我目前正在尝试在Keras中实现自定义损失功能。但是，由于我的模型是VGG和我的自定义网络的结合，所以我无法获得子模型的输出。
这是网络定义：
model = VGG19(weights='imagenet', include_top=False)
#model.summary()
def change_model(model, new_input_shape=(None, 40, 40, 3)):
''' Change the input size of the provided network'''
# replace input shape of first layer
model._layers[0].batch_input_shape = new_input_shape
# rebuild model architecture by exporting and importing via json
new_model = keras.models.model_from_json(model.to_json())
# copy weights from old model to new one
for layer in new_model.layers:
try:
layer.set_weights(model.get_layer(name=layer.name).get_weights())
print("Loaded layer {}".format(layer.name))
except:
print("Could not transfer weights for layer {}".format(layer.name))
return new_model
new_model = change_model(model,new_input_shape=(None, 1024, 1024, 3))
new_model.summary()
for layer in new_model.layers:
layer.trainable = False
vector_1 = new_model.get_layer("block4_conv4").output
def create_detector_network(kernel_reg = 0.):
input = Input(shape=(128, 128, 512))
x = Conv2D(128, kernel_size=3, strides=1, name='detect_1', padding='same', kernel_regularizer=regularizers.l2(kernel_reg))(input)
x = BatchNormalization()(x)
x = Conv2D(1+pow(8,2), kernel_size=1, strides=1, name='detect_2', kernel_regularizer=regularizers.l2(kernel_reg))(x)
x = BatchNormalization()(x)
prob = Activation('softmax')(x)
prob = Lambda(lambda x: x[:,:, :, :-1], output_shape= (128, 128, 64))(prob) #x[:, :, :-1]
prob = keras.layers.UpSampling2D(size=(8, 8), data_format=None, interpolation='nearest')(prob)
prob = Conv2D(1, kernel_size=1, strides=1, name='reduce_dim')(prob)
return Model(input, [prob, x])
detector_model = create_detector_network()
detector_model.summary()
output = detector_model(vector_1)
full_model = Model(inputs=new_model.input, outputs=output)
和摘要可以在这里看到： 我的网络：
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_4 (InputLayer) (None, 128, 128, 512) 0
_________________________________________________________________
detect_1 (Conv2D) (None, 128, 128, 128) 589952
_________________________________________________________________
batch_normalization_3 (Batch (None, 128, 128, 128) 512
_________________________________________________________________
detect_2 (Conv2D) (None, 128, 128, 65) 8385
_________________________________________________________________
batch_normalization_4 (Batch (None, 128, 128, 65) 260
_________________________________________________________________
activation_2 (Activation) (None, 128, 128, 65) 0
_________________________________________________________________
lambda_2 (Lambda) (None, 128, 128, 64) 0
_________________________________________________________________
up_sampling2d_2 (UpSampling2 (None, 1024, 1024, 64) 0
_________________________________________________________________
reduce_dim (Conv2D) (None, 1024, 1024, 1) 65
=================================================================
Total params: 599,174
Trainable params: 598,788
Non-trainable params: 386
_________________________________________________________________
VGG +我的网络：
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_7 (InputLayer) (None, 1024, 1024, 3) 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, 1024, 1024, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 1024, 1024, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 512, 512, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 512, 512, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 512, 512, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 256, 256, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 256, 256, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 256, 256, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 256, 256, 256) 590080
_________________________________________________________________
block3_conv4 (Conv2D) (None, 256, 256, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 128, 128, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 128, 128, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 128, 128, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 128, 128, 512) 2359808
_________________________________________________________________
block4_conv4 (Conv2D) (None, 128, 128, 512) 2359808
_________________________________________________________________
model_7 (Model) (None, 128, 128, 65) 599109
=================================================================
Total params: 11,184,261
Trainable params: 598,723
Non-trainable params: 10,585,538
_________________________________________________________________
这是我训练网络的方式：
losses = {
"detect_2": "mse",
"reduce_dim": "mse",
}
full_model.compile(optimizer='adam',
loss=losses,
loss_weights={'prob': 1.0, 'main': 0})
full_model.summary()
history = full_model.fit_generator(
train_generator,
steps_per_epoch=100,
epochs=7,
validation_data=validation_generator,
validation_steps=80)
现在，我想使用'detect_2'和'reduce_dim'层的输出，并从中计算出损耗/精度。但是，当我运行代码时，出现以下错误：
