Question

我有一个4D张量 x ，例如：

(bs1, bs2, sent_len1, sentlen2)  # bs1 and bs2 are unknown and bs1 >= bs2.

和索引张量 ind ，例如：

(bs1, 1)

如果将 ind 转换为numpy数组 np_ind 并执行操作：

cate_ind = to_categorical(np_ind, num_classes=None)

cate_ind的形状为：

(bs1, bs2)

我想使用张量 ind 来减小 x 的尺寸

(bs1, bs2, sent_len1, sentlen2) to (bs1, sent_len1, sentlen2)

代码示例为：

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):
        super(CustomLayer, self).__init__(**kwargs)

    def build(self, input_shape):
        assert len(input_shape) == 3
        super(CustomLayer, self).build(input_shape)

    # TODO:
    def dimension_reduction(self, tensor, indices):
        pass

    def call(self, x, mask=None):
        tensor1, tensor2, ind = x[0], x[1], x[2]  # (bs1, sent_len1, wd) (bs2, sent_len2, wd) (bs1, 1)
        tensor2 = K.permute_dimensions(tensor2, [0, 2, 1])
        align = K.dot(tensor1, tensor2)
        align = K.permute_dimensions(align, [0, 2, 1, 3])  # (bs1, bs2, sent_len1, sent_len2)
        align = self.dimension_reduction(align, ind)  # (bs1, sen_len1, sent_len2)
        return align

    def compute_output_shape(self, input_shape):
        t1_shape, t2_shape = input_shape[0], input_shape[1]
        return t1_shape[0], t1_shape[1], t2_shape[1]


# model example
t1 = Input(shape=(7, 3))
t2 = Input(shape=(6, 3))
t3 = Input(shape=(1,))
output = CustomLayer()([t1, t2, t3])
model = Model([t1, t2, t3], output)

# data example
tensor1 = np.random.rand(10, 7, 3)  # (bs1, sent_len1, wd)
tensor2 = np.random.rand(4, 6, 3)  # (bs2, sent_len2, wd)
indices = np.array([0, 1, 3, 2, 0, 1, 2, 2, 3, 1])  # (bs1, 1)
# indices = to_categorical(indices)  # (bs1, bs2)
print(model.predict([tensor1, tensor2, indices]))  # (bs1, sen_len1, sent_len2)

将特殊的4D张量减少为keras中的3D张量

0 个答案: