我正在尝试执行混合增强（详细信息here），但出现如下所示的值错误：
ValueError: operands could not be broadcast together with shapes (47,128,128,3) (64,1,1,1)
以下是我正在使用的MixupImageDataGenerator类定义：
class MixupImageDataGenerator():
def __init__(self, generator, directory, batch_size, img_height, img_width, alpha=0.2, subset=None):
"""Constructor for mixup image data generator.
Keyword Arguments:
alpha {float} -- Mixup beta distribution alpha parameter. (default: {0.2})
subset {str} -- 'training' or 'validation' if validation_split is specified in
`generator` (ImageDataGenerator).(default: {None})
"""
self.batch_index = 0
self.batch_size = 64
self.alpha = alpha
# First iterator yielding tuples of (x, y)
self.generator1 = generator.flow_from_directory(directory,
target_size=(
128, 128),
class_mode="categorical",
batch_size=64,
shuffle=True,
subset=subset)
# Second iterator yielding tuples of (x, y)
self.generator2 = generator.flow_from_directory(directory,
target_size=(
128, 128),
class_mode="categorical",
batch_size=64,
shuffle=True,
subset=subset)
# Number of images across all classes in image directory.
self.n = self.generator1.samples
def reset_index(self):
"""Reset the generator indexes array.
"""
self.generator1._set_index_array()
self.generator2._set_index_array()
def on_epoch_end(self):
self.reset_index()
def reset(self):
self.batch_index = 0
def __len__(self):
# round up
return (self.n + self.batch_size - 1) // self.batch_size
def get_steps_per_epoch(self):
"""Get number of steps per epoch based on batch size and
number of images.
Returns:
int -- steps per epoch.
"""
return self.n // self.batch_size
def __next__(self):
"""Get next batch input/output pair.
Returns:
tuple -- batch of input/output pair, (inputs, outputs).
"""
if self.batch_index == 0:
self.reset_index()
#my_list = []
current_index = (self.batch_index * self.batch_size) % self.n
#my_list.insert(0,current_index)
#print("current index" + str(current_index))
if self.n > current_index + self.batch_size:
self.batch_index += 1
else:
self.batch_index = 0
#print("batch_index" + str(self.batch_index))
# random sample the lambda value from beta distribution.
l = np.random.beta(self.alpha, self.alpha, self.batch_size)
print("l")
print(l.shape)
X_l = l.reshape(self.batch_size, 1, 1, 1)
print("X_l")
print(X_l.shape)
y_l = l.reshape(self.batch_size, 1)
print("y_l")
print(y_l.shape)
# Get a pair of inputs and outputs from two iterators.
X1, y1 = self.generator1.next()
print("X1, y1")
print(X1.shape, y1.shape)
X2, y2 = self.generator2.next()
print("X2, y2")
print(X2.shape, y2.shape)
# Perform the mixup.
X = X1 * X_l + X2 * (1 - X_l)
y = y1 * y_l + y2 * (1 - y_l)
print("X")
print(X.shape)
print("y")
print(y.shape)
print("my_list " + str(my_list))
return X, y
def __iter__(self):
while True:
yield next(self)
以下是我的代码，在这里我调用Mixup类并执行实际的增强操作：
# Create training and validation generator.
train_generator = MixupImageDataGenerator(generator=input_datagen,
directory='data/train/',
batch_size=64,
img_height=128,
img_width=128,
subset='training')
validation_generator = input_datagen.flow_from_directory(directory = 'data/train/', target_size=(128, 128), color_mode='rgb', class_mode='categorical',batch_size=64,subset='validation', shuffle=True, seed=42)
print('training steps: ', train_generator.get_steps_per_epoch())
print('validation steps: ', validation_generator.samples // batch_size)
输出：
Found 3503 images belonging to 5 classes.
Found 3503 images belonging to 5 classes.
