def show_probability_chip(
ibs, aid, species=None, fnum=None, config2_=None, blend=False, **kwargs
):
"""
TODO: allow species override in controller
CommandLine:
python -m wbia.viz.viz_hough --exec-show_probability_chip --cnn --show
python -m wbia.viz.viz_hough --exec-show_probability_chip --cnn --show --db PZ_Master1
python -m wbia.viz.viz_hough --exec-show_probability_chip --cnn --show --db PZ_Master1 --aid 9970
Example:
>>> # SCRIPT
>>> from wbia.viz.viz_hough import * # NOQA
>>> import wbia
>>> from wbia.viz import viz_chip
>>> ibs, aid_list, kwargs, config2_ = viz_chip.testdata_showchip()
>>> fnum = 1
>>> species = None
>>> aid = aid_list[0]
>>> fig, ax = show_probability_chip(ibs, aid, species, fnum, blend=True, **kwargs)
>>> ut.show_if_requested()
"""
fnum = pt.ensure_fnum(fnum)
title = 'Probability Chip: ' + ', '.join(vh.get_annot_text(ibs, [aid], True))
hough_cpath = ibs.get_annot_probchip_fpath(aid, config2_=config2_)
img = vt.imread(hough_cpath)
if blend:
chip = ibs.get_annot_chips(aid, config2_=config2_)
img = vt.blend_images_multiply(chip, vt.resize_mask(img, chip))
fig, ax = viz_image2.show_image(img, title=title, fnum=fnum, **kwargs)
return fig, ax
def draw_demo():
r"""
CommandLine:
python -m plottool.interact_impaint --exec-draw_demo --show
Example:
>>> # SCRIPT
>>> from plottool.interact_impaint import * # NOQA
>>> result = draw_demo()
>>> print(result)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> ut.show_if_requested()
"""
fpath = ut.grab_test_imgpath('zebra.png')
img = vt.imread(fpath)
mask = impaint_mask2(img)
print('mask = %r' % (mask, ))
print('mask.sum() = %r' % (mask.sum(), ))
if False:
plt.imshow(vt.blend_images_multiply(img, mask))
ax = plt.gca()
ax.grid(False)
ax.set_xticks([])
ax.set_yticks([])
def draw_demo():
r"""
CommandLine:
python -m plottool.interact_impaint --exec-draw_demo --show
Example:
>>> # SCRIPT
>>> from plottool.interact_impaint import * # NOQA
>>> result = draw_demo()
>>> print(result)
>>> ut.quit_if_noshow()
>>> import plottool as pt
>>> ut.show_if_requested()
"""
fpath = ut.grab_test_imgpath('zebra.png')
img = vt.imread(fpath)
mask = impaint_mask2(img)
print('mask = %r' % (mask,))
print('mask.sum() = %r' % (mask.sum(),))
if False:
plt.imshow(vt.blend_images_multiply(img, mask))
ax = plt.gca()
ax.grid(False)
ax.set_xticks([])
ax.set_yticks([])
def show_probability_chip(ibs, aid, species=None, fnum=None, config2_=None,
blend=False, **kwargs):
"""
TODO: allow species override in controller
CommandLine:
python -m ibeis.viz.viz_hough --exec-show_probability_chip --cnn --show
python -m ibeis.viz.viz_hough --exec-show_probability_chip --cnn --show --db PZ_Master1
python -m ibeis.viz.viz_hough --exec-show_probability_chip --cnn --show --db PZ_Master1 --aid 9970
Example:
>>> # SCRIPT
>>> from ibeis.viz.viz_hough import * # NOQA
>>> import ibeis
>>> from ibeis.viz import viz_chip
>>> ibs, aid_list, kwargs, config2_ = viz_chip.testdata_showchip()
>>> fnum = 1
>>> species = None
>>> aid = aid_list[0]
>>> fig, ax = show_probability_chip(ibs, aid, species, fnum, blend=True, **kwargs)
>>> ut.show_if_requested()
"""
fnum = pt.ensure_fnum(fnum)
title = 'Probability Chip: ' + ', '.join(vh.get_annot_text(ibs, [aid], True))
hough_cpath = ibs.get_annot_probchip_fpath(aid, config2_=config2_)
img = vt.imread(hough_cpath)
if blend:
chip = ibs.get_annot_chips(aid, config2_=config2_)
img = vt.blend_images_multiply(chip, vt.resize_mask(img, chip))
fig, ax = viz_image2.show_image(img, title=title, fnum=fnum, **kwargs)
return fig, ax
def saliency(dream, Xb, yb):
"""
num = 10
Xb = model.prepare_data(X_test[0:num])
yb = y_test[0:num]
dpath = ''
dataset = None
"""
dpath = '.'
import theano.tensor as T
import lasagne
import vtool as vt
import theano
model = dream.model
# Use current weights to find the score of a particular class
Xb_shared = theano.shared(Xb)
yb_shared = theano.shared(yb.astype(np.int32))
# Get the final layer and remove the softmax nonlinearity to access the
# pre-activation. (Softmax encourages minimization of other classes)
import copy
#softmax = copy.copy(model.output_layer)
#softmax.nonlinearity = lasagne.nonlinearities.identity
softmax = copy.copy(model.output_layer)
class_probs = lasagne.layers.get_output(softmax, Xb_shared,
deterministic=True)
# werid way to index into position of target
flat_idx = (T.arange(yb_shared.shape[0]) * class_probs.shape[1]) + yb_shared
class_probs_target = T.flatten(class_probs)[flat_idx]
# Get derivative of scores for the target class wrt the input
d_score_wrt_input = theano.grad(class_probs_target.mean(), Xb_shared)
w = np.array(d_score_wrt_input.eval())
saliency = w.max(axis=1, keepdims=True)
outs = saliency.transpose((0, 2, 3, 1))
X = Xb.transpose((0, 2, 3, 1))
for count in range(len(Xb)):
img = X[count]
y = yb[count]
out = vt.norm01(outs[count])
overlay = vt.blend_images_multiply(out, img)
vt.imwrite(join(dpath, 'out%d_A_image_t=%s.jpg' % (count, y)),
vt.rectify_to_uint8(img))
vt.imwrite(join(dpath, 'out%d_B_heat_t=%s.jpg' % (count, y)),
vt.rectify_to_uint8(out))
vt.imwrite(join(dpath, 'out%d_C_overlay_t=%s.jpg' % (count, y)),
vt.rectify_to_uint8(overlay))
def augment_shadow(Xb, yb=None, rng=np.random, return_shadowmaps=False):
"""
CommandLine:
python -m ibeis_cnn.augment --test-augment_shadow --show --db PZ_MTEST
Example:
>>> # ENABLE_DOCTEST
>>> from ibeis_cnn.augment import * # NOQA
>>> from ibeis_cnn import ingest_data, utils, draw_results
>>> Xb_orig, yb_orig, Xb, yb = testdata_augment()
>>> rng = np.random.RandomState(0)
>>> Xb, yb, shadows = augment_shadow(Xb, yb, rng=rng, return_shadowmaps=True)
>>> ut.quit_if_noshow()
>>> show_augmented_patches(Xb_orig, Xb, yb_orig, yb, shadows)
>>> ut.show_if_requested()
"""
import vtool as vt
# Rotate corresponding patches together
Xb1, Xb2 = Xb[::2], Xb[1::2]
perlin_perteb = .5
def perlin_noise01(size):
#scale = 128.0
#scale = 80.0
scale = size[0] * 1.5
noise = (vt.perlin_noise(size, scale=scale, rng=rng).astype(np.float32)) / 255.0
noise = np.transpose(noise[None, :], (1, 2, 0))
return noise
num = len(Xb1)
perlinperterb_flags = rng.uniform(0.0, 1.0, size=num) < perlin_perteb
#perlin_min = 0.2
#perlin_max = 0.9
#perlin_range = perlin_max - perlin_min
if return_shadowmaps:
shadows = np.empty(Xb.shape, dtype=Xb.dtype)
shadows1, shadows2 = shadows[::2], shadows[1::2]
for index in np.where(perlinperterb_flags)[0]:
img1 = Xb1[index]
img2 = Xb2[index]
# TODO: TAKE IN NORMALIZED POINTS
noise1 = perlin_noise01(img1.shape[0:2])
noise2 = perlin_noise01(img2.shape[0:2])
#noise1 = np.clip(noise1 / .7, 0, 1)
#noise2 = np.clip(noise2 / .7, 0, 1)
noise1 = np.clip((noise1 ** .7 - .15) / .75, .1, 1)
noise2 = np.clip((noise2 ** .7 - .15) / .75, .1, 1)
if return_shadowmaps:
shadows1[index] = noise1
shadows2[index] = noise2
#noise1[noise1 > .6] = 1
#noise2[noise2 > .6] = 1
#alpha1 = ((rng.rand() * perlin_range) + perlin_min)
#alpha2 = ((rng.rand() * perlin_range) + perlin_min)
#Xb1[index] = img1 ** (noise1 * alpha1)
#Xb2[index] = img2 ** (noise2 * alpha2)
#alpha1 = alpha2 = .5
alpha1 = alpha2 = .5
Xb1[index] = vt.blend_images_multiply(img1, noise1, alpha1)
Xb2[index] = vt.blend_images_multiply(img2, noise2, alpha2)
# Linear Burn
#Xb1[index] = np.clip(img1 + noise1 - 1.0, 0, 1)
#Xb2[index] = np.clip(img2 + noise2 - 1.0, 0, 1)
#Xb1[index] = vt.blend_images_average(img1, noise1, alpha1)
#Xb2[index] = vt.blend_images_average(img2, noise2, alpha2)
#Xb1[index] = noise1
#Xb2[index] = noise2
#with ut.embed_on_exception_context:
assert Xb.max() <= 1.0, 'max/min = %r, %r' % (Xb.min(), Xb.max())
assert Xb.min() >= 0.0, 'max/min = %r, %r' % (Xb.min(), Xb.max())
#bad_xs = np.where((Xb > 1).sum(axis=-1).sum(axis=-1).sum(axis=-1) > 0)[0]
if return_shadowmaps:
return Xb, yb, shadows
return Xb, yb
