对于某些损失，tf.tape.gradient()返回None

Question

我试图弄清楚为什么有时tf.GradientTape().gradient会返回None，所以我使用了下面三个损失函数(mmd0()，mmd1()，mmd2())，虽然mmd0和mmd1的格式略有不同，但仍然会返回梯度，但对于mmd2，梯度是None。我打印出了这三个函数的损失，有人知道为什么它会这样吗？

def mmd0(x, y): # a and b are lists of aribiturary lengths
  return x  

def mmd1(x1, x2): # a and b are lists of aribiturary lengths
  dis = sum([x**2 for x in x1])/len(x1) - sum([x**2 for x in x2])/len(x2)
  return dis**2

def mmd2(x, y):
  dis = x-y
  return [tf.convert_to_tensor(elem) for elem in dis]

def get_MMD_norm(errors, sigma=0.1): 
  x2 = np.random.normal(0, sigma, len(errors))
  loss0 = mmd0(errors, x2)
  loss1 = mmd1(errors, x2)
  loss2 = mmd2(errors, x2)
  print("loss0:", loss0)
  print("loss1:", loss1)
  print("loss2:", loss2)
  return tf.cast(loss2, tf.float32)

def loss(model, x, y, sigma=0.1):
  y_ = model(x) # y_.shape is (batch_size, 3) for Iris dataset
  losses = []
  loss_object = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
  for i in range(y.shape[0]):
    loss = loss_object(y_true=y[i], y_pred=y_[i])
    losses.append(loss) 
  batch_loss = get_MMD_norm(losses)
  single_losses_list = [loss.numpy() for loss in losses]
  return tf.convert_to_tensor(batch_loss, dtype=np.float32), single_losses_list

def grad(model, inputs, targets, sigma=0.1):
  with tf.GradientTape() as tape:
    tape.watch(model.trainable_variables)
    batch_loss, single_losses = loss(model, inputs, targets, sigma=0.1)
  return tape.gradient(batch_loss, model.trainable_variables), batch_loss, single_losses 

