我正在尝试在tensorflow 1.14.0中实现Sliding DFT算法,并且正在使用tf.function,这样我就不必担心控制流了,但是我遇到了一个问题。当我尝试将变量的一个元素与该变量中的另一个元素分配在一起时,在跨步切片分配中的不兼容类型方面出现错误。
我尝试使用tf.scatter更新,tf分配,并且仅使用典型的切片分配,但是这些都没有用。
@tf.function
def sdft_func(self,input_tensor):
for i in range(self.N_t):
#retrieving variables so that I have direct access to it
#instead of getting access to the read tensor
_, _, self.in_s = self.get_variables()
last = self.in_s[self.N_t-1]
for j in range(self.N_t,0,-1):
_, _, self.in_s = self.get_variables()
val = self.in_s[j-1]
#The line below gives the error
self.in_s = self.in_s[j].assign(val)
print(self.in_s)
我得到的错误如下:
TypeError:在op'strided_slice_1 / _assign'中,输入类型([tf.complex64,tf.int32,tf.int32,tf.int32,tf.complex64])与预期的类型([tf.complex64_ref, tf.int32,tf.int32,tf.int32,tf.complex64])
提前谢谢您!
答案 0 :(得分:0)
似乎我缩小了问题,因为它与尝试在tf.function函数内设置tf.complex64变量有关。因此,为了克服这一点,我简单地抽象了该操作,以便在tf.function函数之外进行变量设置。解决方案请参见下文:
def sdft_func(self,input_tensor):
@tf.function
def func(input_tensor,N_t,in_s,coeffs,freqs):
in_s = tf.identity(in_s)
coeffs = tf.identity(coeffs)
freqs = tf.identity(freqs)
for i in range(N_t):
last = in_s[self.N_t-1]
in_s = in_s[:-1]
new_val = tf.expand_dims(tf.complex(input_tensor[i],
tf.cast(0.0,dtype=tf.float32)),0)
in_s = tf.concat([new_val,in_s],axis=0)
delta = in_s[0] - last
freqs_2 = tf.TensorArray(tf.complex64,size=self.N)
for j in range(self.N_t):
freqs_2 = freqs_2.write(j,(freqs[j]+delta)*coeffs[j])
freqs = freqs_2.stack()
freqs.set_shape([self.N])
return freqs,in_s
new_freqs, new_in_s = func(input_tensor,self.N_t,
self.in_s,self.coeffs,self.freqs)
self.in_s = self.in_s.assign(new_in_s)
self.freqs = self.freqs.assign(new_freqs)