如何预测，模型需要多少 epochs？

Question

我正在尝试使用自动编码器为图像着色，我有以下模型：

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 256, 256, 32)      320       
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 128, 128, 64)      18496     
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 128, 128, 64)      36928     
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 64, 64, 128)       73856     
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 64, 64, 128)       147584    
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 32, 32, 256)       295168    
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 32, 32, 256)       590080    
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 16, 16, 512)       1180160   
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 16, 16, 512)       2359808   
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 8, 8, 1024)        4719616   
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 8, 8, 1024)        9438208   
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 8, 8, 512)         4719104   
_________________________________________________________________
conv2d_transpose (Conv2DTran (None, 16, 16, 512)       2359808   
_________________________________________________________________
conv2d_12 (Conv2D)           (None, 16, 16, 256)       1179904   
_________________________________________________________________
conv2d_transpose_1 (Conv2DTr (None, 32, 32, 256)       590080    
_________________________________________________________________
conv2d_13 (Conv2D)           (None, 32, 32, 128)       295040    
_________________________________________________________________
conv2d_transpose_2 (Conv2DTr (None, 64, 64, 128)       147584    
_________________________________________________________________
conv2d_14 (Conv2D)           (None, 64, 64, 64)        73792     
_________________________________________________________________
conv2d_transpose_3 (Conv2DTr (None, 128, 128, 64)      36928     
_________________________________________________________________
conv2d_15 (Conv2D)           (None, 128, 128, 32)      18464     
_________________________________________________________________
conv2d_transpose_4 (Conv2DTr (None, 256, 256, 32)      9248      
_________________________________________________________________
conv2d_16 (Conv2D)           (None, 256, 256, 16)      4624      
_________________________________________________________________
conv2d_17 (Conv2D)           (None, 256, 256, 2)       290       
=================================================================
Total params: 28,295,090
Trainable params: 28,295,090
Non-trainable params: 0

我还准备了包含 80 000 张图像的数据集。 40 000 人和 40 000 自然人。

我的尝试：

100 个纪元：

120 个纪元：

160 个纪元：

210 个纪元：

300 个纪元：

我使用 LAB 颜色空间和 MAE 作为损失函数。

如何设置最合适的 epoch 数？

Answer 1

对于需要多少个 epoch 没有确切的答案。重要的是模型不能过度拟合或数据拟合不足。您可以使用 L1 / L2 正则化器来避免记忆训练数据集。另一方面，当loss 或val_loss 没有发生显着变化时，您可以使用tf.keras.callbacks.EarlyStopping 来削减纪元。