Found 1496 images belonging to 5 classes.
training steps: 54
validation steps: 23
我开始了培训，并展示了第42步的输出。先前的所有步骤都具有相似的输出。
42/54 [======================>.......] - ETA: 11s - loss: 1.3245 - acc: 0.4565
current index 3392
batch_index 54
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(64, 128, 128, 3) (64, 5)
\n X2, y2
(64, 128, 128, 3) (64, 5)
\n X
(64, 128, 128, 3)
y
(64, 5)
43/54 [======================>.......] - ETA: 10s - loss: 1.3227 - acc: 0.4557
current index 3456
batch_index 0
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(47, 128, 128, 3) (47, 5)
\n X2, y2
(47, 128, 128, 3) (47, 5)
44/54 [=======================>......] - ETA: 9s - loss: 1.3233 - acc: 0.4538 current index 0
batch_index 1
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(64, 128, 128, 3) (64, 5)
\n X2, y2
(64, 128, 128, 3) (64, 5)
\n X
(64, 128, 128, 3)
y
(64, 5)
45/54 [========================>.....] - ETA: 8s - loss: 1.3214 - acc: 0.4549
current index 64
batch_index 2
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(64, 128, 128, 3) (64, 5)
\n X2, y2
(64, 128, 128, 3) (64, 5)
\n X
(64, 128, 128, 3)
y
(64, 5)
46/54 [========================>.....] - ETA: 7s - loss: 1.3211 - acc: 0.4541
current index 128
batch_index 3
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(64, 128, 128, 3) (64, 5)
\n X2, y2
(64, 128, 128, 3) (64, 5)
\n X
(64, 128, 128, 3)
y
(64, 5)
47/54 [=========================>....] - ETA: 6s - loss: 1.3193 - acc: 0.4561
current index 192
batch_index 4
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(64, 128, 128, 3) (64, 5)
\n X2, y2
(64, 128, 128, 3) (64, 5)
\n X
(64, 128, 128, 3)
y
(64, 5)
.
.
.
.
.
53/54 [============================>.] - ETA: 0s - loss: 1.3072 - acc: 0.4608
current index 576
batch_index 10
\n l
(64,)
X_l
(64, 1, 1, 1)
y_l
(64, 1)
\n X1, y1
(64, 128, 128, 3) (64, 5)
\n X2, y2
(64, 128, 128, 3) (64, 5)
\n X
(64, 128, 128, 3)
y
(64, 5)
54/54 [==============================] - 53s 986ms/step - loss: 1.3058 - acc: 0.4606 - val_loss: 1.2556 - val_acc: 0.4749
Epoch 00001: val_acc improved from -inf to 0.47486, saving model to GoDL-NP_CE_NLS_ABC-MIXUP.h5
Epoch 2/3
此后，出现以下错误：
ValueError: operands could not be broadcast together with shapes (47,128,128,3) (64,1,1,1)
我的观察是
1）除步骤43之外，所有其他步骤均具有以下形状：
X1，y1 （64，128，128，3）（64，5）
X2，y2 （64，128，128，3）（64，5）
但是步骤43具有以下形状：
X1，y1 （47，128，128，3）（47，5）
X2，y2 （47，128，128，3）（47，5）
2）在训练的第一步开始之前，数据生成器似乎已经向前移动了，这可以从以下事实看出：第一步运行时cuurent_index = 124。
3）我觉得可用于步骤43的图像仅为47个，而不是预期的64个（等于批处理大小）。我之所以这样说，是因为对于步骤47，current_index = 3456，batch_index = 0。训练图像的总大小为3503。因此，相差为3503-3456 =47。下一步，当前索引= 0，batch_index = 1。
所以，我想我需要的是以下东西：
答案 0 :(得分：0)
基于此问题：https://github.com/Tony607/keras_mixup_generator/issues/1
您需要更改混合生成器，以便使用X1.shape[0]为最后一批正确计算lambda的大小：
# Get a pair of inputs and outputs from two iterators.
X1, y1 = self.generator1.next()
X2, y2 = self.generator2.next()
# random sample the lambda value from beta distribution.
l = np.random.beta(self.alpha, self.alpha, X1.shape[0])
X_l = l.reshape(X1.shape[0], 1, 1, 1)
y_l = l.reshape(X1.shape[0], 1)
# Perform the mixup.
X = X1 * X_l + X2 * (1 - X_l)
y = y1 * y_l + y2 * (1 - y_l)
return X, y