def make_conv_weight_image(all_weights, limit=144):
""" just makes the image ndarray of the weights """
import vtool as vt
# Try to infer if use_color should be shown
num, channels, height, width = all_weights.shape
# Try to infer if use_color should be shown
use_color = (channels == 3)
# non-use_color features need to be flattened
if not use_color:
all_weights_ = all_weights.reshape(num * channels, height, width, 1)
else:
# convert from theano to cv2 BGR
all_weights_ = utils.convert_theano_images_to_cv2_images(all_weights)
# convert from BGR to RGB
all_weights_ = all_weights_[..., ::-1]
#cv2.cvtColor(all_weights_[-1], cv2.COLOR_BGR2RGB)
# Limit all_weights_
#num = all_weights_.shape[0]
num, height, width, channels = all_weights_.shape
if limit is not None and num > limit:
all_weights_ = all_weights_[:limit]
num = all_weights_.shape[0]
# Convert weight values to image values
normalize_individually = False
if normalize_individually:
# Normalize each feature individually
all_max = vt.multiaxis_reduce(np.amax, all_weights_, startaxis=1)
all_min = vt.multiaxis_reduce(np.amin, all_weights_, startaxis=1)
all_domain = all_max - all_min
extra_dims = (None,) * (len(all_weights_.shape) - 1)
broadcaster = (slice(None),) + extra_dims
all_features = ((all_weights_ - all_min[broadcaster]) *
(255.0 / all_domain[broadcaster])).astype(np.uint8)
else:
# Normalize jointly across all filters
_max = all_weights_.max()
_min = all_weights_.min()
_domain = _max - _min
all_features = ((all_weights_ - _min) * (255.0 / _domain)).astype(np.uint8)
#import scipy.misc
# resize feature, give them a border, and stack them together
new_height, new_width = max(32, height), max(32, width)
nbp_ = 1 # num border pixels
_resized_features = np.array([
cv2.resize(img, (new_width, new_height),
interpolation=cv2.INTER_NEAREST)
for img in all_features
])
resized_features = _resized_features.reshape(
num, new_height, new_width, channels)
border_shape = (num, new_height + (nbp_ * 2),
new_width + (nbp_ * 2), channels)
bordered_features = np.zeros(border_shape, dtype=resized_features.dtype)
bordered_features[:, nbp_:-nbp_, nbp_:-nbp_, :] = resized_features
#img_list = bordered_features
stacked_img = vt.stack_square_images(bordered_features)
return stacked_img
def render_vocab(vocab):
"""
Renders the average patch of each word.
This is a quick visualization of the entire vocabulary.
CommandLine:
python -m wbia.algo.smk.vocab_indexer render_vocab --show
Example:
>>> # DISABLE_DOCTEST
>>> from wbia.algo.smk.vocab_indexer import * # NOQA
>>> vocab = testdata_vocab('PZ_MTEST', num_words=64)
>>> all_words = vocab.render_vocab()
>>> ut.quit_if_noshow()
>>> import wbia.plottool as pt
>>> pt.qt4ensure()
>>> pt.imshow(all_words)
>>> ut.show_if_requested()
"""
import wbia.plottool as pt
wx_list = list(range(len(vocab)))
# wx_list = ut.strided_sample(wx_list, 64)
wx_list = ut.strided_sample(wx_list, 64)
word_patch_list = []
for wx in ut.ProgIter(wx_list, bs=True, lbl='building patches'):
word = vocab.wx_to_word[wx]
word_patch = vt.inverted_sift_patch(word, 64)
word_patch = pt.render_sift_on_patch(word_patch, word)
word_patch_list.append(word_patch)
all_words = vt.stack_square_images(word_patch_list)
return all_words
def visualize_vocab_word(ibs, invassign, wx, fnum=None):
"""
Example:
>>> from ibeis.new_annots import * # NOQA
>>> import plottool as pt
>>> pt.qt4ensure()
>>> ibs, aid_list, vocab = testdata_vocab()
>>> #aid_list = aid_list[0:1]
