def enlarge(img, xTimes, yTimes):
# todo 未调试
if xTimes > 1:
img = np_repeat(img, xTimes, axis=0)
if yTimes > 1:
img = np_repeat(img, yTimes, axis=1)
return img
def train_document_dm(model, doc_words, doctag_indexes, alpha, work=None, neu1=None,
learn_doctags=True, learn_words=True, learn_hidden=True,
word_vectors=None, word_locks=None, doctag_vectors=None, doctag_locks=None):
"""
Update distributed memory model ("PV-DM") by training on a single document.
Called internally from `Doc2Vec.train()` and `Doc2Vec.infer_vector()`. This
method implements the DM model with a projection (input) layer that is
either the sum or mean of the context vectors, depending on the model's
`dm_mean` configuration field. See `train_dm_concat()` for the DM model
with a concatenated input layer.
The document is provided as `doc_words`, a list of word tokens which are looked up
in the model's vocab dictionary, and `doctag_indexes`, which provide indexes
into the doctag_vectors array.
Any of `learn_doctags', `learn_words`, and `learn_hidden` may be set False to
prevent learning-updates to those respective model weights, as if using the
(partially-)frozen model to infer other compatible vectors.
This is the non-optimized, Python version. If you have a C compiler, gensim
will use the optimized version from doc2vec_inner instead.
"""
if word_vectors is None:
word_vectors = model.syn0
if word_locks is None:
word_locks = model.syn0_lockf
if doctag_vectors is None:
doctag_vectors = model.docvecs.doctag_syn0
if doctag_locks is None:
doctag_locks = model.docvecs.doctag_syn0_lockf
word_vocabs = [model.vocab[w] for w in doc_words if w in model.vocab and
model.vocab[w].sample_int > model.random.randint(2**32)]
doctag_sum = np_sum(doctag_vectors[doctag_indexes], axis=0)
doctag_len = len(doctag_indexes)
for pos, word in enumerate(word_vocabs):
reduced_window = model.random.randint(model.window) # `b` in the original doc2vec code
start = max(0, pos - model.window + reduced_window)
window_pos = enumerate(word_vocabs[start:(pos + model.window + 1 - reduced_window)], start)
word2_indexes = [word2.index for pos2, word2 in window_pos if (word2 is not None and pos2 != pos)]
l1 = np_sum(word_vectors[word2_indexes], axis=0) + doctag_sum # 1 x layer1_size
if word2_indexes and model.cbow_mean:
l1 /= (len(word2_indexes) + doctag_len)
neu1e = train_cbow_pair(model, word, word2_indexes, l1, alpha,
learn_vectors=False, learn_hidden=learn_hidden)
if word2_indexes and not model.cbow_mean:
neu1e /= (len(word2_indexes) + doctag_len)
if learn_doctags:
doctag_vectors[doctag_indexes] += neu1e * \
np_repeat(doctag_locks[doctag_indexes], model.vector_size).reshape(-1, model.vector_size)
if learn_words:
word_vectors[word2_indexes] += neu1e * \
np_repeat(word_locks[word2_indexes], model.vector_size).reshape(-1, model.vector_size)
return len(word_vocabs)
def fan_regular_pols(np_verts,
np_pols,
np_distances,
np_faces_id,
custom_normals,
index_offset=0,
use_custom_normals=False,
output_old_v_id=True,
output_old_face_id=True,
output_pols_groups=True):
pols_number = np_pols.shape[0]
pol_sides = np_pols.shape[1]
v_pols = np_verts[np_pols] #shape [num_pols, num_corners, 3]
if (len(np_distances) > 1
and np.any(np_distances != 0)) or np_distances != 0:
if use_custom_normals:
normals = custom_normals
else:
normals = np_faces_normals(v_pols)
average = np.sum(
v_pols, axis=1
) / pol_sides + normals * np_distances[:,
np_newaxis] #shape [num_pols, 3]
