def opt_crf():
from clab.torch.urban_pred import get_snapshot, urban_mapper_eval_dataset, PredictHarness # NOQA
from clab.torch.sseg_train import task_datasets, get_task, SSegInputsWrapper # NOQA
from clab import util
import ubelt as ub
# train_dpath = ub.truepath(
# '~/remote/aretha/data/work/urban_mapper/arch/unet/train/input_4214-yxalqwdk/solver_4214-yxalqwdk_unet_vgg_nttxoagf_a=1,n_ch=5,n_cl=3')
# load_path = get_snapshot(train_dpath, epoch=202)
datasets = task_datasets(get_task('urban_mapper_3d'))
test_dataset = datasets['test']
test_dataset.with_gt = False
test_dataset.inputs.make_dumpsafe_names()
test_dataset.center_inputs = test_dataset._original_urban_mapper_normalizer()
test_dataset.tag = 'test'
prob_folder = ub.truepath(
'~/remote/aretha/data/work/urban_mapper/test/input_4224-rwyxarza/solver_4214-yxalqwdk_unet_vgg_nttxoagf_a=1,n_ch=5,n_cl=3/_epoch_00000202/log_probs')
import glob
subset = slice(300, 310)
prob_paths = test_dataset.inputs.align(glob.glob(prob_folder + '/*.npz'))[subset]
gt_paths = test_dataset.inputs.gt_paths[subset]
im_paths = test_dataset.inputs.im_paths[subset]
import numpy as np
imgs = [util.imread(p) for p in ub.ProgIter(im_paths)]
probs = [np.load(p)['arr_0'] for p in ub.ProgIter(prob_paths)]
gts = [util.imread(p) for p in ub.ProgIter(gt_paths)]
from .torch import filters
# true = gts[4]
import optml
class_median_weights = test_dataset.class_weights()
class_weights = class_median_weights / class_median_weights.sum()
class CRFModel(optml.models.Model):
__model_module__ = 'sklearn' # hack
def __init__(self, **kwargs):
self.kwargs = kwargs
def get_params(self, deep=False):
return self.kwargs
def fit(self, X, y=None):
pass
def predict(self, X):
return [filters.crf_posterior(imgs[i], probs[i], **self.kwargs).argmax(axis=0) for i in ub.ProgIter(X, label='predicting')]
def clf_score(y_true, y_pred):
from .metrics import confusion_matrix, jaccard_score_from_confusion # NOQA
cfsn = np.zeros((3, 3))
for i, pred in zip(y_true, y_pred):
true = gts[i]
cfsn += confusion_matrix(true.ravel(), pred.ravel(), [0, 1, 2])
ious = jaccard_score_from_confusion(cfsn)
weighted_miou = ((ious * class_weights)[0:2]).sum()
return weighted_miou
from optml.bayesian_optimizer import BayesianOptimizer
model = CRFModel()
params = [
# w1 = kwargs.get('w1', 4)
# w2 = kwargs.get('w2', 3)
# sigma_alpha = kwargs.get('sigma_alpha', 100)
# sigma_beta = kwargs.get('sigma_beta', 3)
# sigma_gamma = kwargs.get('sigma_gamma', 3)
# n_iters = kwargs.get('n_iters', 10)
optml.Parameter(name='w1', param_type='integer', lower=1, upper=100),
optml.Parameter(name='sigma_alpha', param_type='integer', lower=1, upper=150),
optml.Parameter(name='sigma_beta', param_type='integer', lower=1, upper=150),
optml.Parameter(name='w2', param_type='integer', lower=3, upper=3),
optml.Parameter(name='sigma_gamma', param_type='integer', lower=3, upper=3),
optml.Parameter(name='n_iters', param_type='integer', lower=10, upper=10),
# optml.Parameter(name='param3', param_type='categorical', possible_values=['val1','val2','val3'])
]
optimizer = BayesianOptimizer(model=model,
hyperparams=params,
eval_func=clf_score)
# optimizer.model = model
X_train = np.arange(len(prob_paths)) # dummy
y_train = np.arange(len(X_train))
# Add some good values to help out initial model
seed_params = [
# Best known so far
{'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 4, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 2, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 5, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 3, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 105, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 95, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 101, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 99, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 61, 'sigma_beta': 11, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 139, 'sigma_beta': 1, 'w1': 50, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 139, 'sigma_beta': 3, 'w1': 50, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 10, 'sigma_alpha': 139, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
