def prepare_data(augment_iter=0):
X = []
y = []
for i in range(n_classes):
if i == n_classes - 1:
char = 'None'
else:
char = vocabulary[i]
res_x = pickle.load(open(root_path + char + ".pkl", 'rb'))
res_y = np.tile(i, (len(res_x), 1)).tolist()
X += res_x
y += res_y
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.15,
random_state=42,
stratify=y)
X_train, y_train = augment_data(X_train, y_train, iterations=augment_iter)
# add features and normalize data
pen_up = []
for i in range(len(X_train)):
sequence = np.asarray(X_train[i])
pen_up.append(sequence[:, 2])
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
X_train[i] = sequence
data_scaler.fit(np.vstack(X_train))
for i in range(len(X_train)):
sequence = np.asarray(X_train[i])
sequence = data_scaler.transform(sequence)
X_train[i] = np.column_stack((sequence, pen_up[i]))
for i in range(len(X_test)):
sequence = np.asarray(X_test[i])
pen_up = sequence[:, 2]
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
sequence = data_scaler.transform(sequence)
X_test[i] = np.column_stack((sequence, pen_up))
return X_train, X_test, y_train, y_test
def predict(sess, model):
# prediction for sample
observer = DataObserver("demo.log")
print("Real-time prediction started.")
while True:
new_entry = observer.step()
if new_entry != None:
sequence = np.asarray(new_entry)
pen_up = sequence[:, 2]
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
if CONFIG.use_normalization:
sequence = data_scaler.transform(sequence)
sequence = np.column_stack((sequence, pen_up)).tolist()
#get my prediction
output = model.predict(sess, [sequence])
prediction = np.argmax(output[0], axis=0)
if prediction < len(vocabulary):
print("Input detected: " + str(vocabulary[prediction]) + ", Probabilities: " + str(output[0]))
time.sleep(0.05)
def prepare_data(pad_length=False):
X = []
y = []
for i in range(CONFIG.n_classes):
if i == CONFIG.n_classes-1:
char = 'None'
else:
char = vocabulary[i]
res_x = pickle.load(open(CONFIG.root_path + char + ".pkl", 'rb'))
res_y = np.tile(np.eye(CONFIG.n_classes)[i], (len(res_x), 1)).tolist()
X += res_x
y += res_y
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42, stratify=y)
if CONFIG.use_augmentation:
X_train, y_train = augment_data(X_train, y_train, iterations=CONFIG.augment_iter)
# add features and normalize data to 0 mean and unit variance
pen_up = []
for i in range(len(X_train)):
sequence = np.asarray(X_train[i])
pen_up.append(sequence[:, 2])
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
if CONFIG.use_normalization:
X_train[i] = sequence
else:
X_train[i] = np.column_stack((sequence, pen_up[i])).tolist()
if CONFIG.use_normalization:
global data_scaler
data_scaler.fit(np.vstack(X_train))
for i in range(len(X_train)):
sequence = np.asarray(X_train[i])
sequence = data_scaler.transform(sequence)
X_train[i] = np.column_stack((sequence, pen_up[i])).tolist()
for i in range(len(X_test)):
sequence = np.asarray(X_test[i])
pen_up = sequence[:, 2]
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
if CONFIG.use_normalization:
sequence = data_scaler.transform(sequence)
X_test[i] = np.column_stack((sequence, pen_up)).tolist()
# # dimensionality reduction with PCA
# pca = PCA(n_components=8)
# pca.fit(np.vstack(X_train))
#
# for i in range(len(X_train)):
# X_train[i] = pca.transform(X_train[i]).tolist()
#
# for i in range(len(X_test)):
# X_test[i] = pca.transform(X_test[i]).tolist()
# plot pca result
# fig, ax1 = plt.subplots()
# ax1.plot()
# ax1.set_xlabel('components')
# ax1.set_ylabel('variance percentage')
# ax1.plot(range(1, len(pca.explained_variance_ratio_) + 1), pca.explained_variance_ratio_, color='tab:blue')
# plt.show()
if pad_length:
max_seqLen = max(len(max(X_train, key=len)), len(max(X_test, key=len)))
# Pad sequences for dimension consistency
padding_mask = np.zeros(CONFIG.n_features).tolist()
for i in range(len(X_train)):
X_train[i] += [padding_mask for _ in range(max_seqLen - len(X_train[i]))]
for i in range(len(X_test)):
X_test[i] += [padding_mask for _ in range(max_seqLen - len(X_test[i]))]
return X_train, X_test, y_train, y_test
if len(char_idx) > 1:
char_idx = list(range(char_idx[0], char_idx[-1] + 1))
label = index_and_label[1].replace(' ', '').replace(
'"', '').replace('\n', '')
sequence = []
for idx in char_idx:
sequence.extend(strokes[idx])
sequence.extend([[0., 0., 1.]])
# add features
sequence = np.asarray(sequence)
pen_up.append(sequence[:, 2])
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
onehot_label = np.zeros(len(vocabulary) + 1)
if label.upper() in vocabulary:
X.append(sequence)
onehot_label[vocabulary.index(label.upper())] = 1.
y.append(onehot_label)
else:
X.append(sequence)
onehot_label[-1] = 1.
y.append(onehot_label)
# normalize data and add pen_up column
unipen_data_scaler = StandardScaler()
unipen_data_scaler.fit(np.vstack(X))
for i in range(len(X)):
def prepare_data(augment_iter=0):
X = []
y = []
for i in range(n_classes):
if i == n_classes - 1:
char = 'None'
else:
char = vocabulary[i]
res_x = pickle.load(open(root_path + char + ".pkl", 'rb'))
res_y = np.tile(i, (len(res_x), 1)).tolist()
X += res_x
y += res_y
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.15,
random_state=42,
stratify=y)
X_train, y_train = augment_data(X_train, y_train, iterations=augment_iter)
# add features and normalize data to 0 mean and unit variance
pen_up = []
for i in range(len(X_train)):
sequence = np.asarray(X_train[i])
pen_up.append(sequence[:, 2])
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
X_train[i] = sequence
data_scaler.fit(np.vstack(X_train))
for i in range(len(X_train)):
sequence = np.asarray(X_train[i])
sequence = data_scaler.transform(sequence)
X_train[i] = np.column_stack((sequence, pen_up[i])).tolist()
for i in range(len(X_test)):
sequence = np.asarray(X_test[i])
pen_up = sequence[:, 2]
sequence = sequence[:, 0:2]
sequence = add_features(sequence)
sequence = data_scaler.transform(sequence)
X_test[i] = np.column_stack((sequence, pen_up)).tolist()
max_seqLen = max(len(max(X_train, key=len)), len(max(X_test, key=len)))
# Pad sequences for dimension consistency
padding_mask = np.zeros(n_features).tolist()
for i in range(len(X_train)):
X_train[i] += [
padding_mask for _ in range(max_seqLen - len(X_train[i]))
]
for i in range(len(X_test)):
X_test[i] += [padding_mask for _ in range(max_seqLen - len(X_test[i]))]
# flat sequence
X_train = np.asarray(X_train)
shape = np.shape(X_train)
X_train = np.reshape(X_train, (shape[0], shape[1] * shape[2]))
X_test = np.asarray(X_test)
shape = np.shape(X_test)
X_test = np.reshape(X_test, (shape[0], shape[1] * shape[2]))
return X_train, X_test, y_train, y_test