def get_instance(ti_line):
j = json.loads(ti_line)
ti = TrainingInstance.from_dict(j)
obs = [
o.l2_word for o in ti.current_sent if o.lang == 'de'
if o.l2_word.strip() != ''
]
guesses = [g.guess for g in ti.current_guesses if g.guess.strip() != '']
guesses += [
o.l2_word for o in ti.current_sent if o.lang == 'en'
if o.l2_word.strip() != ''
] # wink! ;)
guesses += [
g.guess for g in ti.past_correct_guesses if g.guess.strip() != ''
]
guesses += [
g.guess for g in ti.past_guesses_for_current_sent
if g.guess.strip() != ''
]
guesses = [g.split() for g in guesses]
obs = [o.split() for o in obs]
guesses_flat = sum(guesses, [])
#for gf in guesses_flat:
# add_to_tags(gf)
obs_flat = sum(obs, [])
for of in obs_flat:
add_to_obs(of)
return ti, obs_flat, guesses_flat
def batch_predictions(training_instance,
theta_en_en_names, theta_en_de_names,
theta_en_en, theta_en_de,
phi_wapper, lr,
en_domain, de2id, en2id, d2t, qp=False):
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
sent_id = ti.current_sent[0].sent_id
fg = create_factor_graph(ti=ti,
learning_rate=lr,
theta_en_en_names=theta_en_en_names,
theta_en_de_names=theta_en_de_names,
theta_en_de=theta_en_de,
theta_en_en=theta_en_en,
phi_wrapper=phi_wapper,
en_domain=en_domain,
de2id=de2id,
en2id=en2id, d2t=d2t)
fg.initialize()
fg.treelike_inference(3)
p = fg.get_posterior_probs()
if qp:
factor_dist = None
fgs = None
else:
fgs = '\n'.join(['*SENT_ID:' + str(sent_id)] + fg.to_string())
factor_dist = fg.to_dist()
p0,p25, p50, t = fg.get_precision_counts()
return [p, fgs, factor_dist, (p0, p25, p50,t)]
def batch_predictions(training_instance,
f_en_en_theta,
f_en_de_theta,
adapt_phi_en_en,
adapt_phi_en_de, lr,
en_domain,
de2id,
en2id,
basic_f_en_en,
basic_f_en_de,
domain2theta):
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
sent_id = ti.current_sent[0].sent_id
fg = create_factor_graph(ti=ti,
learning_rate=lr,
theta_en_de=f_en_de_theta,
theta_en_en=f_en_en_theta,
phi_en_en=adapt_phi_en_en,
phi_en_de=adapt_phi_en_de,
basic_f_en_en=basic_f_en_en,
basic_f_en_de=basic_f_en_de,
en_domain=en_domain,
de2id=de2id,
en2id=en2id,
domain2theta=domain2theta)
fg.initialize()
fg.treelike_inference(3)
return fg.get_posterior_probs()
def batch_sgd(training_instance,
theta_en_en,
theta_en_de,
phi_en_en,
phi_en_de, lr,
en_domain,
de2id,
en2id,
basic_f_en_en,
basic_f_en_de,
domain2theta):
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
sent_id = ti.current_sent[0].sent_id
fg = create_factor_graph(ti=ti,
learning_rate=lr,
theta_en_de=theta_en_de,
theta_en_en=theta_en_en,
phi_en_en=phi_en_en,
phi_en_de=phi_en_de,
basic_f_en_en=basic_f_en_en,
basic_f_en_de=basic_f_en_de,
en_domain=en_domain,
de2id=de2id,
en2id=en2id,
domain2theta=domain2theta)
fg.initialize()
# sys.stderr.write('.')
fg.treelike_inference(3)
