def msrc_over_weak(n_train_full=40, n_train=276,
C=100, alpha=0.1,
test_method='gco', test_n_iter=5, n_iter=5,
max_iter=1000, verbose=1,
check_every=50, complete_every=100, update_w_every=50,
relaxed_test=False,
use_latent_first_iter=100):
# save parameters as meta
meta_data = locals()
logger = logging.getLogger(__name__)
crf_test = HCRF(n_states=24, n_features=2028, n_edge_features=4, alpha=alpha,
inference_method=test_method, n_iter=test_n_iter)
crf_latent = HCRF(n_states=24, n_features=2028, n_edge_features=4, alpha=alpha,
inference_method='gco', n_iter=n_iter)
trainer = OverWeak(crf_latent, n_states=24, n_features=2028, n_edge_features=4,
C=C, alpha=alpha,
max_iter=max_iter, verbose=verbose,
check_every=check_every, complete_every=complete_every, update_w_every=update_w_every)
x_train, y_train, y_train_full, x_test, y_test = load_msrc(n_train_full, n_train, dense=True)
logger.info('start training')
start = time()
trainer.fit(x_train, y_train,
train_scorer=lambda w: compute_score(crf_test, w, x_train, y_train_full, relaxed=relaxed_test),
test_scorer=lambda w: compute_score(crf_test, w, x_test, y_test, relaxed=relaxed_test),
use_latent_first_iter=use_latent_first_iter)
stop = time()
time_elapsed = stop - start
logger.info('testing')
test_score = compute_score(crf_test, trainer.w, x_test, y_test)
train_score = compute_score(crf_test, trainer.w, x_train, y_train_full)
logger.info('========================================')
logger.info('train score: %f', train_score)
logger.info('test score: %f', test_score)
exp_data = {}
exp_data['timestamps'] = trainer.timestamps
exp_data['objective'] = trainer.objective_curve
exp_data['w'] = trainer.w
exp_data['train_scores'] = trainer.train_score
exp_data['test_scores'] = trainer.test_score
exp_data['w_history'] = trainer.w_history
meta_data['dataset_name'] = 'msrc'
meta_data['annotation_type'] = 'full+weak'
meta_data['label_type'] = 'image-level labelling'
meta_data['trainer'] = 'komodakis+latent+kappa'
meta_data['train_score'] = train_score
meta_data['test_score'] = test_score
meta_data['time_elapsed'] = time_elapsed
return ExperimentResult(exp_data, meta_data)
def msrc_over_weak(n_train_full=40,
n_train=276,
C=100,
alpha=0.1,
test_method='gco',
test_n_iter=5,
n_iter=5,
max_iter=1000,
verbose=1,
check_every=50,
complete_every=100,
update_w_every=50,
relaxed_test=False,
use_latent_first_iter=100):
# save parameters as meta
meta_data = locals()
logger = logging.getLogger(__name__)
crf_test = HCRF(n_states=24,
n_features=2028,
n_edge_features=4,
alpha=alpha,
inference_method=test_method,
n_iter=test_n_iter)
crf_latent = HCRF(n_states=24,
n_features=2028,
n_edge_features=4,
alpha=alpha,
inference_method='gco',
n_iter=n_iter)
trainer = OverWeak(crf_latent,
n_states=24,
n_features=2028,
n_edge_features=4,
C=C,
alpha=alpha,
max_iter=max_iter,
verbose=verbose,
check_every=check_every,
complete_every=complete_every,
update_w_every=update_w_every)
x_train, y_train, y_train_full, x_test, y_test = load_msrc(n_train_full,
n_train,
dense=True)
logger.info('start training')
start = time()
trainer.fit(x_train,
y_train,
train_scorer=lambda w: compute_score(
crf_test, w, x_train, y_train_full, relaxed=relaxed_test),
test_scorer=lambda w: compute_score(
crf_test, w, x_test, y_test, relaxed=relaxed_test),
use_latent_first_iter=use_latent_first_iter)
stop = time()
time_elapsed = stop - start
logger.info('testing')
test_score = compute_score(crf_test, trainer.w, x_test, y_test)
train_score = compute_score(crf_test, trainer.w, x_train, y_train_full)
logger.info('========================================')
logger.info('train score: %f', train_score)
logger.info('test score: %f', test_score)
exp_data = {}
exp_data['timestamps'] = trainer.timestamps
exp_data['objective'] = trainer.objective_curve
exp_data['w'] = trainer.w
exp_data['train_scores'] = trainer.train_score
exp_data['test_scores'] = trainer.test_score
exp_data['w_history'] = trainer.w_history
meta_data['dataset_name'] = 'msrc'
meta_data['annotation_type'] = 'full+weak'
meta_data['label_type'] = 'image-level labelling'
meta_data['trainer'] = 'komodakis+latent+kappa'
meta_data['train_score'] = train_score
meta_data['test_score'] = test_score
meta_data['time_elapsed'] = time_elapsed
return ExperimentResult(exp_data, meta_data)
def syntetic_over_weak(n_train_full=10,
n_train=100,
C=1,
dataset=1,
max_iter=100,
verbose=1,
test_samples=10,
check_every=10,
test_method='gco',
test_n_iter=5,
relaxed_test=False,
alpha=1,
n_iter=5,
complete_every=10,
update_w_every=5,
update_mu=20,
use_latent_first_iter=500,
undergenerating_weak=False,
smd=False):
# save parameters as meta
meta_data = locals()
logger = logging.getLogger(__name__)
crf_test = HCRF(n_states=10,
n_features=10,
n_edge_features=2,
inference_method=test_method,
n_iter=test_n_iter)
crf_latent = HCRF(n_states=10,
n_features=10,
n_edge_features=2,
alpha=alpha,
inference_method='gco',
n_iter=n_iter)
trainer = OverWeak(crf_latent,
n_states=10,
n_features=10,
n_edge_features=2,
C=C,
max_iter=max_iter,
verbose=verbose,
check_every=check_every,
complete_every=complete_every,
alpha=alpha,
update_w_every=update_w_every,
update_mu=update_mu)
x_train, y_train, y_train_full, x_test, y_test = \
load_syntetic(dataset, n_train_full, n_train)
x_test = x_test[:test_samples]
y_test = y_test[:test_samples]
logger.info('start training')
start = time()
trainer.fit(x_train,
y_train,
train_scorer=lambda w: compute_score(
crf_test, w, x_train, y_train_full, relaxed=relaxed_test),
test_scorer=lambda w: compute_score(
crf_test, w, x_test, y_test, relaxed=relaxed_test),
decompose='grid',
use_latent_first_iter=use_latent_first_iter,
undergenerating_weak=undergenerating_weak,
smd=smd)
stop = time()
time_elapsed = stop - start
logger.info('testing')
test_score = compute_score(crf_test, trainer.w, x_test, y_test)
train_score = compute_score(crf_test, trainer.w, x_train, y_train_full)
logger.info('========================================')
logger.info('train score: %f', train_score)
logger.info('test score: %f', test_score)
exp_data = {}
exp_data['timestamps'] = trainer.timestamps
exp_data['objective'] = trainer.objective_curve
exp_data['w'] = trainer.w
exp_data['train_scores'] = trainer.train_score
exp_data['test_scores'] = trainer.test_score
exp_data['w_history'] = trainer.w_history
meta_data['dataset_name'] = 'syntetic'
meta_data['annotation_type'] = 'full+weak'
meta_data['label_type'] = 'image-level labelling'
meta_data['trainer'] = 'komodakis+latent+kappa'
meta_data['train_score'] = train_score
meta_data['test_score'] = test_score
meta_data['time_elapsed'] = time_elapsed
return ExperimentResult(exp_data, meta_data)
def syntetic_over_weak(n_train_full=10, n_train=100, C=1, dataset=1,
max_iter=100, verbose=1,
test_samples=10, check_every=10,
test_method='gco', test_n_iter=5, relaxed_test=False,
alpha=1, n_iter=5, complete_every=10,
update_w_every=5, update_mu=20,
use_latent_first_iter=500, undergenerating_weak=False,
smd=False):
# save parameters as meta
meta_data = locals()
logger = logging.getLogger(__name__)
crf_test = HCRF(n_states=10, n_features=10, n_edge_features=2,
inference_method=test_method, n_iter=test_n_iter)
crf_latent = HCRF(n_states=10, n_features=10, n_edge_features=2,
alpha=alpha, inference_method='gco', n_iter=n_iter)
trainer = OverWeak(crf_latent, n_states=10, n_features=10, n_edge_features=2,
C=C, max_iter=max_iter, verbose=verbose, check_every=check_every,
complete_every=complete_every, alpha=alpha, update_w_every=update_w_every,
update_mu=update_mu)
x_train, y_train, y_train_full, x_test, y_test = \
load_syntetic(dataset, n_train_full, n_train)
x_test = x_test[:test_samples]
y_test = y_test[:test_samples]
logger.info('start training')
start = time()
trainer.fit(x_train, y_train,
train_scorer=lambda w: compute_score(crf_test, w, x_train, y_train_full, relaxed=relaxed_test),
test_scorer=lambda w: compute_score(crf_test, w, x_test, y_test, relaxed=relaxed_test),
decompose='grid',
use_latent_first_iter=use_latent_first_iter, undergenerating_weak=undergenerating_weak,
smd=smd)
stop = time()
time_elapsed = stop - start
logger.info('testing')
test_score = compute_score(crf_test, trainer.w, x_test, y_test)
train_score = compute_score(crf_test, trainer.w, x_train, y_train_full)
logger.info('========================================')
logger.info('train score: %f', train_score)
logger.info('test score: %f', test_score)
exp_data = {}
exp_data['timestamps'] = trainer.timestamps
exp_data['objective'] = trainer.objective_curve
exp_data['w'] = trainer.w
exp_data['train_scores'] = trainer.train_score
exp_data['test_scores'] = trainer.test_score
exp_data['w_history'] = trainer.w_history
meta_data['dataset_name'] = 'syntetic'
meta_data['annotation_type'] = 'full+weak'
meta_data['label_type'] = 'image-level labelling'
meta_data['trainer'] = 'komodakis+latent+kappa'
meta_data['train_score'] = train_score
meta_data['test_score'] = test_score
meta_data['time_elapsed'] = time_elapsed
return ExperimentResult(exp_data, meta_data)