else:
average = np.sum(v_pols, axis=1) / pol_sides
idx_offset = len(np_verts) + index_offset
new_idx = np_arange(idx_offset, pols_number + idx_offset)
new_pols = np.zeros([pols_number, pol_sides, 3], dtype=int)
new_pols[:, :, 0] = np_pols
new_pols[:, :, 1] = np_roll(np_pols, -1, axis=1)
new_pols[:, :, 2] = new_idx[:, np_newaxis]
old_vert_id = np_pols[:, 0].tolist() if output_old_v_id else []
if output_old_face_id:
old_face_id = np_repeat(np_faces_id[:, np_newaxis], pol_sides,
axis=1).tolist()
else:
old_face_id = []
if output_pols_groups:
pols_groups = np_repeat(1, len(new_pols) * pol_sides).tolist()
else:
pols_groups = []
return (
average.tolist(),
new_pols.reshape(-1, 3).tolist(),
old_vert_id,
old_face_id,
pols_groups,
)
def numpy_full_list(array, desired_length):
'''retuns array with desired length by repeating last item'''
if not isinstance(array, ndarray):
array = np_array(array)
length_diff = desired_length - array.shape[0]
if length_diff > 0:
new_part = np_repeat(array[np_newaxis, -1], length_diff, axis=0)
return np_concatenate((array, new_part))[:desired_length]
return array[:desired_length]
def numpy_match_long_repeat(list_of_arrays):
'''match numpy arrays length by repeating last one'''
out = []
maxl = 0
for array in list_of_arrays:
maxl = max(maxl, array.shape[0])
for array in list_of_arrays:
length_diff = maxl - array.shape[0]
if length_diff > 0:
new_part = np_repeat(array[np_newaxis, -1], length_diff, axis=0)
array = np_concatenate((array, new_part))
out.append(array)
return out
def numpy_full_list_cycle(array, desired_length):
'''retuns array with desired length by cycling'''
length_diff = desired_length - array.shape[0]
if length_diff > 0:
if length_diff < array.shape[0]:
return np_concatenate((array, array[:length_diff]))
new_part = np_repeat(array, ceil(length_diff / array.shape[0]), axis=0)
if len(array.shape) > 1:
shape = (ceil(length_diff / array.shape[0]), 1)
else:
shape = ceil(length_diff / array.shape[0])
new_part = np_tile(array, shape)
return np_concatenate((array, new_part[:length_diff]))
return array[:desired_length]
def numpy_match_long_cycle(list_of_arrays):
'''match numpy arrays length by cycling over the array'''
out = []
maxl = 0
for array in list_of_arrays:
maxl = max(maxl, array.shape[0])
for array in list_of_arrays:
length_diff = maxl - array.shape[0]
if length_diff > 0:
if length_diff < array.shape[0]:
array = np_concatenate((array, array[:length_diff]))
else:
new_part = np_repeat(array, ceil(length_diff / array.shape[0]), axis=0)
if len(array.shape) > 1:
shape = (ceil(length_diff / array.shape[0]), 1)
else:
shape = ceil(length_diff / array.shape[0])
new_part = np_tile(array, shape)
array = np_concatenate((array, new_part[:length_diff]))
out.append(array)
return out
def train_document_dm_concat(model, doc_words, doctag_indexes, alpha, work=None, neu1=None,
learn_doctags=True, learn_words=True, learn_hidden=True,
word_vectors=None, word_locks=None, doctag_vectors=None, doctag_locks=None):
"""
Update distributed memory model ("PV-DM") by training on a single document, using a
concatenation of the context window word vectors (rather than a sum or average).
Called internally from `Doc2Vec.train()` and `Doc2Vec.infer_vector()`.
The document is provided as `doc_words`, a list of word tokens which are looked up
in the model's vocab dictionary, and `doctag_indexes`, which provide indexes
into the doctag_vectors array.
