def __compute_vocabulary(cx2_desc, train_cxs, vocab_size, cache_dir=None):
'''Computes a vocabulary of size vocab_size given a set of training data'''
# Read params
akm_flann_params = params.BOW_AKMEANS_FLANN_PARAMS
words_flann_params = params.BOW_WORDS_FLANN_PARAMS
max_iters = params.AKMEANS_MAX_ITERS
# Make a training set of descriptors to build the vocabulary
tx2_desc = cx2_desc[train_cxs]
train_desc = np.vstack(tx2_desc)
num_train_desc = train_desc.shape[0]
if vocab_size > num_train_desc:
msg = '[mc2] vocab_size(%r) > #train_desc(%r)' % (vocab_size,
num_train_desc)
helpers.printWARN(msg)
vocab_size = num_train_desc / 2
# Cluster descriptors into a visual vocabulary
matcher_uid = params.get_matcher_uid(with_train=True, with_indx=False)
words_uid = 'words_' + matcher_uid
_, words = algos.precompute_akmeans(train_desc,
vocab_size,
max_iters,
akm_flann_params,
cache_dir,
force_recomp=False,
same_data=False,
uid=words_uid)
# Index the vocabulary for fast nearest neighbor search
words_flann = algos.precompute_flann(words,
cache_dir,
uid=words_uid,
flann_params=words_flann_params)
return words, words_flann
def __compute_vocabulary(cx2_desc, train_cxs, vocab_size, cache_dir=None):
'''Computes a vocabulary of size vocab_size given a set of training data'''
# Read params
akm_flann_params = params.BOW_AKMEANS_FLANN_PARAMS
words_flann_params = params.BOW_WORDS_FLANN_PARAMS
max_iters = params.AKMEANS_MAX_ITERS
# Make a training set of descriptors to build the vocabulary
tx2_desc = cx2_desc[train_cxs]
train_desc = np.vstack(tx2_desc)
num_train_desc = train_desc.shape[0]
if vocab_size > num_train_desc:
msg = '[mc2] vocab_size(%r) > #train_desc(%r)' % (vocab_size, num_train_desc)
helpers.printWARN(msg)
vocab_size = num_train_desc / 2
# Cluster descriptors into a visual vocabulary
matcher_uid = params.get_matcher_uid(with_train=True, with_indx=False)
words_uid = 'words_'+matcher_uid
_, words = algos.precompute_akmeans(train_desc, vocab_size, max_iters,
akm_flann_params, cache_dir,
force_recomp=False, same_data=False,
uid=words_uid)
# Index the vocabulary for fast nearest neighbor search
words_flann = algos.precompute_flann(words, cache_dir, uid=words_uid,
flann_params=words_flann_params)
return words, words_flann
def __index_database_to_vocabulary(cx2_desc, words, words_flann, indexed_cxs,
cache_dir):
'''Assigns each database chip a visual-vector and returns
data for the inverted file'''
# TODO: Save precomputations here
print('[mc2] Assigning each database chip a bag-of-words vector')
num_indexed = len(indexed_cxs)
ax2_cx, ax2_fx, ax2_desc = __aggregate_descriptors(cx2_desc, indexed_cxs)
# Build UID
matcher_uid = params.get_matcher_uid()
data_uid = helpers.hashstr(ax2_desc)
uid = data_uid + '_' + matcher_uid
try:
cx2_vvec = io.smart_load(cache_dir, 'cx2_vvec', uid, '.cPkl') #sparse
wx2_cxs = io.smart_load(cache_dir, 'wx2_cxs', uid, '.npy')
wx2_fxs = io.smart_load(cache_dir, 'wx2_fxs', uid, '.npy')
wx2_idf = io.smart_load(cache_dir, 'wx2_idf', uid, '.npy')
print('[mc2] successful cache load: vocabulary indexed databased.')
return cx2_vvec, wx2_cxs, wx2_fxs, wx2_idf
#helpers.CacheException as ex:
except IOError as ex:
print(repr(ex))
print('[mc2] quantizing each descriptor to a word')
# Assign each descriptor to its nearest visual word
print('[mc2] ...this may take awhile with no indication of progress')
tt1 = helpers.Timer('quantizing each descriptor to a word')
ax2_wx, _ = words_flann.nn_index(ax2_desc, 1, checks=128)
tt1.toc()
# Build inverse word to ax
tt2 = helpers.Timer('database_indexing')
print('')
print('[mc2] building inverse word to ax map')
wx2_axs = [[] for _ in xrange(len(words))]
for ax, wx in enumerate(ax2_wx):
wx2_axs[wx].append(ax)
# Compute inverted file: words -> database
print('[mc2] building inverted file word -> database')
wx2_cxs = np.array([[ax2_cx[ax] for ax in ax_list] for ax_list in wx2_axs])
wx2_fxs = np.array([[ax2_fx[ax] for ax in ax_list] for ax_list in wx2_axs])
# Build sparse visual vectors with term frequency weights
print('[mc2] building sparse visual words')
coo_cols = ax2_wx
coo_rows = ax2_cx
coo_values = np.ones(len(ax2_cx), dtype=BOW_DTYPE)
