def test_model_builder_with_class_weights_sgd(self):
# get ambiguity map
ambig_fname = config["test_data_dir"] +"/de-en_ambig.tab"
ambig_map = AmbiguityMap(ambig_fname)
# get samples
samp_hdf_fname = config["test_data_dir"] + "/de-en_samples.hdf5_"
samp_hdfile = h5py.File(samp_hdf_fname, "r")
# create data generator
data_gen = DataSetGenerator(ambig_map, samp_hdfile)
# get lemma counts
# TODO: reading all counts is slow - file with subset of counts needed
counts_fname = config["count"]["lemma"]["en"]["pkl_fname"]
from sklearn.linear_model import SGDClassifier
# build
models_hdf_fname = tempfile.NamedTemporaryFile().name
builder = ModelBuilder(
data_gen,
models_hdf_fname = models_hdf_fname,
classifier = SGDClassifier(),
counts_fname=counts_fname)
builder.run()
def nb_build_model(ns):
"""
build NB models with class weighting
"""
ns.models_fname = ns.fname_prefix + "_models.hdf5"
counts_fname = config["count"]["lemma"][ns.target_lang]["pkl_fname"]
model_builder = ModelBuilder(ns.get_data_generator(ns),
ns.models_fname,
ns.classifier,
counts_fname=counts_fname)
model_builder.run()
# clean up params
delattr(ns, "ambig_map")
def test_model_builder(self):
# get ambiguity map
ambig_fname = config["test_data_dir"] +"/de-en_ambig.tab"
ambig_map = AmbiguityMap(ambig_fname)
# get samples
samp_hdf_fname = config["test_data_dir"] + "/de-en_samples.hdf5_"
samp_hdfile = h5py.File(samp_hdf_fname, "r")
# create data generator
data_gen = DataSetGenerator(ambig_map, samp_hdfile)
# build
models_hdf_fname = tempfile.NamedTemporaryFile().name
builder = ModelBuilder(
data_gen,
models_hdf_fname = models_hdf_fname,
classifier = MultinomialNB() )
builder.run()
def make_classifiers():
for lang_pair in "de-en",:
ambig_fname = "{}/{}_ambig.tab".format(
config["test_data_dir"], lang_pair)
ambig_map = AmbiguityMap(ambig_fname)
samp_fname = "{}/{}_samples.hdf5_".format(
config["test_data_dir"], lang_pair)
samp_hdfile = h5py.File(samp_fname, "r")
data_gen = DataSetGenerator(ambig_map, samp_hdfile)
models_fname = "{}/{}_models.hdf5_".format(
config["test_data_dir"], lang_pair)
builder = ModelBuilder(
data_generator = data_gen,
models_hdf_fname = models_fname,
classifier = MultinomialNB() )
builder.run()
def nb_exp(data_sets=config["eval"]["data_sets"],
lang_pairs=(),
text=False,
draw=False,
diff=False,
trash_models=False):
n_components = 10
descriptor = [ ("data", "S16"),
("lang", "S8"),
("nist", "f"),
("blue", "f"),
("name", "S256") ]
results = np.zeros(9999, dtype=descriptor)
exp_count = 0
script_fname = os.path.splitext(os.path.basename(__file__))[0]
results_fname = "_" + script_fname + "_results.txt"
results_outf = open(results_fname, "w")
for data in data_sets:
for lang in lang_pairs or config["eval"][data].keys():
ambig_fname = config["sample"][lang]["ambig_fname"]
try:
samples_fname = config["sample"][lang]["samples_filt_fname"]
except KeyError:
samples_fname = config["sample"][lang]["samples_fname"]
log.warn("backing off to unfiltered samples from " +
samples_fname)
graphs_fname = config["eval"][data][lang]["graphs_fname"]
name = "{}_{}_{}".format(
script_fname, data, lang)
exp_dir = "_" + name
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
models_fname = exp_dir + "/" + name + ".hdf5"
classifier = Pipeline( [("MCF", MinCountFilter(5)),
("MFF", MaxFreqFilter(0.05)),
("CHI2", SelectFpr(chi2, alpha=0.001 )),
("NMF", NMF(n_components=n_components)),
("MNB", MultinomialNB()),
])
# get ambiguity map
ambig_map = AmbiguityMap(ambig_fname, graphs_fname=graphs_fname)
#ambig_map = AmbiguityMap(ambig_fname, subset={"klar/adj"})
# train classifier
model_builder = ModelBuilder( ambig_map, samples_fname,
models_fname, classifier) #,with_vocab_mask=True)
model_builder.run()
# apply classifier
model = TranslationClassifier(models_fname)
score_attr="nb_score"
source_lang = lang.split("-")[0]
scorer = ClassifierScore(model,
score_attr=score_attr,
filter=filter_functions(source_lang),
vectorizer="mft")
graph_list = cPickle.load(open(graphs_fname))
scorer(graph_list)
best_scorer = BestScorer(["nb_score", "freq_score"])
best_scorer(graph_list)
scored_graphs_fname = exp_dir + "/" + name + "_graphs.pkl"
log.info("saving scored graphs to " + scored_graphs_fname)
cPickle.dump(graph_list, open(scored_graphs_fname, "w"))
#graph_list = cPickle.load(open(scored_graphs_fname))
nist_score, bleu_score = postprocess(
name, data, lang, graph_list,
