我正在尝试使用暹罗神经网络来检查两个图像是否相似。作为参考，我从标准示例here开始，虽然我更改了某些帖子中建议的 contrastive_loss 和 compute_accuracy 中的返回值，但它们似乎运行良好（我将准确度提高到0.98并将损失降低到0.01）。
现在我想使用卷积基础VGG和我自己的数据集（大小为200x200和3个通道），并将最后一层设置为可训练（尽管我尝试使用整个网络训练）我所做的是here
我的模型如下：
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_2 (InputLayer) (None, 200, 200, 3) 0
__________________________________________________________________________________________________
input_3 (InputLayer) (None, 200, 200, 3) 0
__________________________________________________________________________________________________
vgg16 (Model) (None, 6, 6, 512) 14714688 input_2[0][0]
input_3[0][0]
__________________________________________________________________________________________________
lambda_1 (Lambda) (None, 1) 0 vgg16[1][0]
vgg16[2][0]
==================================================================================================
Total params: 14,714,688
Trainable params: 7,079,424
Non-trainable params: 7,635,264
当我开始训练程序时，每个时期的损失和准确度值始终为0.5。
Train on 2000 samples, validate on 800 samples
Epoch 1/10
2000/2000 [==============================] - 2561s 1s/step - loss: 0.5005 - acc: 0.4998 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 2/10
2000/2000 [==============================] - 2563s 1s/step - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 3/10
1999/2000 [============================>.] - ETA: 0s - loss: 0.4997 - acc: 0.5003Epoch 4/10
2000/2000 [==============================] - 2553s 1s/step - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 5/10
2000/2000 [==============================] - 2554s 1s/step - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 6/10
2000/2000 [==============================] - 2562s 1s/step - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 7/10
763/2000 [==========>...................] - ETA: 19:54 - loss: 0.4980 - acc: 0.5020
如果我在训练后预测模型，则所有输出均为0，形状为(2000, 1, 6, 512)，用于训练数据
关于我做错了什么的任何建议？
谢谢