %.1f%% (missed %r)\n\n' % (fmt1_tup)
scalar_summary += '# InTrain Ranks <= 5: %r/%r = %.1f%% (missed %r)\n' % (fmt5_in_tup)
scalar_summary += '# InTrain Ranks <= 1: %r/%r = %.1f%% (missed %r)\n\n' % (fmt1_in_tup)
scalar_summary += '# OutTrain Ranks <= 5: %r/%r = %.1f%% (missed %r)\n' % (fmt5_out_tup)
scalar_summary += '# OutTrain Ranks <= 1: %r/%r = %.1f%% (missed %r)\n\n' % (fmt1_out_tup)
header += scalar_summary
# Experiment parameters
#header += '# Full Parameters: \n' + helpers.indent(params.param_string(), '#') + '\n\n'
# More Metadata
header += textwrap.dedent('''
# Rank Result Metadata:
# QCX = Query chip-index
# QGNAME = Query images name
# NUMGT = Num ground truth matches
# TT = top true
# BT = bottom true
# TF = top false''').strip()
# Build the CSV table
test_sample_gx = hs.tables.cx2_gx[test_samp]
test_sample_gname = hs.tables.gx2_gname[test_sample_gx]
test_sample_gname = [g.replace('.jpg', '') for g in test_sample_gname]
column_labels = ['QCX', 'NUM GT',
'TT CX', 'BT CX', 'TF CX',
'TT SCORE', 'BT SCORE', 'TF SCORE',
'TT RANK', 'BT RANK', 'TF RANK',
'QGNAME', ]
column_list = [
test_samp, qcx2_numgt[test_samp],
qcx2_top_true_cx[test_samp], qcx2_bot_true_cx[test_samp],
qcx2_top_false_cx[test_samp], qcx2_top_true_score[test_samp],
qcx2_bot_true_score[test_samp], qcx2_top_false_score[test_samp],
qcx2_top_true_rank[test_samp], qcx2_bot_true_rank[test_samp],
qcx2_top_false_rank[test_samp], test_sample_gname, ]
column_type = [int, int, int, int, int,
float, float, float, int, int, int, str, ]
rankres_str = ld2.make_csv_table(column_labels, column_list, header, column_type)
# Put some more data at the end
problem_true_pairs = zip(allres.problem_true.qcxs, allres.problem_true.cxs)
problem_false_pairs = zip(allres.problem_false.qcxs, allres.problem_false.cxs)
problem_str = '\n'.join( [
'#Problem Cases: ',
'# problem_true_pairs = ' + repr(problem_true_pairs),
'# problem_false_pairs = ' + repr(problem_false_pairs)])
rankres_str += '\n' + problem_str
# Attach results to allres structure
allres.rankres_str = rankres_str
allres.scalar_summary = scalar_summary
allres.problem_false_pairs = problem_false_pairs
allres.problem_true_pairs = problem_true_pairs
allres.problem_false_pairs = problem_false_pairs
allres.problem_true_pairs = problem_true_pairs
def build_rankres_str(allres):
'Builds csv files showing the cxs/scores/ranks of the query results'
hs = allres.hs
#SV = allres.SV
#qcx2_res = allres.qcx2_res
cx2_cid = hs.tables.cx2_cid
#cx2_nx = hs.tables.cx2_nx
test_samp = hs.test_sample_cx
train_samp = hs.train_sample_cx
indx_samp = hs.indexed_sample_cx
# Get organized data for csv file
(qcx2_top_true_rank, qcx2_top_true_score,
qcx2_top_true_cx) = allres.top_true_qcx_arrays
(qcx2_bot_true_rank, qcx2_bot_true_score,
qcx2_bot_true_cx) = allres.bot_true_qcx_arrays
(qcx2_top_false_rank, qcx2_top_false_score,
qcx2_top_false_cx) = allres.top_false_qcx_arrays
# Number of groundtruth per query
qcx2_numgt = np.zeros(len(cx2_cid)) - 2
for qcx in test_samp:
qcx2_numgt[qcx] = len(hs.get_other_indexed_cxs(qcx))
# Easy to digest results
num_chips = len(test_samp)
num_nonquery = len(np.setdiff1d(indx_samp, test_samp))
# Find the test samples WITH ground truth
test_samp_with_gt = np.array(test_samp)[qcx2_numgt[test_samp] > 0]
if len(test_samp_with_gt) == 0:
warnings.warn('[rr2] there were no queries with ground truth')
#train_nxs_set = set(cx2_nx[train_samp])
flag_cxs_fn = hs.flag_cxs_with_name_in_sample
def ranks_less_than_(thresh, intrain=None):
#Find the number of ranks scoring more than thresh
# Get statistics with respect to the training set
if len(test_samp_with_gt) == 0:
test_cxs_ = np.array([])
elif intrain is None: # report all
test_cxs_ = test_samp_with_gt
else: # report either or
in_train_flag = flag_cxs_fn(test_samp_with_gt, train_samp)
if intrain is False:
in_train_flag = True - in_train_flag
