Question

我是深度学习的新手。我的代码中发生了什么我将发布代码和错误我想做resnet +三重损失

ValueError：尺寸必须相等，但输入形状为[？，2,1]，[?, 0,1]的'loss / dropout_1_loss / sub'（op：'Sub'）的尺寸必须为2和0。


# define model -----------------------------------------
input_dim = X1.shape[1:]
#print(input_dim)
input_a = Input(shape=input_dim) # <class 'tensorflow.python.framework.ops.Tensor'>
input_b = Input(shape=input_dim)
input_c = Input(shape=input_dim)

model_a = resnet50.ResNet50(weights="imagenet", include_top=True, input_tensor=input_a)
model_b = resnet50.ResNet50(weights="imagenet", include_top=True, input_tensor=input_b)
model_c = resnet50.ResNet50(weights="imagenet", include_top=True, input_tensor=input_c)

for layer in model_a.layers:
    if(layer.name != 'fc1000'):
        layer.trainable = False
    layer.name = layer.name + "_1"

for layer in model_b.layers:
    if(layer.name != 'fc1000'):
        layer.trainable = False
    layer.name = layer.name + "_2"

for layer in model_c.layers:
    if(layer.name != 'fc1000'):
        layer.trainable = False
    layer.name = layer.name + "_3"

shared_fc1000 = model_a.get_layer('fc1000_1')
# weight transfer

a = model_a.get_layer('avg_pool_1').output
b = model_b.get_layer('avg_pool_2').output
c = model_c.get_layer('avg_pool_3').output

processed_a = shared_fc1000(a)
processed_b = shared_fc1000(b)
processed_c = shared_fc1000(c)

positive_dist = Lambda(euclidean_distance, output_shape=eucl_dist_output_shape)([processed_a, processed_b])
negative_dist = Lambda(euclidean_distance, output_shape=eucl_dist_output_shape)([processed_a, processed_c])

stacked_dists = Lambda( 
            lambda vects: K.stack(vects, axis=1),
            output_shape=eucl_dist_output_shape
)([positive_dist, negative_dist])


model = Model([input_a, input_b, input_c], output=Dropout(0.25)(stacked_dists))


model.summary()

rms = RMSprop()
model.compile(loss=triplet_loss, optimizer=rms)

ValueError：尺寸必须相等，但输入形状为[？，2,1]，[?, 0,1]的'loss / dropout_1_loss / sub'（op：'Sub'）的尺寸应为2和0

0 个答案: