2
我發現fit_generator()
會比它應該運行更多的步驟。
我設置了steps_per_epoch=100
。我和k都從0開始。但是在訓練過程結束時,它會打印出k = 109
。只有在添加驗證數據時纔會出現這種情況。Keras fit_generator上的問題,運行幾個步驟比應該多
def data_generate(xfd, yfd, x_line_offset, y_line_offset):
while True:
k = 0
x_line_offset, y_line_offset = shuffle_list(x_line_offset, y_line_offset)
for i in range(100):
print('i = {}'.format(i))
print('k = {}'.format(k))
k += 1
x_train = get_line_by_offset(xfd, x_line_offset[i])
x_train = rescaling(x_train, 0, 65535, 0, 1)
y_train = get_line_by_offset(yfd, y_line_offset[i])
yield x_train, y_train
train_generator = data_generate(xfd_train, yfd_train, x_train_line_offset, y_train_line_offset)
validation_generator = data_generate(xfd_valid, yfd_valid, x_valid_line_offset, y_valid_line_offset)
model.fit_generator(train_generator, steps_per_epoch=100,
validation_data=validation_generator,
validation_steps=len(fix_y_valid_line_offset), epochs=1)
因爲它會打印出k = 109
,我認爲它運行幾個步驟。我不知道它是否有缺陷。但是在k = 99
之後keras日誌消息不顯示。