coo_format = (coo_values, (coo_rows, coo_cols))
coo_cx2_vvec = spsparse.coo_matrix(coo_format, dtype=np.float, copy=True)
cx2_tf_vvec = spsparse.csr_matrix(coo_cx2_vvec, copy=False)
# Compute idf_w = log(Number of documents / Number of docs containing word_j)
print('[mc2] computing tf-idf')
wx2_df = np.array([len(set(cxs)) + 1 for cxs in wx2_cxs], dtype=np.float)
wx2_idf = np.array(np.log2(np.float(num_indexed) / wx2_df))
# Compute tf-idf
print('[mc2] preweighting with tf-idf')
cx2_tfidf_vvec = algos.sparse_multiply_rows(cx2_tf_vvec, wx2_idf)
# Normalize
print('[mc2] normalizing')
cx2_tfidf_vvec = algos.sparse_multiply_rows(cx2_tf_vvec, wx2_idf)
cx2_vvec = algos.sparse_normalize_rows(cx2_tfidf_vvec)
tt2.toc()
# Save to cache
print('[mc2] saving to cache')
r'''
input_data = ax2_desc
data = cx2_vvec
uid='cx2_vvec'+matcher_uid
'''
io.smart_save(cx2_vvec, cache_dir, 'cx2_vvec', uid, '.cPkl') #sparse
io.smart_save(wx2_cxs, cache_dir, 'wx2_cxs', uid, '.npy')
io.smart_save(wx2_fxs, cache_dir, 'wx2_fxs', uid, '.npy')
io.smart_save(wx2_idf, cache_dir, 'wx2_idf', uid, '.npy')
return cx2_vvec, wx2_cxs, wx2_fxs, wx2_idf
def __index_database_to_vocabulary(cx2_desc, words, words_flann, indexed_cxs, cache_dir):
'''Assigns each database chip a visual-vector and returns
data for the inverted file'''
# TODO: Save precomputations here
print('[mc2] Assigning each database chip a bag-of-words vector')
num_indexed = len(indexed_cxs)
ax2_cx, ax2_fx, ax2_desc = __aggregate_descriptors(cx2_desc, indexed_cxs)
# Build UID
matcher_uid = params.get_matcher_uid()
data_uid = helpers.hashstr(ax2_desc)
uid = data_uid + '_' + matcher_uid
try:
cx2_vvec = io.smart_load(cache_dir, 'cx2_vvec', uid, '.cPkl') #sparse
wx2_cxs = io.smart_load(cache_dir, 'wx2_cxs', uid, '.npy')
wx2_fxs = io.smart_load(cache_dir, 'wx2_fxs', uid, '.npy')
wx2_idf = io.smart_load(cache_dir, 'wx2_idf', uid, '.npy')
print('[mc2] successful cache load: vocabulary indexed databased.')
return cx2_vvec, wx2_cxs, wx2_fxs, wx2_idf
#helpers.CacheException as ex:
except IOError as ex:
print(repr(ex))
print('[mc2] quantizing each descriptor to a word')
# Assign each descriptor to its nearest visual word
print('[mc2] ...this may take awhile with no indication of progress')
tt1 = helpers.Timer('quantizing each descriptor to a word')
ax2_wx, _ = words_flann.nn_index(ax2_desc, 1, checks=128)
tt1.toc()
# Build inverse word to ax
tt2 = helpers.Timer('database_indexing')
print('')
print('[mc2] building inverse word to ax map')
wx2_axs = [[] for _ in xrange(len(words))]
for ax, wx in enumerate(ax2_wx):
wx2_axs[wx].append(ax)
# Compute inverted file: words -> database
print('[mc2] building inverted file word -> database')
wx2_cxs = np.array([[ax2_cx[ax] for ax in ax_list] for ax_list in wx2_axs])
wx2_fxs = np.array([[ax2_fx[ax] for ax in ax_list] for ax_list in wx2_axs])
# Build sparse visual vectors with term frequency weights
print('[mc2] building sparse visual words')
coo_cols = ax2_wx
coo_rows = ax2_cx
coo_values = np.ones(len(ax2_cx), dtype=BOW_DTYPE)
coo_format = (coo_values, (coo_rows, coo_cols))
coo_cx2_vvec = spsparse.coo_matrix(coo_format, dtype=np.float, copy=True)
cx2_tf_vvec = spsparse.csr_matrix(coo_cx2_vvec, copy=False)
# Compute idf_w = log(Number of documents / Number of docs containing word_j)
print('[mc2] computing tf-idf')
wx2_df = np.array([len(set(cxs))+1 for cxs in wx2_cxs], dtype=np.float)
wx2_idf = np.array(np.log2(np.float(num_indexed) / wx2_df))
# Compute tf-idf
print('[mc2] preweighting with tf-idf')
cx2_tfidf_vvec = algos.sparse_multiply_rows(cx2_tf_vvec, wx2_idf)
# Normalize
print('[mc2] normalizing')
cx2_tfidf_vvec = algos.sparse_multiply_rows(cx2_tf_vvec, wx2_idf)
cx2_vvec = algos.sparse_normalize_rows(cx2_tfidf_vvec)
tt2.toc()
# Save to cache
print('[mc2] saving to cache')
r'''
input_data = ax2_desc
data = cx2_vvec
uid='cx2_vvec'+matcher_uid
'''
io.smart_save(cx2_vvec, cache_dir, 'cx2_vvec', uid, '.cPkl') #sparse
io.smart_save(wx2_cxs, cache_dir, 'wx2_cxs', uid, '.npy')
io.smart_save(wx2_fxs, cache_dir, 'wx2_fxs', uid, '.npy')
io.smart_save(wx2_idf, cache_dir, 'wx2_idf', uid, '.npy')
return cx2_vvec, wx2_cxs, wx2_fxs, wx2_idf