grads, batch_loss, single_losses = grad(model, features, labels)
print("grads:", grads)
print("batch_loss:", batch_loss)
##########################################################
loss0: [<tf.Tensor: id=39621, shape=(), dtype=float32, numpy=2.1656876>, <tf.Tensor: id=39659, shape=(), dtype=float32, numpy=2.057112>, <tf.Tensor: id=39697, shape=(), dtype=float32, numpy=2.2769136>, <tf.Tensor: id=39735, shape=(), dtype=float32, numpy=2.0263004>, <tf.Tensor: id=39773, shape=(), dtype=float32, numpy=2.1568372>, <tf.Tensor: id=39811, shape=(), dtype=float32, numpy=0.7392154>, <tf.Tensor: id=39849, shape=(), dtype=float32, numpy=0.7742219>, <tf.Tensor: id=39887, shape=(), dtype=float32, numpy=2.2176154>, <tf.Tensor: id=39925, shape=(), dtype=float32, numpy=1.0187237>, <tf.Tensor: id=39963, shape=(), dtype=float32, numpy=2.160415>, <tf.Tensor: id=40001, shape=(), dtype=float32, numpy=0.80997854>, <tf.Tensor: id=40039, shape=(), dtype=float32, numpy=0.70803094>, <tf.Tensor: id=40077, shape=(), dtype=float32, numpy=0.8207226>, <tf.Tensor: id=40115, shape=(), dtype=float32, numpy=0.82957774>, <tf.Tensor: id=40153, shape=(), dtype=float32, numpy=0.88732547>, <tf.Tensor: id=40191, shape=(), dtype=float32, numpy=0.90633464>, <tf.Tensor: id=40229, shape=(), dtype=float32, numpy=0.7932346>, <tf.Tensor: id=40267, shape=(), dtype=float32, numpy=2.1767666>, <tf.Tensor: id=40305, shape=(), dtype=float32, numpy=0.80166155>, <tf.Tensor: id=40343, shape=(), dtype=float32, numpy=0.7831647>, <tf.Tensor: id=40381, shape=(), dtype=float32, numpy=0.77431095>, <tf.Tensor: id=40419, shape=(), dtype=float32, numpy=0.82067406>, <tf.Tensor: id=40457, shape=(), dtype=float32, numpy=0.74510425>, <tf.Tensor: id=40495, shape=(), dtype=float32, numpy=2.1666338>, <tf.Tensor: id=40533, shape=(), dtype=float32, numpy=0.7922478>, <tf.Tensor: id=40571, shape=(), dtype=float32, numpy=0.73235756>, <tf.Tensor: id=40609, shape=(), dtype=float32, numpy=2.1792874>, <tf.Tensor: id=40647, shape=(), dtype=float32, numpy=0.919183>, <tf.Tensor: id=40685, shape=(), dtype=float32, numpy=0.761979>, <tf.Tensor: id=40723, shape=(), dtype=float32, numpy=2.1664479>, <tf.Tensor: id=40761, shape=(), dtype=float32, numpy=0.77892226>, <tf.Tensor: id=40799, shape=(), dtype=float32, numpy=0.99058735>]
loss1: tf.Tensor(4.158007, shape=(), dtype=float32)
loss2: [<tf.Tensor: id=40935, shape=(), dtype=float64, numpy=2.325676997771268>, <tf.Tensor: id=40936, shape=(), dtype=float64, numpy=1.9988182000798667>, <tf.Tensor: id=40937, shape=(), dtype=float64, numpy=2.303379813455908>, <tf.Tensor: id=40938, shape=(), dtype=float64, numpy=2.0615775258879356>, <tf.Tensor: id=40939, shape=(), dtype=float64, numpy=2.2949723624257774>, <tf.Tensor: id=40940, shape=(), dtype=float64, numpy=0.7019287657319235>, <tf.Tensor: id=40941, shape=(), dtype=float64, numpy=0.8522054859739794>, <tf.Tensor: id=40942, shape=(), dtype=float64, numpy=2.0819949907118125>, <tf.Tensor: id=40943, shape=(), dtype=float64, numpy=1.065878291073558>, <tf.Tensor: id=40944, shape=(), dtype=float64, numpy=2.1225998300026805>, <tf.Tensor: id=40945, shape=(), dtype=float64, numpy=0.9485520218242218>, <tf.Tensor: id=40946, shape=(), dtype=float64, numpy=0.7221746903906889>, <tf.Tensor: id=40947, shape=(), dtype=float64, numpy=0.9985009994522388>, <tf.Tensor: id=40948, shape=(), dtype=float64, numpy=0.9143119687525019>, <tf.Tensor: id=40949, shape=(), dtype=float64, numpy=0.9230117922853999>, <tf.Tensor: id=40950, shape=(), dtype=float64, numpy=1.0220225043292934>, <tf.Tensor: id=40951, shape=(), dtype=float64, numpy=0.8735972169951878>, <tf.Tensor: id=40952, shape=(), dtype=float64, numpy=2.1279260795512753>, <tf.Tensor: id=40953, shape=(), dtype=float64, numpy=0.9597649765787801>, <tf.Tensor: id=40954, shape=(), dtype=float64, numpy=0.8338326272407959>, <tf.Tensor: id=40955, shape=(), dtype=float64, numpy=0.6674084331022461>, <tf.Tensor: id=40956, shape=(), dtype=float64, numpy=0.8679296826013285>, <tf.Tensor: id=40957, shape=(), dtype=float64, numpy=0.8174893483228802>, <tf.Tensor: id=40958, shape=(), dtype=float64, numpy=2.212290299049252>, <tf.Tensor: id=40959, shape=(), dtype=float64, numpy=0.7304098620074719>, <tf.Tensor: id=40960, shape=(), dtype=float64, numpy=0.8463413221121661>, <tf.Tensor: id=40961, shape=(), dtype=float64, numpy=2.3081013094190443>, <tf.Tensor: id=40962, shape=(), dtype=float64, numpy=1.0314178020997722>, <tf.Tensor: id=40963, shape=(), dtype=float64, numpy=0.774951045805575>, <tf.Tensor: id=40964, shape=(), dtype=float64, numpy=2.127838465488091>, <tf.Tensor: id=40965, shape=(), dtype=float64, numpy=0.909498425717612>, <tf.Tensor: id=40966, shape=(), dtype=float64, numpy=1.0217239989370837>]
grads: [None, None, None, None, None, None]
batch_loss: tf.Tensor(
[2.325677   1.9988182  2.3033798  2.0615776  2.2949724  0.7019288
 0.8522055  2.081995   1.0658783  2.1225998  0.948552   0.7221747
 0.998501   0.91431195 0.9230118  1.0220225  0.8735972  2.127926
 0.95976496 0.8338326  0.6674084  0.8679297  0.8174893  2.2122903
 0.73040986 0.8463413  2.3081014  1.0314178  0.77495104 2.1278384
 0.90949845 1.021724  ], shape=(32,), dtype=float32)

Answer 1

你看过this answer吗？我想我也有类似的问题，我相信你的问题可能和我的问题有关。它必须与损失有关，该损失是通过过程中的某个步骤计算的，其中感兴趣的张量从磁带的开始到结束是“丢失的”。引用的答案指出，原始海报有一个区域，其中返回了numpy数组而不是tensorflow张量，因此导致梯度Tape无法计算梯度。

我可能错了，因为我根本不是tensorflow专家，但这是我在寻找类似问题的解决方案时不断出现的东西。