Question

我很确定这是一个错误-尽管正在寻找解决方法。我将注意到，我使用CNN作为特征提取器，然后使用tf.signal的滑动窗口使CNN的输出倍数，然后馈入密集网络。因此，这是一个有点奇怪的体系结构（有充分的理由）-但我希望渐变仅能在整个窗口框架内平均（真的，我不在乎-只需以某种合理的方式进行处理即可）。无论如何-一个示例模型.....假设args.num_frames = 6。

import sys
import pygame

class Skel():

    def __init__(self, screen):
        #Initialize skeleteon and set screen
        self.screen = screen

        #Get rectangle of skeleton and screen
        self.image = pygame.image.load("EM17_skeleton.bmp")
        self.rect = self.image.get_rect()
        self.screen_rect = screen.get_rect()

        #Place skeleton at center
        self.rect.centerx = self.screen_rect.centerx
        self.rect.centery = self.screen_rect.centery

    def draw(self,screen):
        """Draw skeleton at location"""
        screen.blit(self.image, self.rect)

skel = Skel(screen)

def run_game(self):
    pygame.init()
    screen = pygame.display.set_mode((400,400))

    #Set background color
    bg_color = (0,0,160)

    #Set loop
    while True:
        for event in pygame.event.get():
            if event.type == pygame.QUIT:
                sys.exit()
        screen.fill(bg_color)
        skel.draw(screen)
        pygame.display.flip()


run_game()

现在，如果我们这样做......

 inputs = tf.keras.Input(shape=data_loader_train[0][0][0].shape, name='img') ## (108, 192, 3)
    x = layers.Conv2D(32, 3, activation='relu')(inputs)
    x = layers.Conv2D(16, 3, activation='relu')(x)
    block_1_output = layers.MaxPooling2D(2)(x)

    x = layers.Conv2D(16, 3, activation='relu', padding='same')(block_1_output)
    block_2_output = layers.add([x, block_1_output])
    block_2_output = layers.MaxPooling2D(2)(block_2_output)

    x = layers.Conv2D(16, 3, activation='relu', padding='same')(block_2_output)
    block_3_output = layers.add([x, block_2_output])
    block_3_output = layers.MaxPooling2D(2)(block_3_output)

    x = layers.Conv2D(32, 3, activation='relu')(block_3_output)
    x = layers.GlobalAveragePooling2D()(x)

### if this part is included we get an error ##########
    x = layers.Flatten()(x)
    x = tf.signal.frame(x,args.num_frames,1, axis=0)
    x = layers.Flatten()(x)
##############################
    x = layers.Dense(16, activation='relu')(x)
    x = layers.Dense(1)(x)
    counts = tf.keras.activations.softplus(x)

    model = tf.keras.Model(inputs, counts, name='toy_resnet')

在“ grads = ...”处出现以下错误

Error = AssertionError：期望所有args是张量或变量；但是得到了CompositeTensor：[

当tf.signal.frame用于TF2.0时无法获得tape.gradient

0 个答案: