def _data(data_pth, split_val=True, verbose=1):
data = np.load(data_pth, allow_pickle=True)
x, y = data['x'], data['y']
x = x[:, :, np.newaxis]
x_train, y_train, x_test, y_test = split_data(x, y)
class_weights_dict = calc_class_weights(y_train)
if split_val:
x_train, y_train, x_val, y_val = split_data(x_train, y_train)
y_train = to_one_hot(y_train, dimension=10)
y_test = to_one_hot(y_test, dimension=10)
y_val = to_one_hot(y_val, dimension=10)
if verbose:
print('\nx_train shape: %s'
'\ny_train shape: %s'
'\nx_test shape: %s'
'\ny_test shape: %s'
'\nx_val shape: %s'
'\ny_val shape: %s' %
(x_train.shape, y_train.shape, x_test.shape, y_test.shape,
x_val.shape, y_val.shape))
return x_train, y_train, x_test, y_test, x_val, y_val, class_weights_dict
else:
y_train = to_one_hot(y_train, dimension=10)
y_test = to_one_hot(y_test, dimension=10)
if verbose:
print('\nx_train shape: %s'
'\ny_train shape: %s'
'\nx_test shape: %s'
'\ny_test shape: %s' %
(x_train.shape, y_train.shape, x_test.shape, y_test.shape))
return x_train, y_train, x_test, y_test, class_weights_dict
def train_model(model_svm_path,
tr_features,
tr_labels,
verbose=False,
**model_kwargs):
c_value = model_kwargs.get("c_value", 1)
gamma_value = model_kwargs.get("gamma_value", 0)
multi_mode = model_kwargs.get("multi_model", "ovo")
use_probability = False
class_weights = calc_class_weights(tr_labels)
# class_weights = "balanced",
gamma_value = gamma_value if gamma_value != 0.0 else "auto"
# TODO load best params from cross validation!
# NOTE 0.0 means 1/n_features default value
svm_model = svm.SVC(C=c_value,
kernel='rbf',
degree=3,
gamma=gamma_value,
coef0=0.0,
shrinking=True,
probability=use_probability,
tol=0.001,
cache_size=200,
class_weight=class_weights,
verbose=verbose,
max_iter=-1,
decision_function_shape=multi_mode,
random_state=None)
with PrintTime("train the model", verbose=verbose):
svm_model.fit(tr_features, tr_labels)
if model_svm_path:
# Export model: save/write trained SVM model
if not os.path.exists(os.path.dirname(model_svm_path)):
os.makedirs(os.path.dirname(model_svm_path))
joblib.dump(svm_model, model_svm_path)
# TODO Export StandardScaler()
return svm_model
A.shape).toarray(), csr_matrix(A.shape).toarray()
orc = OllivierRicci(G, alpha=0.5, verbose="INFO")
orc.compute_ricci_curvature()
frc = FormanRicci(G)
frc.compute_ricci_curvature()
for tup in orc.G.edges:
i, j = tup[0], tup[1]
ollivier_curv_vals[i][j] = map_curvature_val(
orc.G[i][j]['ricciCurvature'], alpha=4)
forman_curv_vals[i][j] = map_curvature_val(
frc.G[i][j]['formanCurvature'], alpha=4)
K = A.shape[1] if X is None else X.shape[0]
nC = Y.shape[1]
W = None
if args.weighted:
W = utils.calc_class_weights(Y[..., idx_train, :])
# ************************************************************
# calculate node features
# ************************************************************
vals = []
X = X.astype(np.float32)
EYE = scipy.sparse.eye(K, dtype=np.float32, format='coo')
A = A.astype(np.float32)
# normalized x
rowsum = sparse.as_coo(X).sum(axis=-1, keepdims=True)
#rowsum.data[rowsum.data==0] = 1e-10
rowsum.data = 1.0 / rowsum.data
vals.append((sparse.as_coo(X) * rowsum).to_scipy_sparse())
nodes = scipy.sparse.hstack(vals)
vocab[w] for w in re.findall('\w+', x_.lower()) if w in vocab
] for x_ in dataset.x_test]
dev_x_int = [[
vocab[w] for w in re.findall('\w+', x_.lower()) if w in vocab
] for x_ in dataset.x_dev]
max_sent_length = max([len(x_) for x_ in train_x_int])
train_x = pad_sequences(train_x_int, max_sent_length)
test_x = pad_sequences(test_x_int, max_sent_length)
dev_x = pad_sequences(dev_x_int, max_sent_length)
if do_train:
model = build_model(max_sent_length, vocab, embeddings, num_classes)
print(model.summary())
class_weights = calc_class_weights(train_y, mlb.classes_)
class_weights = {i: w for i, w in enumerate(class_weights)}
early_stopping = EarlyStopping(monitor='val_loss',
patience=3,
restore_best_weights=True)
try:
model.fit(train_x,
train_y,
batch_size=32,
epochs=50,
verbose=1,
validation_data=(dev_x, dev_y),
class_weight=class_weights,
callbacks=[early_stopping])