>>> fstack = StackedFeatures(ibs, aid_list)
>>> nAssign = 2
>>> invassign = fstack.inverted_assignment(vocab, nAssign)
>>> sortx = ut.argsort(invassign.num_list)[::-1]
>>> wx_list = ut.take(invassign.wx_list, sortx)
>>> wx = wx_list[0]
"""
import plottool as pt
pt.qt4ensure()
vecs = invassign.get_vecs(wx)
word = invassign.vocab.wx2_word[wx]
word_patches = invassign.get_patches(wx)
average_patch = np.mean(word_patches, axis=0)
average_vec = vecs.mean(axis=0)
average_vec = word
word
with_sift = True
fnum = 2
fnum = pt.ensure_fnum(fnum)
if with_sift:
patch_img = pt.render_sift_on_patch(average_patch, average_vec)
#sift_word_patches = [pt.render_sift_on_patch(patch, vec) for patch, vec in ut.ProgIter(list(zip(word_patches, vecs)))]
#stacked_patches = vt.stack_square_images(word_patches)
#stacked_patches = vt.stack_square_images(sift_word_patches)
else:
patch_img = average_patch
stacked_patches = vt.stack_square_images(word_patches)
solidbar = np.zeros((patch_img.shape[0], int(patch_img.shape[1] * .1), 3),
dtype=patch_img.dtype)
border_color = (100, 10, 10) # bgr, darkblue
if ut.is_float(solidbar):
solidbar[:, :, :] = (np.array(border_color) / 255)[None, None]
else:
solidbar[:, :, :] = np.array(border_color)[None, None]
word_img = vt.stack_image_list([patch_img, solidbar, stacked_patches],
vert=False,
modifysize=True)
pt.imshow(word_img, fnum=fnum)
#pt.imshow(patch_img, pnum=(1, 2, 1), fnum=fnum)
#patch_size = 64
#half_size = patch_size / 2
#pt.imshow(stacked_patches, pnum=(1, 2, 2), fnum=fnum)
pt.iup()
def visualize_vocab_word(ibs, invassign, wx, fnum=None):
"""
Example:
>>> from ibeis.new_annots import * # NOQA
>>> import plottool as pt
>>> pt.qt4ensure()
>>> ibs, aid_list, vocab = testdata_vocab()
>>> #aid_list = aid_list[0:1]
>>> fstack = StackedFeatures(ibs, aid_list)
>>> nAssign = 2
>>> invassign = fstack.inverted_assignment(vocab, nAssign)
>>> sortx = ut.argsort(invassign.num_list)[::-1]
>>> wx_list = ut.take(invassign.wx_list, sortx)
>>> wx = wx_list[0]
"""
import plottool as pt
pt.qt4ensure()
vecs = invassign.get_vecs(wx)
word = invassign.vocab.wx2_word[wx]
word_patches = invassign.get_patches(wx)
average_patch = np.mean(word_patches, axis=0)
average_vec = vecs.mean(axis=0)
average_vec = word
word
with_sift = True
fnum = 2
fnum = pt.ensure_fnum(fnum)
if with_sift:
patch_img = pt.render_sift_on_patch(average_patch, average_vec)
#sift_word_patches = [pt.render_sift_on_patch(patch, vec) for patch, vec in ut.ProgIter(list(zip(word_patches, vecs)))]
#stacked_patches = vt.stack_square_images(word_patches)
#stacked_patches = vt.stack_square_images(sift_word_patches)
else:
patch_img = average_patch
stacked_patches = vt.stack_square_images(word_patches)
solidbar = np.zeros((patch_img.shape[0], int(patch_img.shape[1] * .1), 3), dtype=patch_img.dtype)
border_color = (100, 10, 10) # bgr, darkblue
if ut.is_float(solidbar):
solidbar[:, :, :] = (np.array(border_color) / 255)[None, None]
else:
solidbar[:, :, :] = np.array(border_color)[None, None]
word_img = vt.stack_image_list([patch_img, solidbar, stacked_patches], vert=False, modifysize=True)
pt.imshow(word_img, fnum=fnum)
#pt.imshow(patch_img, pnum=(1, 2, 1), fnum=fnum)
#patch_size = 64
#half_size = patch_size / 2
#pt.imshow(stacked_patches, pnum=(1, 2, 2), fnum=fnum)
pt.iup()
def render_inverted_vocab(inva, ibs, use_data=False):
"""
Renders the average patch of each word.
This is a visualization of the entire vocabulary.