test_cxs_ = test_samp_with_gt[in_train_flag]
# number of test samples with ground truth
num_with_gt = len(test_cxs_)
if num_with_gt == 0:
return [], ('NoGT', 'NoGT', -1, 'NoGT')
# find tests with ranks greater and less than thresh
testcx2_ttr = qcx2_top_true_rank[test_cxs_]
greater_cxs = test_cxs_[np.where(testcx2_ttr > thresh)[0]]
num_greater = len(greater_cxs)
num_less = num_with_gt - num_greater
num_greater = num_with_gt - num_less
frac_less = 100.0 * num_less / num_with_gt
fmt_tup = (num_less, num_with_gt, frac_less, num_greater)
return greater_cxs, fmt_tup
greater5_cxs, fmt5_tup = ranks_less_than_(5)
greater1_cxs, fmt1_tup = ranks_less_than_(1)
#
gt5_intrain_cxs, fmt5_in_tup = ranks_less_than_(5, intrain=True)
gt1_intrain_cxs, fmt1_in_tup = ranks_less_than_(1, intrain=True)
#
gt5_outtrain_cxs, fmt5_out_tup = ranks_less_than_(5, intrain=False)
gt1_outtrain_cxs, fmt1_out_tup = ranks_less_than_(1, intrain=False)
#
allres.greater1_cxs = greater1_cxs
allres.greater5_cxs = greater5_cxs
#print('greater5_cxs = %r ' % (allres.greater5_cxs,))
#print('greater1_cxs = %r ' % (allres.greater1_cxs,))
# CSV Metadata
header = '# Experiment allres.title_suffix = ' + allres.title_suffix + '\n'
header += helpers.get_timestamp(format_='comment') + '\n'
# Scalar summary
scalar_summary = '# Num Query Chips: %d \n' % num_chips
scalar_summary += '# Num Query Chips with at least one match: %d \n' % len(
test_samp_with_gt)
scalar_summary += '# Num NonQuery Chips: %d \n' % num_nonquery
scalar_summary += '# Ranks <= 5: %r/%r = %.1f%% (missed %r)\n' % (fmt5_tup)
scalar_summary += '# Ranks <= 1: %r/%r = %.1f%% (missed %r)\n\n' % (
fmt1_tup)
scalar_summary += '# InTrain Ranks <= 5: %r/%r = %.1f%% (missed %r)\n' % (
fmt5_in_tup)
scalar_summary += '# InTrain Ranks <= 1: %r/%r = %.1f%% (missed %r)\n\n' % (
fmt1_in_tup)
scalar_summary += '# OutTrain Ranks <= 5: %r/%r = %.1f%% (missed %r)\n' % (
fmt5_out_tup)
scalar_summary += '# OutTrain Ranks <= 1: %r/%r = %.1f%% (missed %r)\n\n' % (
fmt1_out_tup)
header += scalar_summary
# Experiment parameters
#header += '# Full Parameters: \n' + helpers.indent(params.param_string(), '#') + '\n\n'
# More Metadata
header += textwrap.dedent('''
# Rank Result Metadata:
# QCX = Query chip-index
# QGNAME = Query images name
# NUMGT = Num ground truth matches
# TT = top true
# BT = bottom true
# TF = top false''').strip()
# Build the CSV table
test_sample_gx = hs.tables.cx2_gx[test_samp]
test_sample_gname = hs.tables.gx2_gname[test_sample_gx]
test_sample_gname = [g.replace('.jpg', '') for g in test_sample_gname]
column_labels = [
'QCX',
'NUM GT',
'TT CX',
'BT CX',
'TF CX',
'TT SCORE',
'BT SCORE',
'TF SCORE',
'TT RANK',
'BT RANK',
'TF RANK',
'QGNAME',
]
column_list = [
test_samp,
qcx2_numgt[test_samp],
qcx2_top_true_cx[test_samp],
qcx2_bot_true_cx[test_samp],
qcx2_top_false_cx[test_samp],
qcx2_top_true_score[test_samp],
qcx2_bot_true_score[test_samp],
qcx2_top_false_score[test_samp],
qcx2_top_true_rank[test_samp],
qcx2_bot_true_rank[test_samp],
qcx2_top_false_rank[test_samp],
test_sample_gname,
]
column_type = [
int,
int,
int,
int,
int,
float,
float,
float,
int,
int,
int,
str,
]
rankres_str = ld2.make_csv_table(column_labels, column_list, header,
column_type)
# Put some more data at the end
problem_true_pairs = zip(allres.problem_true.qcxs, allres.problem_true.cxs)
problem_false_pairs = zip(allres.problem_false.qcxs,
allres.problem_false.cxs)
problem_str = '\n'.join([
'#Problem Cases: ',
'# problem_true_pairs = ' + repr(problem_true_pairs),
'# problem_false_pairs = ' + repr(problem_false_pairs)
])
rankres_str += '\n' + problem_str
# Attach results to allres structure
allres.rankres_str = rankres_str
allres.scalar_summary = scalar_summary
allres.problem_false_pairs = problem_false_pairs
allres.problem_true_pairs = problem_true_pairs
allres.problem_false_pairs = problem_false_pairs
allres.problem_true_pairs = problem_true_pairs