best_score_attr="best_score",
base_score_attrs=["nb_score","freq_score"],
out_dir=exp_dir,
base_fname=name,
text=text,
draw=draw,
diff=diff
)
results[exp_count] = (data, lang, nist_score, bleu_score, name)
results_fname = exp_dir + "/" + name + ".npy"
log.info("saving result to " + results_fname)
np.save(results_fname, results[exp_count])
exp_count += 1
if trash_models:
log.info("Trashing models file " + models_fname)
os.remove(models_fname)
# add to table of results per data set & language pair
sub_results = results[(results["lang"] == lang) &
(results["data"] == data)]
sub_results = np.sort(sub_results,
axis=0,
order=("lang", "blue"))[::-1]
text_table(sub_results, results_outf)
results_outf.write("\n\n")
results_outf.close()
results = results[:exp_count]
results_fname = "_" + script_fname + "_results.npy"
log.info("saving pickled results to " + results_fname)
np.save(results_fname, results)
text_table(results)
return results
def centroid_exp(data_sets=config["eval"]["data_sets"],
lang_pairs=(),
text=False,
draw=False,
diff=False,
trash_models=False,
dump_centroids=False):
descriptor = [ ("data", "S16"),
("lang", "S8"),
("min_count", "f"),
("max_freq", "f"),
("nist", "f"),
("blue", "f"),
("name", "S256") ]
results = np.zeros(9999, dtype=descriptor)
exp_count = 0
script_fname = os.path.splitext(os.path.basename(__file__))[0]
results_fname = "_" + script_fname + "_results.txt"
results_outf = open(results_fname, "w")
for data in data_sets:
for lang in lang_pairs or config["eval"][data].keys():
ambig_fname = config["sample"][lang]["ambig_fname"]
try:
samples_fname = config["sample"][lang]["samples_filt_fname"]
except KeyError:
samples_fname = config["sample"][lang]["samples_fname"]
log.warn("backing off to unfiltered samples from " +
samples_fname)
graphs_fname = config["eval"][data][lang]["graphs_fname"]
#for min_count in (1, 5, 10, 25, 50, 100, 250, 1000, 2500, 5000):
# for max_freq in (0.0001, 0.001, 0.005, 0.01, 0.05, 0.10, 0.25, 0.5, 1.0):
for min_count in (5,):
for max_freq in (0.01,):
name = "{}_{}_{}_min_count={:d}_max_freq={:f}".format(
script_fname, data, lang, min_count, max_freq)
exp_dir = "_" + name
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
models_fname = exp_dir + "/" + name + ".hdf5"
classifier = Pipeline( [("MCF", MinCountFilter(min_count)),
("MFF", MaxFreqFilter(max_freq)),
("CHI2", SelectFpr()),
#("TFIDF", TfidfTransformer()),
("CNC", CosNearestCentroid())
#("NC", NearestCentroidProb())
])
# train classifier
model_builder = ModelBuilder(
ambig_fname, samples_fname, models_fname, classifier,
graphs_fname, with_vocab_mask=True)
model_builder.run()
# print the centroids to a file, only the 50 best features
if dump_centroids:
print_fname = exp_dir + "/" + name + "_centroids.txt"
print_centroids(models_fname,
n=50,
outf=print_fname)
# apply classifier
model = TranslationClassifier(models_fname)
score_attr="centroid_score"
source_lang = lang.split("-")[0]
scorer = ClassifierScore(model,
score_attr=score_attr,
filter=filter_functions(source_lang))
graph_list = cPickle.load(open(graphs_fname))
scorer(graph_list)
best_scorer = BestScorer(["centroid_score", "freq_score"])
best_scorer(graph_list)
scored_graphs_fname = exp_dir + "/" + name + "_graphs.pkl"
log.info("saving scored graphs to " + scored_graphs_fname)
cPickle.dump(graph_list, open(scored_graphs_fname, "w"))
#graph_list = cPickle.load(open(scored_graphs_fname))
nist_score, bleu_score = postprocess(
name, data, lang, graph_list,
best_score_attr="best_score",
base_score_attrs=["centroid_score","freq_score"],
out_dir=exp_dir,
base_fname=name,
text=text,
draw=draw,
diff=diff
)
results[exp_count] = (data, lang, min_count, max_freq,
nist_score, bleu_score, name)
results_fname = exp_dir + "/" + name + ".npy"
log.info("saving result to " + results_fname)
np.save(results_fname, results[exp_count])
exp_count += 1
if trash_models:
log.info("Trashing models file " + models_fname)
os.remove(models_fname)
sub_results = results[(results["lang"] == lang) &
(results["data"] == data)]
sub_results = np.sort(sub_results,
axis=0,
order=("lang", "blue"))[::-1]
text_table(sub_results, results_outf)
results_outf.write("\n\n")
results_outf.close()
results = results[:exp_count]
results_fname = "_" + script_fname + "_results.npy"
log.info("saving pickled results to " + results_fname)
np.save(results_fname, results)
text_table(results)
return results