ValueError: Unknown entry in loss dictionary: "detect_2". Only expected the following keys: ['model_7', 'model_7']
很显然，某处一定有一个错误，因为字典不能两次具有相同的键。 那么有人可以告诉我需要进行哪些更改才能获得图层的输出吗？
答案 0 :(得分：0)
如果'detect_2'和'reduce_dim'是模型的图层，则需要将其输出指定为网络的输出。然后，您可以通过“ y_true”和“ y_pred”以自定义丢失的方式访问它们，它们将分别保存所有网络输出，用于地面真实性和预测。像这样：
import pygame
import time
import random
pygame.init()
window = pygame.display.set_mode((500,500))
pygame.display.set_caption("I am a hacker")
# player class
class players(object):
def __init__(self,x,y,height,width):
self.x = x
self.y = y
self.height = height
self.width = width
self.isJump = False
self.JumpCount = 10
self.fall = 0
self.speed = 5
# enemy class
class enemys(object):
def __init__(self,cordx,cordy,heights,widths):
self.cordx = cordx
self.cordy = cordy
self.heights = heights
self.widths = widths
# color blue for player
blue = (32,207,173)
red = (255,0,0)
orange = (207,135,32)
# FPS
FPS = 60
clock = pygame.time.Clock()
display_width = 50
display_height = 50
font_style = pygame.font.SysFont("bahnschrift", 25)
score_font = pygame.font.SysFont("comicsansms", 35)
# -----------------------------------------------------
# scoring and apple varabiles etc
snake_block = 10
snake_speed = 15
def Your_score(score):
value = score_font.render("Your Score: " + str(score), True, red)
window.blit(value, [0, 0])
def our_snake(snake_block, snake_list):
for x in snake_list:
pygame.draw.rect(window, red, [x[0], x[1], snake_block, snake_block])
def message(msg, color):
mesg = font_style.render(msg, True, color)
window.blit(mesg, [500 / 6, 500 / 3])
game_over = False
game_close = False
x1 = 500 / 2
y1 = 500 / 2
x1_change = 0
y1_change = 0
snake_List = []
Length_of_snake = 1
foodx = round(random.randrange(0, 500 - snake_block) / 10.0) * 10.0
foody = round(random.randrange(0, 500 - snake_block) / 10.0) * 10.0
# ------------------------------------------------------------------------------------------
# Main Loop
playerman = players(50,390,50,50)
enemyman = enemys(190,390,150,10)
runninggame = True
while runninggame:
clock.tick(FPS)
for event in pygame.event.get():
if event.type == pygame.QUIT:
runninggame = False
# ------------------------------------------------------------------------
# Scoring and Apple System
if x1 >= 500 or x1 < 0 or y1 >= 500 or y1 < 0:
game_close = True
x1 += x1_change
y1 += y1_change
window.fill(red)
pygame.draw.rect(window, red, [foodx, foody, snake_block, snake_block])
snake_Head = []
snake_Head.append(x1)
snake_Head.append(y1)
snake_List.append(snake_Head)
if len(snake_List) > Length_of_snake:
del snake_List[0]
for x in snake_List[:-1]:
if x == snake_Head:
game_close = True
our_snake(snake_block, snake_List)
Your_score(Length_of_snake - 1)
pygame.display.update()
if x1 == foodx and y1 == foody:
foodx = round(random.randrange(0, 500 - snake_block) / 10.0) * 10.0
foody = round(random.randrange(0, 500 - snake_block) / 10.0) * 10.0
Length_of_snake += 1
# ------------------------------------------------------------------------------
window.fill((0,0,0))
player = pygame.draw.rect(window,(blue),(playerman.x,playerman.y,playerman.height,playerman.width))
enemy = pygame.draw.rect(window,(orange),(enemyman.cordx,enemyman.cordy,enemyman.heights,enemyman.widths))
keys = pygame.key.get_pressed()
if keys[pygame.K_LEFT] and playerman.x > playerman.speed:
playerman.x -= playerman.speed
if keys[pygame.K_RIGHT] and playerman.x < 500 - playerman.width - playerman.speed:
playerman.x += playerman.speed
if not playerman.isJump:
playerman.y += playerman.fall
playerman.fall += 1
# ----------------------------------------------------- # enem1 collisio
# both of my 2 enemy squares collisions push me back when ever I Jump on the top of them on there sides but when I jump on the middle of of both of them it seems to work if I just want it so when I jump on both of my squares I just don't get pushed back
player.topleft = (playerman.x, playerman.y)
collide = False
playerman.isJump = False
if player.colliderect(enemy):
collide = True
playerman.isJump = False
playerman.y = enemy.top - player.height
if player.right > enemy.left and player.left < enemy.left - player.width:
playerman.x = enemy.left - player.width
if player.left < enemy.right and player.right > enemy.right + player.width:
playerman.x = enemy.right
if player.bottom >= 500:
collide = True
playerman.isJump = False
playerman.JumpCount = 10
playerman.y = 500 - player.height
if collide:
if keys[pygame.K_SPACE]:
playerman.isJump = True
playerman.fall = 0
else:
if playerman.JumpCount > 0:
playerman.y -= (playerman.JumpCount*abs(playerman.JumpCount)) * 0.5
playerman.JumpCount -= 1
else:
playerman.JumpCount = 10
playerman.isJump = False
pygame.display.update()
pygame.quit()
``
然后根据需要随意使用它们
return Model([input], [detect_2_output, reduce_dim_output])
答案 1 :(得分：0)
我认为模型和所有串联都可以。...
尝试以这种方式进行编译，避免使用图层名称，因为它们被隐藏在另一个模型中
losses = ['mse','mse']
full_model.compile(optimizer='adam',
loss=losses ,
loss_weights=[1,0])