]
seed_params = [
# Best known so far
{'n_iters': 5, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
{'n_iters': 20, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 4, 'w2': 3, 'sigma_gamma': 3},
# {'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 4, 'w2': 1, 'sigma_gamma': 1},
# {'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 4, 'w2': 2, 'sigma_gamma': 2},
# {'n_iters': 10, 'sigma_alpha': 100, 'sigma_beta': 3, 'w1': 4, 'w2': 4, 'sigma_gamma': 4},
]
for seed in seed_params:
print('seed = {}'.format(ub.repr2(seed, nl=0, precision=2)))
print(optimizer._try_params(seed, X_train, y_train, X_train, y_train))
bayes_best_params, bayes_best_model = optimizer.fit(X_train=X_train,
y_train=y_train,
n_iters=10,
verbose=True)
names = [p.name for p in optimizer.hyperparams]
names = ['w1', 'sigma_alpha', 'sigma_beta']
xs = np.array([list(ub.take(params, names)) for score, params in optimizer.hyperparam_history])
ys = np.array([score for score, params in optimizer.hyperparam_history])
xs.T[0]
import plottool as pt
pt.qtensure()
pt.plt.clf()
for i in range(len(names)):
pt.plt.plot(xs.T[i], ys, 'o', label=names[i])
pt.plt.legend()
def junk():
pass
import optml
from clab import util
from os.path import join, splitext, basename # NOQA
import glob
import ubelt as ub
import itertools as it
import numpy as np
from clab.live.urban_mapper import instance_fscore
path = ub.truepath(
'~/remote/aretha/data/work/urban_mapper2/test/input_4224-exkudlzu/'
'solver_4214-guwsobde_unet_mmavmuou_eqnoygqy_a=1,c=RGB,n_ch=5,n_cl=4/'
'_epoch_00000154/restiched/pred')
mode_paths = sorted(glob.glob(path + '/*.png'))
results = ub.odict()
"""
# 40/40... rate=0.86 Hz, eta=0:00:00, total=0:00:46, wall=16:29 EST
# ------------------------------------
# best_score = '[0.60784, 0.84042, 0.47841]'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# res = '[0.60784, 0.84042, 0.47841]'
# 40/40... rate=0.90 Hz, eta=0:00:00, total=0:00:44, wall=16:29 EST
# ------------------------------------
# best_score = '[0.60784, 0.84042, 0.47841]'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 93))'
# res = '[0.60543, 0.86952, 0.46611]'
# 40/40... rate=0.89 Hz, eta=0:00:00, total=0:00:44, wall=16:30 EST
# ------------------------------------
# best_score = '[0.60784, 0.84042, 0.47841]'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 100))'
# res = '[0.60316, 0.87241, 0.46250]'
# 40/40... rate=0.93 Hz, eta=0:00:00, total=0:00:42, wall=16:31 EST
# ------------------------------------
# best_score = '[0.60784, 0.84042, 0.47841]'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# key = '((only_inner, True), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# res = '[0.57623, 0.77876, 0.45955]'
# 40/40... rate=0.98 Hz, eta=0:00:00, total=0:00:40, wall=16:32 EST
# ------------------------------------
# best_score = '[0.60784, 0.84042, 0.47841]'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# key = '((only_inner, True), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 93))'
# res = '[0.58619, 0.83016, 0.45472]'
# 40/40... rate=1.01 Hz, eta=0:00:00, total=0:00:39, wall=16:32 EST
# ------------------------------------
# best_score = '[0.60784, 0.84042, 0.47841]'
# key = '((only_inner, False), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 50))'
# key = '((only_inner, True), (inner_k, 0), (post_k, 0), (outer_k, 0), (min_seed_size, 100))'
# res = '[0.58660, 0.83831, 0.45268]'
"""