# sys.stderr.write('.')
# f_en_en_theta, f_en_de_theta = fg.update_theta()
g_en_en, g_en_de = fg.get_unregularized_gradeint()
sample_ag = {}
for f_type, u in fg.active_domains:
g = g_en_en.copy() if f_type == 'en_en' else g_en_de.copy()
t = domain2theta[f_type, u]
r = fg.regularization_param
l = fg.learning_rate
sample_ag[f_type, u] = apply_regularization(r * 0.001, g, l, t) # use a smaller regularization term
g_en_en = apply_regularization(r, g_en_en, l, fg.theta_en_en)
g_en_de = apply_regularization(r, g_en_de, l, fg.theta_en_de)
# turn off adapt_phi
return [sent_id, g_en_en, g_en_de, sample_ag]
def get_instance(ti_line):
j = json.loads(ti_line)
ti = TrainingInstance.from_dict(j)
obs = [o.l2_word for o in ti.current_sent if o.lang == 'de' if o.l2_word.strip() != '']
guesses = [g.guess for g in ti.current_guesses if g.guess.strip() != '']
guesses += [o.l2_word for o in ti.current_sent if o.lang == 'en' if o.l2_word.strip() != ''] # wink! ;)
guesses += [g.guess for g in ti.past_correct_guesses if g.guess.strip() != '']
guesses += [g.guess for g in ti.past_guesses_for_current_sent if g.guess.strip() != '']
guesses = [g.split() for g in guesses]
obs = [o.split() for o in obs]
guesses_flat = sum(guesses, [])
#for gf in guesses_flat:
# add_to_tags(gf)
obs_flat = sum(obs, [])
for of in obs_flat:
add_to_obs(of)
return ti, obs_flat, guesses_flat
def batch_sgd(training_instance,
theta_en_en_names, theta_en_de_names,
theta_en_en, theta_en_de,
phi_wrapper, lr,
en_domain, de2id, en2id, d2t):
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
sent_id = ti.current_sent[0].sent_id
fg = create_factor_graph(ti=ti,
learning_rate=lr,
theta_en_en_names=theta_en_en_names,
theta_en_de_names=theta_en_de_names,
theta_en_de=theta_en_de,
theta_en_en=theta_en_en,
phi_wrapper=phi_wrapper,
en_domain=en_domain,
de2id=de2id,
en2id=en2id,
d2t=d2t)
fg.initialize()
fg.treelike_inference(3)
# f_en_en_theta, f_en_de_theta = fg.update_theta()
if options.user_adapt or options.experience_adapt:
g_en_en, g_en_de = fg.get_unregularized_gradeint()
sample_ag = {}
for f_type, d in fg.active_domains:
g = g_en_en.copy() if f_type == 'en_en' else g_en_de.copy()
t = domain2theta[f_type, d]
r = fg.regularization_param
l = fg.learning_rate
scale_reg = float(options.reg_param_ua_scale)
sample_ag[f_type, d] = apply_regularization(r * scale_reg, g, l, t) # use a smaller regularization term
g_en_en = apply_regularization(r, g_en_en, l, fg.theta_en_en)
g_en_de = apply_regularization(r, g_en_de, l, fg.theta_en_de)
else:
sample_ag = None
g_en_en, g_en_de = fg.return_gradient()
fg.display_timing_info()
p = fg.get_posterior_probs()
return [sent_id, p, g_en_en, g_en_de, sample_ag]
def batch_predictions(training_instance, f_en_en_theta, f_en_de_theta,
adapt_phi_en_en, adapt_phi_en_de, lr, en_domain, de2id,
en2id, basic_f_en_en, basic_f_en_de, domain2theta):
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
sent_id = ti.current_sent[0].sent_id
fg = create_factor_graph(ti=ti,
learning_rate=lr,
theta_en_de=f_en_de_theta,
theta_en_en=f_en_en_theta,
phi_en_en=adapt_phi_en_en,
phi_en_de=adapt_phi_en_de,
basic_f_en_en=basic_f_en_en,
basic_f_en_de=basic_f_en_de,
en_domain=en_domain,
de2id=de2id,
en2id=en2id,
domain2theta=domain2theta)
fg.initialize()
fg.treelike_inference(3)
return fg.get_posterior_probs()
def batch_sgd(training_instance, theta_en_en, theta_en_de, phi_en_en,
phi_en_de, lr, en_domain, de2id, en2id, basic_f_en_en,
basic_f_en_de, domain2theta):
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
sent_id = ti.current_sent[0].sent_id
fg = create_factor_graph(ti=ti,
learning_rate=lr,
theta_en_de=theta_en_de,
theta_en_en=theta_en_en,
phi_en_en=phi_en_en,
phi_en_de=phi_en_de,
basic_f_en_en=basic_f_en_en,
basic_f_en_de=basic_f_en_de,
en_domain=en_domain,
de2id=de2id,
en2id=en2id,
domain2theta=domain2theta)
fg.initialize()