Any of `learn_doctags', `learn_words`, and `learn_hidden` may be set False to
prevent learning-updates to those respective model weights, as if using the
(partially-)frozen model to infer other compatible vectors.
This is the non-optimized, Python version. If you have a C compiler, gensim
will use the optimized version from doc2vec_inner instead.
"""
if word_vectors is None:
word_vectors = model.syn0
if word_locks is None:
word_locks = model.syn0_lockf
if doctag_vectors is None:
doctag_vectors = model.docvecs.doctag_syn0
if doctag_locks is None:
doctag_locks = model.docvecs.doctag_syn0_lockf
word_vocabs = [model.vocab[w] for w in doc_words if w in model.vocab and
model.vocab[w].sample_int > model.random.rand() * 2**32]
doctag_len = len(doctag_indexes)
if doctag_len != model.dm_tag_count:
return 0 # skip doc without expected number of doctag(s) (TODO: warn/pad?)
null_word = model.vocab['\0']
pre_pad_count = model.window
post_pad_count = model.window
padded_document_indexes = (
(pre_pad_count * [null_word.index]) # pre-padding
+ [word.index for word in word_vocabs if word is not None] # elide out-of-Vocabulary words
+ (post_pad_count * [null_word.index]) # post-padding
)
for pos in range(pre_pad_count, len(padded_document_indexes) - post_pad_count):
word_context_indexes = (
padded_document_indexes[(pos - pre_pad_count): pos] # preceding words
+ padded_document_indexes[(pos + 1):(pos + 1 + post_pad_count)] # following words
)
word_context_len = len(word_context_indexes)
predict_word = model.vocab[model.index2word[padded_document_indexes[pos]]]
# numpy advanced-indexing copies; concatenate, flatten to 1d
l1 = concatenate((doctag_vectors[doctag_indexes], word_vectors[word_context_indexes])).ravel()
neu1e = train_cbow_pair(model, predict_word, None, l1, alpha,
learn_hidden=learn_hidden, learn_vectors=False)
# filter by locks and shape for addition to source vectors
e_locks = concatenate((doctag_locks[doctag_indexes], word_locks[word_context_indexes]))
neu1e_r = (neu1e.reshape(-1, model.vector_size)
* np_repeat(e_locks, model.vector_size).reshape(-1, model.vector_size))
if learn_doctags:
np_add.at(doctag_vectors, doctag_indexes, neu1e_r[:doctag_len])
if learn_words:
np_add.at(word_vectors, word_context_indexes, neu1e_r[doctag_len:])
return len(padded_document_indexes) - pre_pad_count - post_pad_count
def distrib_nn_for_cdf(self, ntss_tmp, bool_print: bool = False):
"""
Calculates the two indicators, average and standard deviation of the distances, necessary for the use of the CDF of the normal distribution.
The computation of these indicators are described in `Scoring Message Stream Anomalies in Railway Communication Systems, L.Foulon et al., 2019, ICDMWorkshop <https://ieeexplore.ieee.org/abstract/document/8955558>`_.
:param numpy.ndarray ntss_tmp: Reference sequences
:param boolean bool_print: and True, Displays the nodes stats on the standard output
:returns:
:rtype: list(numpy.ndarray, numpy.array)
"""
start_time = time_time()
node_list, node_list_leaf, node_leaf_ndarray_mean = self.get_list_nodes_and_barycentre(
)
if bool_print:
print("pretrait node --- %s seconds ---" %
(time_time() - start_time))
stdout.flush()
print(len(node_list), " nodes whose ", len(node_list_leaf),
" leafs in tree")
stdout.flush()
nb_leaf = len(node_list_leaf)
cdf_mean = np_zeros((nb_leaf, len(ntss_tmp)))
cdf_std = np_zeros(nb_leaf)
nb_ts_by_node = np_zeros(nb_leaf, dtype=np_uint32)
centroid_dist = np_square(cdist(node_leaf_ndarray_mean, ntss_tmp))
for num, node in enumerate(node_list_leaf):
cdf_std[node.id_numpy_leaf] = np_mean(node.std)
nb_ts_by_node[node.id_numpy_leaf] = node.get_nb_sequences()
dist_list = np_array([np_zeros(i) for i in nb_ts_by_node],
dtype=object)
# calcul distance au carre entre [barycentre et ts] du meme nœud
""" TODO np.vectorize ?"""
