急切模式下的张量流梯度返回零

Question

问题：我正在从保存的检查点加载一个简单的VGG16。我想在推理过程中为图像生成显着性。当我计算所需的梯度（输入图像的损失）时，我将所有梯度都恢复为零。关于我在这里失踪的任何想法都非常感谢！

tf版本：tensorflow-2.0alpha-gpu

该模型：

import tensorflow as tf
from tensorflow.keras.applications.vgg16 import VGG16 as KerasVGG16
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Flatten, Dense

class VGG16(Model):

    def __init__(self, num_classes, use_pretrained=True):

        super(VGG16, self).__init__()
        self.num_classes = num_classes
        self.use_pretrained = use_pretrained

        if use_pretrained:
            self.base_model = KerasVGG16(weights='imagenet', include_top=False)
            for layer in self.base_model.layers:
                layer.trainable = False
        else:
            self.base_model = KerasVGG16(include_top=False)

        self.flatten1 = Flatten(name='flatten')
        self.dense1 = Dense(4096, activation='relu', name='fc1')
        self.dense2 = Dense(100, activation='relu', name='fc2')
        self.dense3 = Dense(self.num_classes, activation='softmax', name='predictions')

    def call(self, inputs):

        x = self.base_model(tf.cast(inputs, tf.float32))
        x = self.flatten1(x)
        x = self.dense1(x)
        x = self.dense2(x)
        x = self.dense3(x)
        return x

我训练这个模型并将其保存到检查点并通过以下方式加载：

model = VGG16(num_classes=2, use_pretrained=False)
checkpoint = tf.train.Checkpoint(net=model)
        status = checkpoint.restore(tf.train.latest_checkpoint('./my_checkpoint'))
status.assert_consumed()

我验证权重是否正确加载。

获得测试图像

# load my image and make sure its float
img = tf.convert_to_tensor(image, dtype=tf.float64)
support_class = tf.convert_to_tensor(support_class, dtype=tf.float64)

得到渐变：

with tf.GradientTape(persistent=True) as g_tape:
    g_tape.watch(img)
    #g_tape.watch(model.base_model.trainable_variables)
    #g_tape.watch(model.trainable_variables)
    loss = tf.losses.CategoricalCrossentropy()(support_class, model(img))    
    gradients_wrt_image = g_tape.gradient(loss,
                                    img, unconnected_gradients=tf.UnconnectedGradients.NONE)

当我检查我的渐变时，它们都是零！知道我错过了什么吗？提前致谢！

Answer 1

渐变不是零，虽然它们非常小：

def almost_equals(a, b, decimal=6):
    try:
        np.testing.assert_almost_equal(a, b, decimal=decimal)
    except AssertionError:
        return False
    return True

image = [abs(np.random.normal(size=(32, 32, 3))) for _ in range(20)]
label = [[0, 1] if i % 3 == 0 else [1, 0] for i in range(20)]
img = tf.convert_to_tensor(image, dtype=tf.float64)
support_class = tf.convert_to_tensor(label, dtype=tf.float64)
loss_fn = tf.losses.CategoricalCrossentropy()

with tf.GradientTape(persistent=True) as tape:
    tape.watch(img)
    softmaxed = model(img)
    loss = loss_fn(support_class, softmaxed)
    grads = tape.gradient(loss, img, unconnected_gradients=tf.UnconnectedGradients.NONE)
    # summing up all gradients with reduction over all dimension:
    print(tf.reduce_sum(grads, axis=None).numpy()) # 0.07137820225818814
    # comparing to zeros:
    zeros_like_grads = np.zeros_like(grads.numpy())  
    for decimal in range(10, 0, -1):
        print('decimal: {0}: {1}'.format(decimal,
                                         almost_equals(zeros_like_grads,
                                                       grads.numpy(),
                                                       decimal=decimal)))
# decimal: 10: False
# decimal: 9: False
# decimal: 8: False
# decimal: 7: False
# decimal: 6: False
# decimal: 5: False
# decimal: 4: False
# decimal: 3: True
# decimal: 2: True
# decimal: 1: True

如你所见，只有从decimal=3开始才开始返回True。

Answer 2

因此，事实证明网络没有任何问题。问题与我在最终的Dense层中使用的softmax激活的行为有关。我没有考虑这样一个事实：softmax的非常自信的预测（例如我的一个预测[[1.0000000e + 00 1.9507678e-25]]）会使梯度为零（理论上非常接近零但实际上为零）。讨论这个以及如何应对它的有用线程：https://github.com/keras-team/keras/issues/5881

我的解决方案：当我想计算输入图像的渐变时，关闭softmax激活