# sys.stderr.write('.')
fg.treelike_inference(3)
# sys.stderr.write('.')
# f_en_en_theta, f_en_de_theta = fg.update_theta()
g_en_en, g_en_de = fg.get_unregularized_gradeint()
sample_ag = {}
for f_type, u in fg.active_domains:
g = g_en_en.copy() if f_type == 'en_en' else g_en_de.copy()
t = domain2theta[f_type, u]
r = fg.regularization_param
l = fg.learning_rate
sample_ag[f_type, u] = apply_regularization(
r * 0.001, g, l, t) # use a smaller regularization term
g_en_en = apply_regularization(r, g_en_en, l, fg.theta_en_en)
g_en_de = apply_regularization(r, g_en_de, l, fg.theta_en_de)
# turn off adapt_phi
return [sent_id, g_en_en, g_en_de, sample_ag]
# pre_fire_en_en = sparse.csr_matrix(pre_fire_en_en)
f_en_de_theta = np.zeros((1, len(f_en_de)))
phi_en_de = np.random.rand(len(en_domain) * len(de_domain), len(f_en_de))
phi_en_de[phi_en_de > 0.5] = 1.0
phi_en_de[phi_en_de < 0.2] = 0.0
load_times = []
grad_times = []
inference_times = []
mp_times = []
fg = None
split_ratio = int(len(training_instances) * 0.33)
test_instances = training_instances[:split_ratio]
all_training_instances = training_instances[split_ratio:]
lr = 0.1
for t_idx, training_instance in enumerate(training_instances):
print t_idx
j_ti = json.loads(training_instance)
ti = TrainingInstance.from_dict(j_ti)
lt = time.time()
fg = FactorGraph(
theta_en_en=f_en_en_theta if fg is None else fg.theta_en_en,
theta_en_de=f_en_de_theta if fg is None else fg.theta_en_de,
phi_en_en=phi_en_en,
phi_en_de=phi_en_de)
fg.learning_rate = lr
fg = load_fg(fg, ti, en_domain, de2id=de2id, en2id=en2id)
print 'done checking', len(training_instances)
tag_list = list(set(to.split('\n')[0].split()))
obs_list = list(set(to.split('\n')[1].split()))
fac_cell_2feat = {}
feat2id = {}
for features_fired_in_factor in factors.strip().split('FACTOR:'):
feature_lines = features_fired_in_factor.strip().split('\n')
fac_type = feature_lines[0].strip()
for fl in feature_lines[1:]:
items = fl.split()
label1, label2 = items[0], items[1]
fac_cell_2feat[fac_type, label1, label2] = [(f_name, 1.0) for f_name in items[2:]]
for f_fired in items[2:]:
feat2id[f_fired] = feat2id.get(f_fired, len(feat2id))
return tag_list, obs_list, fac_cell_2feat, feat2id
if __name__ == '__main__':
opt = OptionParser()
# insert options here
opt.add_option('--ti', dest='training_instances', default='')
(options, _) = opt.parse_args()
if options.training_instances == '':
sys.stderr.write("Usage: jython macaronic-tagger.py --ti [training instances file]\n")
exit(1)
for line in codecs.open(options.training_instances, 'r', 'utf8').readlines():
jti = json.loads(line)
ti = TrainingInstance.from_dict(jti)
pdb.set_trace()
