def eval_butterfly(n_samples, trained_batterfly=None):
if trained_batterfly is None:
# выбираем модель-классификатор
filepath_model = easygui.fileopenbox(
"выберите файл с обученной моделью классиикатором.h5")
trained_batterfly = load_model(filepath_model)
trained_batterfly.summary()
# вытаскиваем полный непорезанный датасет
_, x_test, _, y_test = prepare_data(seg_len=None)
# убеждаемся что датасет подходит к модели
output_len = trained_batterfly.output_shape[1]
num_labels = y_test.shape[1]
assert output_len == num_labels
# включаем его порезку на сегменты
test_generator = ecg_batches_generator_for_classifier(
segment_len=ecg_segment_len,
batch_size=n_samples,
ecg_dataset=x_test,
diagnodses=y_test)
xy = next(test_generator)
# заставляем модель предсказать
prediction = trained_batterfly.predict_on_batch(x=xy[0])
print("ответ модели:")
print(prediction)
print("правильный ответ:")
print(xy[1])
return xy[1], prediction
def TEST_generator_for_ae():
segment_len = 3
batch_size = 1
x_train, _, _, _ = prepare_data(seg_len=None)
train_generator = ecg_batches_generator(segment_len=segment_len,
batch_size=batch_size,
ecg_dataset=x_train)
batch = next(train_generator)
print("батч имеет форму: " + str(batch[0].shape))
def train_batterfly(name):
x_train, x_test, y_train, y_test = prepare_data(
seg_len=None) # вытаскиваем полный непорезанный датасет
num_labels = y_train.shape[1]
model = create_batterfly(num_labels=num_labels)
model.summary()
batch_size = 30
train_generator = ecg_batches_generator_for_classifier(
segment_len=ecg_segment_len,
batch_size=batch_size,
ecg_dataset=x_train,
diagnodses=y_train)
test_generator = ecg_batches_generator_for_classifier(
segment_len=ecg_segment_len,
batch_size=300,
ecg_dataset=x_test,
diagnodses=y_test)
steps_per_epoch = 40
print("батчей за эпоху будет:" + str(steps_per_epoch))
print("в одном батче " + str(batch_size) + " кардиограмм.")
#уменьшение learning rate автоматически на плато
#learning_rate_reduction = ReduceLROnPlateau(monitor='val_loss', factor = 0.1, patience = 5, verbose = 1)
#изменение LR по методу SGDR
#change_lr = cosine_lr.SGDRScheduler(min_lr=0.0001, max_lr=0.1, steps_per_epoch=np.ceil(15/batch_size), lr_decay=0.8, cycle_length=1, mult_factor=1)
#tb_callback = TensorBoard(log_dir='./butterfly_logs', histogram_freq=20, write_graph=True, write_grads=True)
history = model.fit_generator(generator=train_generator,
steps_per_epoch=steps_per_epoch,
epochs=50,
validation_data=test_generator,
validation_steps=1)
save_history(history, name)
model.save(name + '.h5')
eval_generator = ecg_batches_generator_for_classifier(
segment_len=ecg_segment_len,
batch_size=700,
ecg_dataset=x_test,
diagnodses=y_test)
xy = next(eval_generator)
# заставляем модель предсказать
prediction = model.predict_on_batch(x=xy[0])
print("ответ модели:")
print(prediction)
print("правильный ответ:")
print(xy[1])
return xy[1], prediction
def TEST_generator_for_classifier():
segment_len = 3
batch_size = 10
x_train, _, y_train, _ = prepare_data(seg_len=None)
train_generator = ecg_batches_generator_for_classifier(
segment_len=segment_len,
batch_size=batch_size,
ecg_dataset=x_train,
diagnodses=y_train)
batch_xy = next(train_generator)
print("батч имеет форму: \\n x.shape=" + str(batch_xy[0].shape))
print("y.shape=" + str(batch_xy[1].shape))
pprint.pprint(batch_xy[1])
from keras import backend as K
from keras.callbacks import (TensorBoard)
from keras.layers import (LSTM, Bidirectional, Dense, Conv1D, MaxPooling1D,
UpSampling1D)
from keras.models import (Sequential, load_model)
from dataset_getter import prepare_data
num_decoder_tokens = 4
Size = 2992
num_cores = 32
kernel = 8
num_latent = 50
leads = 12
x_train, x_test, y_train, y_test = prepare_data(seg_len=None)
# отрезаем куски без разметки
x_train = x_train[:, 1000:4000, :, :]
x_test = x_test[:, 1000:4000, :, :]
y_train = y_train[:, 1000:4000]
y_test = y_test[:, 1000:4000]
def one_hot(y):
"""
переводит небинарную маску в one hot
:param y:
:return:
"""
y_new = []
for i in range(y.shape[0]):
def prepare_data_for_canterpillar(segment_len=None):
x_train, x_test, _, _ = prepare_data(segment_len)
return x_train, x_test