param_space = [
# optml.Parameter(name='s', param_type='boolean'),
# optml.Parameter(name='w', param_type='boolean'),
optml.Parameter(name='only_inner', param_type='categorical', possible_values=[False, True]),
optml.Parameter(name='inner_k', param_type='categorical', possible_values=[0]),
optml.Parameter(name='post_k', param_type='categorical', possible_values=[0, 3, 5, 7, 9, 12, 15][0:1]),
optml.Parameter(name='outer_k', param_type='categorical', possible_values=[0, 3, 5, 7, 9, 12, 15][0:1]),
optml.Parameter(name='min_seed_size', param_type='categorical',
possible_values=[50, 93, 100]),
# optml.Parameter(name='d', param_type='integer', lower=1, upper=5),
# optml.Parameter(name='n', param_type='integer', lower=1, upper=1),
]
from clab.tasks.urban_mapper_3d import UrbanMapper3D
task = UrbanMapper3D('', '', boundary=True)
def itergrid():
names = [p.name for p in param_space]
for values in it.product(*map(iter, param_space)):
yield ub.odict(zip(names, values))
def instance_label2(pred_seg, only_inner=False, inner_k=0, outer_k=0,
post_k=0, min_seed_size=0):
import cv2
inner = (pred_seg == task.classname_to_id['inner_building']).astype(np.uint8)
# outer = (pred_seg == task.classname_to_id['outer_building']).astype(np.uint8)
# outer_k = 15
# outer = cv2.morphologyEx(outer, cv2.MORPH_ERODE,
# np.ones((outer_k, outer_k), np.uint8),
# iterations=1)
if inner_k > 0:
kernel = np.ones((inner_k, inner_k), np.uint8)
inner = cv2.morphologyEx(inner, cv2.MORPH_OPEN, kernel,
iterations=1)
def cc_locs(mask):
ccs = cv2.connectedComponents(mask.astype(np.uint8), connectivity=4)[1]
rc_locs = np.where(mask > 0)
rc_ids = ccs[rc_locs]
rc_arrs = np.ascontiguousarray(np.vstack(rc_locs).T)
cc_to_loc = util.group_items(rc_arrs, rc_ids, axis=0)
return cc_to_loc
if min_seed_size > 0:
# Remove very small seeds
for inner_id, inner_rcs in cc_locs(inner).items():
if len(inner_rcs) < min_seed_size:
inner[tuple(inner_rcs.T)] = 0
seeds = cv2.connectedComponents(inner, connectivity=4)[1]
if only_inner:
return seeds
mask = (pred_seg > 0).astype(np.uint8)
if outer_k > 1:
mask = cv2.morphologyEx(mask, cv2.MORPH_ERODE,
np.ones((outer_k, outer_k), np.uint8),
iterations=1)
# Ensure we dont clobber a seed
mask[inner] = 1
dmask1 = cv2.dilate(mask, np.ones((3, 3)))
dmask2 = cv2.dilate(dmask1, np.ones((3, 3)))
# Build a topological wall between mask components
twall = dmask1 - mask
# Pixels beyond the wall region are sure background
sure_bg = 1 - dmask2
# prepare watershed seeds
# Label sure background as 1
wseed = sure_bg.astype(np.int)
# Add the seeds starting at 2
seed_mask = seeds > 0
seed_labels = seeds[seed_mask]
wseed[seed_mask] = seed_labels + 1
# The unsure region is now labeled as zero
"""
from clab.torch.urban_mapper import draw_gt_contours
wseed_color = task.instance_colorize(wseed)
twall_alpha = util.ensure_alpha_channel(twall * 255, alpha=1)
twall_alpha[twall == 0, 3] = 0
twall_alpha[twall == 1, 3]
color_seed_wall = util.imutil.overlay_alpha_images(twall_alpha, wseed_color)
pt.imshow(color_seed_wall)
draw_img = draw_gt_contours(color_seed_wall, gti)
pt.imshow(draw_img)
"""
topology = np.dstack([twall * 255] * 3)
markers = np.ascontiguousarray(wseed.astype(np.int32).copy())
markers = cv2.watershed(topology, markers)
# Remove background and border labels
markers[markers <= 1] = 0
"""
color_markers = task.instance_colorize(markers)
pt.imshow(draw_gt_contours(color_markers, gti))
color_seed_wall_ = util.ensure_alpha_channel(color_seed_wall[:, :, 0:3], alpha=.6)
overlay_markers = util.imutil.overlay_alpha_images(color_seed_wall_, color_markers)
pt.imshow(overlay_markers)
pt.imshow(color_markers)
draw_img = draw_gt_contours(overlay_markers, gti)
pt.imshow(draw_img)
"""
instance_mask = (markers > 0).astype(np.uint8)
if post_k > 0:
mask = cv2.morphologyEx(mask, cv2.MORPH_ERODE,
