def setUp(self):
data_path = (os.path.join(os.path.dirname(__file__),
"test_participant.json"))
TFIDF = False
N_GRAMS = (1, 2)
self.data = TrainData.load(data_path)
train_ids = [t.ids for t in self.data.data["1.ABC"]]
flat_train_ids = list(self.data.data["1.ABC"].ids)
self.train_cntxt, self.train_acts = self._format_cntxt_indices(
train_ids)
self.X_speech, vectorizer = get_bow_features(self.data,
tfidf=TFIDF,
n_grams=N_GRAMS,
max_features=None)
self.X_speech = self.X_speech[flat_train_ids, :]
model_gen = SpeechModel.model_generator(SGDClassifier,
loss='log',
average=True,
penalty='l2',
alpha=0.0002)
self.combined_model = CombinedModel(vectorizer,
model_gen,
ALL_ACTIONS,
speech_eps=0.15,
context_eps=0.15)
self.test_utter = "The green piece with two black stripes"
self.test_cntxt = []
def plot_trial(trial, bag):
model_path = os.path.join(args.model_path, args.participant, str(trial),
"model_initial")
fig_path = os.path.join(
os.path.dirname(__file__), "figs", args.participant, str(trial))
if not os.path.exists(fig_path):
os.makedirs(fig_path)
model = CombinedModel.load_from_path(model_path, ALL_ACTIONS,
JointModel.model_generator(
SGDClassifier,
**SPEECH_MODEL_PARAMETERS),
SPEECH_EPS, CONTEXT_EPS)
row = 4
col = 6
cntxt = []
i = 0
for m in bag.read_messages():
if m.topic == TOPIC:
model.predict(cntxt, m.message.utter, plot=True)
cntxt.append(m.message.action)
i += 1
plt.tight_layout()
path = os.path.join(fig_path, "sample_{}_{}".format(
m.message.result, i))
plt.savefig(path, format="svg")
plt.clf()
def _get_model(self, participant, trial):
if trial == 0:
model_type = "model_initial"
t = 0
else:
model_type = "model_final"
t = trial - 1
model_path = os.path.join(args.model_path, participant, str(t + 1),
model_type)
return CombinedModel.load_from_path(
model_path, ALL_ACTIONS,
JointModel.model_generator(SGDClassifier,
**SPEECH_MODEL_PARAMETERS), SPEECH_EPS,
CONTEXT_EPS)
def train_combined_model(speech_eps,
context_eps,
fit_type="incremental",
tfidf=False,
n_grams=(1, 2),
speech_model_class=JointModel,
speech_model_parameters={},
init_new_speech_actions=False):
path = defaults.DATA_PATH
print("PATH: ", os.path.join(path, "train.json"))
data = TrainData.load(os.path.join(path, "train.json"))
flat_train_ids = [i for p in TRAIN_PARTICIPANTS for i in data.data[p].ids]
train_ids_by_trial = [
trial.ids for p in TRAIN_PARTICIPANTS for trial in data.data[p]
]
# Get features
train_context, labels = format_cntxt_indices(data, train_ids_by_trial)
X_speech, vectorizer = get_bow_features(data,
tfidf=tfidf,
n_grams=n_grams,
max_features=None)
X_speech = X_speech[flat_train_ids, :]
model_gen = JointModel.model_generator(speech_model_class,
**speech_model_parameters)
combined_model = CombinedModel(vectorizer,
model_gen,
ALL_ACTIONS,
speech_eps=speech_eps,
context_eps=context_eps)
if "incremental" in fit_type:
combined_model.partial_fit(train_context, X_speech, labels)
elif "offline" in fit_type:
combined_model.fit(train_context, X_speech, labels)
if init_new_speech_actions:
if "incremental" not in fit_type:
raise NotImplementedError(
"Can't add speech data on offline speech")
update_speech_for_new_actions(combined_model.speech_model,
combined_model._vectorizer,
weight=len(labels) * 1. /
len(ALL_ACTIONS))
return combined_model
class TestCombinedModel(TestCase):
def setUp(self):
data_path = (os.path.join(os.path.dirname(__file__),
"test_participant.json"))
TFIDF = False
N_GRAMS = (1, 2)
self.data = TrainData.load(data_path)
train_ids = [t.ids for t in self.data.data["1.ABC"]]
flat_train_ids = list(self.data.data["1.ABC"].ids)
self.train_cntxt, self.train_acts = self._format_cntxt_indices(
train_ids)
self.X_speech, vectorizer = get_bow_features(self.data,
tfidf=TFIDF,
n_grams=N_GRAMS,
max_features=None)
self.X_speech = self.X_speech[flat_train_ids, :]
model_gen = SpeechModel.model_generator(SGDClassifier,
loss='log',
average=True,
penalty='l2',
alpha=0.0002)
self.combined_model = CombinedModel(vectorizer,
model_gen,
ALL_ACTIONS,
speech_eps=0.15,
context_eps=0.15)
self.test_utter = "The green piece with two black stripes"
self.test_cntxt = []
def test_n_children(self):
self.combined_model.fit(self.train_cntxt, self.X_speech,
self.train_acts)
self.assertEqual(self.combined_model.context_model.root.n_children, 2)
def test_predict(self):
self.combined_model.fit(self.train_cntxt, self.X_speech,
self.train_acts)
act, probs = self.combined_model.predict(self.test_cntxt,
self.test_utter)
self.assert_("foot" in act)
def test_prob_normilization(self):
self.combined_model.fit(self.train_cntxt, self.X_speech,
self.train_acts)
c_probs = self.combined_model.get_context_probs(self.test_cntxt)
s_probs = self.combined_model.get_speech_probs(self.test_utter)
self.assertAlmostEqual(sum(c_probs), 1.0, places=4)
self.assertAlmostEqual(sum(s_probs), 1.0, places=4)
def test_incremental_learning(self):
train_ctxt_1 = [[]]
train_ctxt_2 = [[]]
train_utter_1 = "blue piece with two white stripes"
train_utter_2 = "I am feeling fat and sassy"
act_1 = ["top_2"]
act_2 = ["top_1"]
self.combined_model.partial_fit(train_ctxt_1, train_utter_1, act_1)
self.combined_model.partial_fit(train_ctxt_2, train_utter_2, act_2)
self.assertEqual(self.combined_model.context_model.root.n_children, 2)
def _format_cntxt_indices(self, indices):
cntxts = []
actions = []
all_labels = list(self.data.labels)
for t in indices:
cntxt = []
for i in t:
label = all_labels[i]
cntxts.append([c for c in cntxt])
actions.append(label)
cntxt.append(label)
return cntxts, actions
def _load_model(self, model_path, speech_eps, context_eps):
self.model = CM.load_from_path(model_path, ALL_ACTIONS,
JointModel.model_generator(
SGDClassifier,
**SPEECH_MODEL_PARAMETERS),
speech_eps, context_eps)