for node_nn in node_list_leaf:
dist_list[node_nn.id_numpy_leaf] = cdist(
[node_nn.mean], node_nn.get_sequences())[0]
dist_list = np_square(dist_list)
""" TODO np.vectorize ?"""
for num, node in enumerate(node_list_leaf):
node_id = node.id_numpy_leaf
centroid_dist_tmp = centroid_dist[node_id]
centroid_dist_tmp = centroid_dist_tmp.reshape(
centroid_dist_tmp.shape + (1, ))
centroid_dist_tmp = np_repeat(centroid_dist_tmp,
nb_ts_by_node[node_id],
axis=1)
cdf_mean_tmp = np_add(centroid_dist_tmp, dist_list[node_id])
cdf_mean[node_id] = np_sum(cdf_mean_tmp, axis=1)
del dist_list
del cdf_mean_tmp
del centroid_dist_tmp
cdf_mean = np_divide(cdf_mean.T, nb_ts_by_node)
cdf_mean = np_sqrt(cdf_mean)
self.cdf_mean = cdf_mean
self.cdf_std = cdf_std
def inset_regular_pols(np_verts,
np_pols,
np_distances,
np_inset_rate,
np_make_inners,
np_faces_id,
custom_normals,
matrices,
offset_mode='CENTER',
proportional=False,
concave_support=True,
index_offset=0,
use_custom_normals=False,
output_old_face_id=True,
output_old_v_id=True,
output_pols_groups=True):
pols_number = np_pols.shape[0]
pol_sides = np_pols.shape[1]
v_pols = np_verts[np_pols] #shape [num_pols, num_corners, 3]
if offset_mode == 'SIDES':
inner_points = sides_mode_inset(v_pols, np_inset_rate, np_distances,
concave_support, proportional,
use_custom_normals, custom_normals)
elif offset_mode == 'MATRIX':
inner_points = matrix_mode_inset(v_pols, matrices, use_custom_normals,
custom_normals)
else:
if any(np_distances != 0):
if use_custom_normals:
normals = custom_normals
else:
normals = np_faces_normals(v_pols)
average = np.sum(
v_pols, axis=1
) / pol_sides #+ normals*np_distances[:, np_newaxis] #shape [num_pols, 3]
inner_points = average[:, np_newaxis, :] + (
v_pols - average[:, np_newaxis, :]
) * np_inset_rate[:, np_newaxis,
np_newaxis] + normals[:,
np_newaxis, :] * np_distances[:,
np_newaxis,
np_newaxis]
else:
average = np.sum(v_pols, axis=1) / pol_sides #shape [num_pols, 3]
inner_points = average[:, np_newaxis, :] + (
v_pols - average[:, np_newaxis, :]
) * np_inset_rate[:, np_newaxis, np_newaxis]
idx_offset = len(np_verts) + index_offset
new_v_idx = np_arange(idx_offset,
pols_number * pol_sides + idx_offset).reshape(
pols_number, pol_sides)
side_pols = np.zeros([pols_number, pol_sides, 4], dtype=int)
side_pols[:, :, 0] = np_pols
side_pols[:, :, 1] = np_roll(np_pols, -1, axis=1)
side_pols[:, :, 2] = np_roll(new_v_idx, -1, axis=1)
side_pols[:, :, 3] = new_v_idx
side_faces = side_pols.reshape(-1, 4)
new_insets = new_v_idx[np_make_inners]
if pol_sides == 4:
new_faces = np_concatenate([side_faces, new_insets]).tolist()
else:
new_faces = side_faces.tolist() + new_insets.tolist()
old_v_id = np_pols.flatten().tolist() if output_old_v_id else []
if output_old_face_id:
side_ids = np.repeat(np_faces_id[:, np_newaxis], pol_sides, axis=1)
inset_ids = np_faces_id[np_make_inners]
old_face_id = np.concatenate((side_ids.flatten(), inset_ids)).tolist()
else:
old_face_id = []
if output_pols_groups:
pols_groups = np_repeat(
[1, 2], [len(side_faces), len(new_insets)]).tolist()
else:
pols_groups = []
return (inner_points.reshape(-1, 3).tolist(), new_faces,
new_insets.tolist(), old_v_id, old_face_id, pols_groups)
def train_document_dm(model,
doc_words,
doctag_indexes,
alpha,
work=None,
neu1=None,
learn_doctags=True,
learn_words=True,
learn_hidden=True,
word_vectors=None,
word_locks=None,
doctag_vectors=None,
doctag_locks=None):
"""
Update distributed memory model ("PV-DM") by training on a single document.