CommandLine:
python -m wbia.algo.smk.inverted_index render_inverted_vocab --show
python -m wbia.algo.smk.inverted_index render_inverted_vocab --show --use-data
python -m wbia.algo.smk.inverted_index render_inverted_vocab --show --debug-depc
Example:
>>> # DISABLE_DOCTEST
>>> from wbia.algo.smk.inverted_index import * # NOQA
>>> qreq_, inva = testdata_inva()
>>> ibs = qreq_.ibs
>>> all_words = inva.render_inverted_vocab(ibs)
>>> ut.quit_if_noshow()
>>> import wbia.plottool as pt
>>> pt.qt4ensure()
>>> pt.imshow(all_words)
>>> ut.show_if_requested()
"""
import wbia.plottool as pt
# Get words with the most assignments
vocab = ibs.depc['vocab'].get_row_data([inva.vocab_rowid], 'words')[0]
wx_list = ut.strided_sample(inva.wx_list, 64)
word_patch_list = []
for wx in ut.ProgIter(wx_list, bs=True, lbl='building patches'):
word = vocab.wx_to_word[wx]
word_patch = inva.get_word_patch(wx, ibs)
word_patch = pt.render_sift_on_patch(word_patch, word)
word_patch_list.append(word_patch)
all_words = vt.stack_square_images(word_patch_list)
return all_words
def make_wordfigures(ibs, metrics, invindex, figdir, wx_sample, wx2_dpath):
"""
Builds mosaics of patches assigned to words in sample
ouptuts them to disk
"""
from plottool import draw_func2 as df2
import vtool as vt
import parse
vocabdir = join(figdir, 'vocab_patches2')
ut.ensuredir(vocabdir)
dump_word_patches(ibs, vocabdir, invindex, wx_sample, metrics)
# COLLECTING PART --- collects patches in word folders
#vocabdir
seldpath = vocabdir + '_selected'
ut.ensurepath(seldpath)
# stack for show
for wx, dpath in ut.progiter(six.iteritems(wx2_dpath),
lbl='Dumping Word Images:',
num=len(wx2_dpath),
freq=1,
backspace=False):
#df2.rrr()
fpath_list = ut.ls(dpath)
fname_list = [basename(fpath_) for fpath_ in fpath_list]
patch_list = [vt.imread(fpath_) for fpath_ in fpath_list]
# color each patch by nid
nid_list = [
int(parse.parse('{}_nid={nid}_{}', fname)['nid'])
for fname in fname_list
]
nid_set = set(nid_list)
nid_list = np.array(nid_list)
if len(nid_list) == len(nid_set):
# no duplicate names
newpatch_list = patch_list
else:
# duplicate names. do coloring
sortx = nid_list.argsort()
patch_list = np.array(patch_list, dtype=object)[sortx]
fname_list = np.array(fname_list, dtype=object)[sortx]
nid_list = nid_list[sortx]
colors = (255 *
np.array(df2.distinct_colors(len(nid_set)))).astype(
np.int32)
color_dict = dict(zip(nid_set, colors))
wpad, hpad = 3, 3
newshape_list = [
tuple(
(np.array(patch.shape) + (wpad * 2, hpad * 2, 0)).tolist())
for patch in patch_list
]
color_list = [color_dict[nid_] for nid_ in nid_list]
newpatch_list = [
np.zeros(shape) + color[None, None]
for shape, color in zip(newshape_list, color_list)
]
for patch, newpatch in zip(patch_list, newpatch_list):
newpatch[wpad:-wpad, hpad:-hpad, :] = patch
#img_list = patch_list
#bigpatch = vt.stack_image_recurse(patch_list)
#bigpatch = vt.stack_image_list(patch_list, vert=False)
bigpatch = vt.stack_square_images(newpatch_list)
bigpatch_fpath = join(seldpath, basename(dpath) + '_patches.png')
#
def _dictstr(dict_):
str_ = ut.dict_str(dict_, newlines=False)
str_ = str_.replace('\'', '').replace(': ', '=').strip('{},')
return str_
figtitle = '\n'.join([
'wx=%r' % wx,
'stat(pdist): %s' % _dictstr(metrics.wx2_pdist_stats[wx]),
'stat(wdist): %s' % _dictstr(metrics.wx2_wdist_stats[wx]),
])
metrics.wx2_nMembers[wx]
df2.figure(fnum=1, doclf=True, docla=True)
fig, ax = df2.imshow(bigpatch, figtitle=figtitle)
#fig.show()
df2.set_figtitle(figtitle)
df2.adjust_subplots(top=.878, bottom=0)
df2.save_figure(1, bigpatch_fpath)
def make_wordfigures(ibs, metrics, invindex, figdir, wx_sample, wx2_dpath):
"""
Builds mosaics of patches assigned to words in sample
ouptuts them to disk
"""
from plottool import draw_func2 as df2
import vtool as vt
import parse
vocabdir = join(figdir, 'vocab_patches2')
ut.ensuredir(vocabdir)
dump_word_patches(ibs, vocabdir, invindex, wx_sample, metrics)
# COLLECTING PART --- collects patches in word folders
#vocabdir
seldpath = vocabdir + '_selected'
ut.ensurepath(seldpath)
# stack for show
for wx, dpath in ut.progiter(six.iteritems(wx2_dpath), lbl='Dumping Word Images:', num=len(wx2_dpath), freq=1, backspace=False):
