【发布时间】：2018-09-14 08:58:43
【问题描述】：
我正在尝试在 keras 中自定义一个特殊的注意力层。但是我很困惑为什么在尝试了很多方法后总是会出现这个错误。
Traceback (most recent call last):
File "D:/Users/LawLi/PyCharmProjects/fixed_talentDNA/adx.py", line 52, in <module>
print(model.predict([tensor1, tensor2, indices])) # (bs1, sl1, sl2)
File "D:\Users\LawLi\Anaconda3\lib\site-packages\keras\engine\training.py", line 1172, in predict
steps=steps)
File "D:\Users\LawLi\Anaconda3\lib\site-packages\keras\engine\training_arrays.py", line 293, in predict_loop
ins_batch = slice_arrays(ins, batch_ids)
File "D:\Users\LawLi\Anaconda3\lib\site-packages\keras\utils\generic_utils.py", line 507, in slice_arrays
return [None if x is None else x[start] for x in arrays]
File "D:\Users\LawLi\Anaconda3\lib\site-packages\keras\utils\generic_utils.py", line 507, in <listcomp>
return [None if x is None else x[start] for x in arrays]
IndexError: index 4 is out of bounds for axis 0 with size 4
这是我关于自定义层的测试代码，代码中包含一个预测示例。可以直接运行。
from keras.layers import *
from keras.models import Model
from keras.utils import to_categorical
from keras.layers.merge import *
class CustomLayer(Layer):
def __init__(self, **kwargs):
self.supports_masking = True
super(CustomLayer, self).__init__(**kwargs)
def build(self, input_shape):
assert len(input_shape) == 3
super(CustomLayer, self).build(input_shape)
def compute_mask(self, inputs, mask=None):
return None
def call(self, x, mask=None):
tensor1, tensor2, ind = x[0], x[1], x[2] # (bs1, sl1, wd) (bs2, sl2, wd) (bs1, bs2. sl1, sl2)
tensor2 = K.permute_dimensions(tensor2, [0, 2, 1])
align = K.dot(tensor1, tensor2)
align = K.permute_dimensions(align, [0, 2, 1, 3]) # (bs1, bs2, sl1, sl2)
align = align + ind
align = K.max(align, axis=1)
align = K.sum(align, axis=2)
align = K.softmax(align, axis=1)
weighted_ans = tensor1 * K.expand_dims(align, 2)
return K.sum(weighted_ans, axis=1)
def compute_output_shape(self, input_shape):
t1_shape, t2_shape = input_shape[0], input_shape[1]
return t1_shape[0], t1_shape[1], t1_shape[2]
# model example
t1 = Input(shape=(7, 3))
t2 = Input(batch_shape=(4, 6, 3))
t3 = Input(shape=(4, 7, 6))
output = CustomLayer()([t1, t2, t3])
model = Model([t1, t2, t3], output)
# data example
tensor1 = np.random.rand(10, 7, 3) # (bs1, sl1, wd)
tensor2 = np.random.rand(4, 6, 3) # (bs2, sl2, wd)
indices = np.array([0, 1, 3, 2, 0, 1, 2, 2, 3, 1]) # (bs1, 1)
indices = to_categorical(indices, num_classes=4) * 999 - 999 # (bs1, bs2)
indices = np.expand_dims(indices, axis=2)
indices = np.expand_dims(indices, axis=3)
indices = np.repeat(indices, 7, axis=2).repeat(6, axis=3)
print(model.predict([tensor1, tensor2, indices])) # (bs1, sl1, wd)
谢谢你的帮助。
【问题讨论】：
-
你能打印
indices = to_categorical(indices, num_classes=4)的输出吗？它可能编码为 [1,2,3,4] 而不是您所期望的 [0,1,2,3]。 -
我只是打印你所说的那一行。输出是一个形状为 (10, 4) 的热矩阵，如下所示：
[0. 1. 0. 0.][0. 0. 0. 1.]...[0. 0. 1. 0.][1. 0. 0. 0.] -
嗯，如果你在
weighted_ans = tensor1 * K.expand_dims(align, 2)之后立即添加print(weighted_ans.shape)，它会在第一次调用指令时输出(?, 7, 3)，到时候应该还没有确定吗？
标签： python keras deep-learning artificial-intelligence