np.ones((post_k, post_k), np.uint8),
iterations=1)
pred_ccs = cv2.connectedComponents(instance_mask, connectivity=4)[1]
return pred_ccs
best_key = None
best_score = [-np.inf]
for params in itergrid():
key = tuple(params.items())
if key not in results:
scores = []
for pred_fpath in ub.ProgIter(mode_paths[10:50]):
gtl_fname = basename(pred_fpath).replace('.png', '_GTL.tif')
gti_fname = basename(pred_fpath).replace('.png', '_GTI.tif')
dsm_fname = basename(pred_fpath).replace('.png', '_DSM.tif')
# bgr_fname = basename(pred_fpath).replace('.png', '_RGB.tif')
gtl_fpath = join(ub.truepath('~/remote/aretha/data/UrbanMapper3D/training/'), gtl_fname)
gti_fpath = join(ub.truepath('~/remote/aretha/data/UrbanMapper3D/training/'), gti_fname)
dsm_fpath = join(ub.truepath('~/remote/aretha/data/UrbanMapper3D/training/'), dsm_fname)
# bgr_fpath = join(ub.truepath('~/remote/aretha/data/UrbanMapper3D/training/'), bgr_fname)
pred_seg = util.imread(pred_fpath)
gti = util.imread(gti_fpath)
gtl = util.imread(gtl_fpath)
dsm = util.imread(dsm_fpath)
# bgr = util.imread(bgr_fpath)
pred = instance_label2(pred_seg, **params)
uncertain = (gtl == 65)
score = instance_fscore(gti, uncertain, dsm, pred)
scores.append(score)
res = np.array(scores).mean(axis=0)
if res[0] > best_score[0]:
best_score = res
best_key = key
print('------------------------------------')
print('best_score = {!r}'.format(ub.repr2(list(best_score), precision=5, nl=0)))
print('key = {!r}'.format(ub.repr2(best_key, si=True, nl=0)))
print('key = {!r}'.format(ub.repr2(key, si=True, nl=0)))
print('res = {!r}'.format(ub.repr2(list(res), precision=5, nl=0)))
results[key] = res
import pandas as pd
rdf = pd.DataFrame(list(results.values()),
index=results.keys(),
columns=['f1', 'precision', 'recall'])
rdf.sort_values('f1').index[-1]
import optml
from optml.bayesian_optimizer import BayesianOptimizer
hyperparams = [
optml.Parameter(name='mask_thresh', param_type='continuous', lower=.15, upper=.85),
optml.Parameter(name='seed_thresh', param_type='continuous', lower=.15, upper=.85),
optml.Parameter(name='min_seed_size', param_type='integer', lower=0, upper=100),
]
self = BayesianOptimizer(model=None, hyperparams=hyperparams, eval_func=None)
for path, path1 in zip(prob_paths, prob1_paths):
pass
def func(**new_hyperparams):
probs = np.load(path)['arr_0']
probs1 = np.load(path1)['arr_0']
seed_probs = probs[:, :, task.classname_to_id['inner_building']]
mask_probs = probs1[:, :, 1]
seed_thresh, mask_thresh, min_seed_size = ub.take(
new_hyperparams, ['seed_thresh', 'mask_thresh', 'min_seed_size'])
seed = (seed_probs > mask_thresh).astype(np.uint8)
mask = (mask_probs > seed_thresh).astype(np.uint8)
pred = seeded_instance_label(seed, mask, min_seed_size=min_seed_size)
scores = instance_fscore(gti, uncertain, dsm, pred)
fscore = scores[0]
return fscore
history = self.simple(func, n_iters=10, verbose=True)
best = max(history)
print('best = {!r}'.format(best))
# {'mask_thresh': 0.45318221555100013, 'seed_thresh': 0.69172340500683327, 'min_seed_size': 41}
func(**{'mask_thresh': 0.5, 'seed_thresh': 0.7, 'min_seed_size': 20})
data_format='channels_first'))
model.add(Convolution2D(filters=32, kernel_size=(
3,
3,
), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(units=128, activation='relu'))
#model.add(Dropout(0.5))
model.add(Dense(10, activation='softmax'))
params = [
optml.Parameter(name='filters', param_type='integer', lower=20, upper=50),
optml.Parameter(name='units', param_type='integer', lower=100, upper=150)
]
def clf_score(y_true, y_pred):
mse = ((y_true - y_pred)**2).mean(axis=None)
cce = sklearn.metrics.log_loss(y_true, y_pred)
return cce
bayesOpt = BayesianOptimizer(model=model,
hyperparams=params,
eval_func=clf_score)
bayes_best_params, bayes_best_model = bayesOpt.fit(X_train=X_train,