#df2.rrr()
fpath_list = ut.ls(dpath)
fname_list = [basename(fpath_) for fpath_ in fpath_list]
patch_list = [gtool.imread(fpath_) for fpath_ in fpath_list]
# color each patch by nid
nid_list = [int(parse.parse('{}_nid={nid}_{}', fname)['nid']) for fname in fname_list]
nid_set = set(nid_list)
nid_list = np.array(nid_list)
if len(nid_list) == len(nid_set):
# no duplicate names
newpatch_list = patch_list
else:
# duplicate names. do coloring
sortx = nid_list.argsort()
patch_list = np.array(patch_list, dtype=object)[sortx]
fname_list = np.array(fname_list, dtype=object)[sortx]
nid_list = nid_list[sortx]
colors = (255 * np.array(df2.distinct_colors(len(nid_set)))).astype(np.int32)
color_dict = dict(zip(nid_set, colors))
wpad, hpad = 3, 3
newshape_list = [tuple((np.array(patch.shape) + (wpad * 2, hpad * 2, 0)).tolist()) for patch in patch_list]
color_list = [color_dict[nid_] for nid_ in nid_list]
newpatch_list = [np.zeros(shape) + color[None, None] for shape, color in zip(newshape_list, color_list)]
for patch, newpatch in zip(patch_list, newpatch_list):
newpatch[wpad:-wpad, hpad:-hpad, :] = patch
#img_list = patch_list
#bigpatch = vt.stack_image_recurse(patch_list)
#bigpatch = vt.stack_image_list(patch_list, vert=False)
bigpatch = vt.stack_square_images(newpatch_list)
bigpatch_fpath = join(seldpath, basename(dpath) + '_patches.png')
#
def _dictstr(dict_):
str_ = ut.dict_str(dict_, newlines=False)
str_ = str_.replace('\'', '').replace(': ', '=').strip('{},')
return str_
figtitle = '\n'.join([
'wx=%r' % wx,
'stat(pdist): %s' % _dictstr(metrics.wx2_pdist_stats[wx]),
'stat(wdist): %s' % _dictstr(metrics.wx2_wdist_stats[wx]),
])
metrics.wx2_nMembers[wx]
df2.figure(fnum=1, doclf=True, docla=True)
fig, ax = df2.imshow(bigpatch, figtitle=figtitle)
#fig.show()
df2.set_figtitle(figtitle)
df2.adjust_subplots(top=.878, bottom=0)
df2.save_figure(1, bigpatch_fpath)
def render_inverted_vocab_word(inva, wx, ibs, fnum=None):
"""
Creates a visualization of a visual word. This includes the average patch,
the SIFT-like representation of the centroid, and some of the patches that
were assigned to it.
CommandLine:
python -m wbia.algo.smk.inverted_index render_inverted_vocab_word --show
Example:
>>> # DISABLE_DOCTEST
>>> from wbia.algo.smk.inverted_index import * # NOQA
>>> import wbia.plottool as pt
>>> qreq_, inva = testdata_inva()
>>> ibs = qreq_.ibs
>>> wx_list = list(inva.wx_to_aids.keys())
>>> wx = wx_list[0]
>>> ut.qtensure()
>>> fnum = 2
>>> fnum = pt.ensure_fnum(fnum)
>>> # Interactive visualization of many words
>>> for wx in ut.InteractiveIter(wx_list):
>>> word_img = inva.render_inverted_vocab_word(wx, ibs, fnum)
>>> pt.imshow(word_img, fnum=fnum, title='Word %r/%r' % (wx, '?'))
>>> pt.update()
"""
import wbia.plottool as pt
# Create the contributing patch image
word_patches = inva.get_patches(wx, ibs)
word_patches_ = ut.strided_sample(word_patches, 64)
stacked_patches = vt.stack_square_images(word_patches_)
# Create the average word image
vocab = ibs.depc['vocab'].get_row_data([inva.vocab_rowid], 'words')[0]
word = vocab.wx_to_word[wx]
average_patch = np.mean(word_patches, axis=0)
# vecs = inva.get_vecs(wx)
# assert np.allclose(word, vecs.mean(axis=0))
with_sift = True
if with_sift:
patch_img = pt.render_sift_on_patch(average_patch, word)
else:
patch_img = average_patch
# Stack them together
solidbar = np.zeros(
(patch_img.shape[0], int(patch_img.shape[1] * 0.1), 3), dtype=patch_img.dtype,
)
border_color = (100, 10, 10) # bgr, darkblue
if ut.is_float(solidbar):
solidbar[:, :, :] = (np.array(border_color) / 255)[None, None]
else:
solidbar[:, :, :] = np.array(border_color)[None, None]
patch_img2 = vt.inverted_sift_patch(word)
# Fix types
patch_img = vt.rectify_to_uint8(patch_img)
patch_img2 = vt.rectify_to_uint8(patch_img2)
solidbar = vt.rectify_to_uint8(solidbar)
stacked_patches = vt.rectify_to_uint8(stacked_patches)
# Stack everything together
patch_img2, patch_img = vt.make_channels_comparable(patch_img2, patch_img)
img_list = [patch_img, solidbar, patch_img2, solidbar, stacked_patches]
word_img = vt.stack_image_list(img_list, vert=False, modifysize=True)
return word_img
