def run_all():
language_pairs = "de-en", "en-de"
tab_fnames = ( join(config["private_data_dir"], "corpmod/de/en/de-en_ambig.tab"),
join(config["private_data_dir"], "corpmod/en/de/en-de_ambig.tab"))
descriptor = {'names': ('lang_pair', 'k', 'NIST', 'BLUE'),
'formats': ('S8', 'i4', 'f4', 'f4')}
#percentiles = [0.25, 0.5, 1, 2.5, 5, 10, 20, 25, 30, 40, 50, 60, 70, 80, 90]
#p_values = [0.1, 0.05, 0.01, 0.005, 0.001]
#p_values = [0.0005, 0.0001, 0.00005, 0.00001]
#p_values = [0.00005, 0.00001, 0.000005, 0.000001, 0.0000005, 0.0000001]
#p_values = [0.00000005, 0.00000001, 0.000000005, 0.000000001, 0.0000000005, 0.000000001]
#k_values = [1,2,3,4,5,10,25,50,100,250,500,1000]
k_values = [150,200,300,350,400,450]
results = np.zeros(9999, dtype=descriptor)
exp_count = 0
for lang_pair, tab_fname in zip(language_pairs, tab_fnames):
target_lang = lang_pair.split("-")[1]
samp_fname = target_lang + "_samples_subset_filtered.hdf5"
graphs_pkl_fname = "prep/{}_graphs.pkl".format(lang_pair)
target_lang = lang_pair.split("-")[1]
counts_pkl_fname = config["count"]["lemma"][target_lang]["pkl_fname"]
for k in k_values:
results[exp_count] = ( lang_pair,
k,
0,
0
)
exp_dir = "exp_" + "_".join("{}={}".format(var, value)
for var, value in zip(results.dtype.names,
results[exp_count])[:-2])
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
#else:
# continue
classifier = MultinomialNB()
models_fname = join(exp_dir, "nb_models.hdf5")
builder = NBModelBuilder(tab_fname, samp_fname, models_fname,
classifier, graphs_pkl_fname=graphs_pkl_fname,
counts_pkl_fname=counts_pkl_fname,
feat_selector=SelectKBest(chi2, k))
builder.run()
nist, blue = score_model(lang_pair, exp_dir, draw=False)
results[exp_count]["NIST"] = nist
results[exp_count]["BLUE"] = blue
exp_count += 1
results = results[:exp_count]
print results
results.dump("nb_feat_select_bst_results_2.pkl")
def run_all():
language_pairs = "en-de", "de-en"
tab_fnames = ( join(config["private_data_dir"], "corpmod/en/de/en-de_ambig.tab"),
join(config["private_data_dir"], "corpmod/de/en/de-en_ambig.tab") )
descriptor = {'names': ('lang_pair', 'vocab_i', 'vocab_j', 'NIST', 'BLUE'),
'formats': ('S8','i4', 'i4', 'f4', 'f4')}
results = np.zeros(9999, dtype=descriptor)
exp_count = 0
for lang_pair, tab_fname in zip(language_pairs, tab_fnames):
target_lang = lang_pair.split("-")[1]
samp_fname = target_lang + "_samples_subset_filtered.hdf5"
graphs_pkl_fname = "prep/{}_graphs.pkl".format(lang_pair)
lempos_subset = extract_source_lempos_subset(graphs_pkl_fname)
target_lang = lang_pair.split("-")[1]
counts_pkl_fname = config["count"]["lemma"][target_lang]["pkl_fname"]
for vocab_i, vocab_j in cut_vocab(samp_fname, counts_pkl_fname):
results[exp_count] = ( lang_pair,
vocab_i,
vocab_j,
0,
0
)
exp_dir = "exp_" + "_".join("{}={}".format(var, value)
for var, value in zip(results.dtype.names,
results[exp_count])[:-2])
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
classifier = MultinomialNB()
models_fname = join(exp_dir, "nb_models.hdf5")
make_models(tab_fname, samp_fname, models_fname, classifier,
source_lempos_subset=lempos_subset,
counts_pkl_fname=counts_pkl_fname,
vocab_i=vocab_i, vocab_j=vocab_j)
nist, blue = score_model(lang_pair, exp_dir, draw=False)
results[exp_count]["NIST"] = nist
results[exp_count]["BLUE"] = blue
exp_count += 1
results = results[:exp_count]
print results
results.dump("nb_cut_vocab_results.pkl")
def run_all():
language_pairs = "en-de", "de-en"
tab_fnames = ( join(config["private_data_dir"], "corpmod/en/de/en-de_ambig.tab"),
join(config["private_data_dir"], "corpmod/de/en/de-en_ambig.tab") )
extended_vectors = True, #False
classifier_types = MultinomialNB, BernoulliNB
alpha_values = 1.0, 0.1, 0.01, np.finfo(np.double).eps
corpus_prior_values = False, True
# the fit_prior parameters seems to make absolutely no difference
# fit_prior_values = True, False
descriptor = {'names': ('lang_pair', 'classifier', 'alpha',
'corpus_priors', 'extended', 'NIST', 'BLUE'),
'formats': ('S8', 'S64', 'f4', 'b', 'b', 'f4', 'f4')}
results = np.zeros(9999, dtype=descriptor)
exp_count = 0
for lang_pair, tab_fname in zip(language_pairs, tab_fnames):
graphs_pkl_fname = "prep/{}_graphs.pkl".format(lang_pair)
lempos_subset = extract_source_lempos_subset(graphs_pkl_fname)
target_lang = lang_pair.split("-")[1]
for extended in extended_vectors:
if extended:
samp_fname = target_lang + "_samples_subset_filtered_extended.hdf5"
else:
samp_fname = target_lang + "_samples_subset_filtered.hdf5"
for corpus_prior in corpus_prior_values:
if corpus_prior:
counts_pkl_fname = config["count"]["lemma"][target_lang]["pkl_fname"]
else:
counts_pkl_fname = None
for classifier_class in classifier_types:
for alpha in alpha_values:
results[exp_count] = ( lang_pair,
classifier_class.__name__,
alpha,
corpus_prior,
extended,
0,
0
)
exp_dir = "exp_" + "_".join("{}={}".format(var, value)
for var, value in zip(results.dtype.names,
results[exp_count])[:-2])
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
classifier = classifier_class(alpha=alpha)
models_fname = join(exp_dir, "nb_models.hdf5")
make_models(tab_fname, samp_fname, models_fname, classifier,
source_lempos_subset=lempos_subset,
counts_pkl_fname=counts_pkl_fname)
nist, blue = score_model(lang_pair, exp_dir, draw=False)
results[exp_count]["NIST"] = nist
results[exp_count]["BLUE"] = blue
exp_count += 1
results = results[:exp_count]
print results
results.dump("results.pkl")