Called internally from `Doc2Vec.train()` and `Doc2Vec.infer_vector()`. This
method implements the DM model with a projection (input) layer that is
either the sum or mean of the context vectors, depending on the model's
`dm_mean` configuration field. See `train_dm_concat()` for the DM model
with a concatenated input layer.
The document is provided as `doc_words`, a list of word tokens which are looked up
in the model's vocab dictionary, and `doctag_indexes`, which provide indexes
into the doctag_vectors array.
Any of `learn_doctags', `learn_words`, and `learn_hidden` may be set False to
prevent learning-updates to those respective model weights, as if using the
(partially-)frozen model to infer other compatible vectors.
This is the non-optimized, Python version. If you have a C compiler, gensim
will use the optimized version from doc2vec_inner instead.
"""
if word_vectors is None:
word_vectors = model.syn0
if word_locks is None:
word_locks = model.syn0_lockf
if doctag_vectors is None:
doctag_vectors = model.docvecs.doctag_syn0
if doctag_locks is None:
doctag_locks = model.docvecs.doctag_syn0_lockf
word_vocabs = [
model.vocab[w] for w in doc_words if w in model.vocab
and model.vocab[w].sample_int > model.random.randint(2**32)
]
doctag_sum = np_sum(doctag_vectors[doctag_indexes], axis=0)
doctag_len = len(doctag_indexes)
for pos, word in enumerate(word_vocabs):
reduced_window = model.random.randint(
model.window) # `b` in the original doc2vec code
start = max(0, pos - model.window + reduced_window)
window_pos = enumerate(
word_vocabs[start:(pos + model.window + 1 - reduced_window)],
start)
word2_indexes = [
word2.index for pos2, word2 in window_pos
if (word2 is not None and pos2 != pos)
]
l1 = np_sum(word_vectors[word2_indexes],
axis=0) + doctag_sum # 1 x layer1_size
if word2_indexes and model.cbow_mean:
l1 /= (len(word2_indexes) + doctag_len)
neu1e = train_cbow_pair(model,
word,
word2_indexes,
l1,
alpha,
learn_vectors=False,
learn_hidden=learn_hidden)
if word2_indexes and not model.cbow_mean:
neu1e /= (len(word2_indexes) + doctag_len)
if learn_doctags:
doctag_vectors[doctag_indexes] += neu1e * \
np_repeat(doctag_locks[doctag_indexes], model.vector_size).reshape(-1, model.vector_size)
if learn_words:
word_vectors[word2_indexes] += neu1e * \
np_repeat(word_locks[word2_indexes], model.vector_size).reshape(-1, model.vector_size)
return len(word_vocabs)
def mock_noise(scale, size):
"""Return `size` copies of the `scale` parameter."""
return np_repeat(scale, size)
