Keras model gives different prediction on the same input during fit() and predict()

Question

I'm training a simple adversarial image to break a pretrained model. However, the result I obtained during the fit() process is different from calling predict() on the same input (constant input).

model.trainable = False
gan = Sequential()
gan.add(Dense( 256 * 256 * 3, use_bias=False, input_shape=(1,)))
gan.add(Reshape((256, 256, 3)))
gan.add(model)

gan.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_2 (Dense)              (None, 196608)            196608    
_________________________________________________________________
reshape_2 (Reshape)          (None, 256, 256, 3)       0         
_________________________________________________________________
sequential_1 (Sequential)    (None, 2)                 24952610  
=================================================================
Total params: 25,149,218
Trainable params: 196,608
Non-trainable params: 24,952,610
_________________________________________________________________

img = img.reshape(256, 256, 3)

def custom_loss(layer):

    # Create a loss function that adds the MSE loss to the mean of all squared activations of a specific layer
    def loss(y_true,y_pred):
        y_true = K.print_tensor(y_true, message='y_true = ')
        y_pred = K.print_tensor(y_pred, message='y_pred = ')
        label_diff = K.square(y_pred - y_true)

        return K.mean(label_diff)


    # Return a function
    return loss

gan.compile(optimizer='adam',
              loss=custom_loss(gan.layers[1]), # Call the loss function with the selected layer
              metrics=['accuracy'])

x = np.ones((1,1))
goal = np.array([0, 1])
y = goal.reshape((1,2))

gan.fit(x, y, epochs=300, verbose=1)

During fit(), the loss is decreasing nicely

Epoch 1/300
1/1 [==============================] - 5s 5s/step - loss: 0.9950 - acc: 0.0000e+00
...
Epoch 300/300
1/1 [==============================] - 0s 46ms/step - loss: 0.0045 - acc: 1.0000

In the backend, the y_pred and y_true were also correct

......
y_true = [[0 1]]
y_pred = [[0.100334756 0.899665236]]
y_true = [[0 1]]
y_pred = [[0.116679631 0.883320332]]
y_true = [[0 1]]
y_pred = [[0.0832592845 0.916740656]]
y_true = [[0 1]]
y_pred = [[0.098835744 0.901164234]]
y_true = [[0 1]]
y_pred = [[0.0979194269 0.902080595]]
y_true = [[0 1]]
y_pred = [[0.057831794 0.942168236]]
y_true = [[0 1]]y_pred = [[0.0760448873 0.923955142]]

y_true = [[0 1]]
y_pred = [[0.041532293 0.958467722]]
y_true = [[0 1]]
y_pred = [[0.0667938739 0.933206141]]

print(gan.predict(x))

Gives

[[0.99923825 0.00076174]]

---

Tried with both pretrained Resnet and InceptionV3 and both are experiencing the same problem. Attached is model.summary()

For Inception:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
inception_v3 (Model)         (None, None, None, 2048)  21802784  
_________________________________________________________________
global_average_pooling2d_1 ( (None, 2048)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 1024)              2098176   
_________________________________________________________________
dropout_1 (Dropout)          (None, 1024)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 1024)              1049600   
_________________________________________________________________
dropout_2 (Dropout)          (None, 1024)              0         
_________________________________________________________________
dense_3 (Dense)              (None, 2)                 2050      
=================================================================
Total params: 24,952,610
Trainable params: 14,264,706
Non-trainable params: 10,687,904
_________________________________________________________________

For Resnet:

Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 256, 256, 3)  0                                            
__________________________________________________________________________________________________
conv1_pad (ZeroPadding2D)       (None, 262, 262, 3)  0           input_1[0][0]                    
__________________________________________________________________________________________________
conv1 (Conv2D)                  (None, 128, 128, 64) 9472        conv1_pad[0][0]                  
__________________________________________________________________________________________________
bn_conv1 (BatchNormalization)   (None, 128, 128, 64) 256         conv1[0][0]                      
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 128, 128, 64) 0           bn_conv1[0][0]                   
__________________________________________________________________________________________________
pool1_pad (ZeroPadding2D)       (None, 130, 130, 64) 0           activation_1[0][0]               
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D)  (None, 64, 64, 64)   0           pool1_pad[0][0]                  
__________________________________________________________________________________________________
res2a_branch2a (Conv2D)         (None, 64, 64, 64)   4160        max_pooling2d_1[0][0]            
__________________________________________________________________________________________________
bn2a_branch2a (BatchNormalizati (None, 64, 64, 64)   256         res2a_branch2a[0][0]             
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 64, 64, 64)   0           bn2a_branch2a[0][0]              
__________________________________________________________________________________________________
res2a_branch2b (Conv2D)         (None, 64, 64, 64)   36928       activation_2[0][0]               
__________________________________________________________________________________________________
bn2a_branch2b (BatchNormalizati (None, 64, 64, 64)   256         res2a_branch2b[0][0]             
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 64, 64, 64)   0           bn2a_branch2b[0][0]              
__________________________________________________________________________________________________
res2a_branch2c (Conv2D)         (None, 64, 64, 256)  16640       activation_3[0][0]               
__________________________________________________________________________________________________
res2a_branch1 (Conv2D)          (None, 64, 64, 256)  16640       max_pooling2d_1[0][0]            
__________________________________________________________________________________________________
bn2a_branch2c (BatchNormalizati (None, 64, 64, 256)  1024        res2a_branch2c[0][0]             
__________________________________________________________________________________________________
bn2a_branch1 (BatchNormalizatio (None, 64, 64, 256)  1024        res2a_branch1[0][0]              
__________________________________________________________________________________________________
add_1 (Add)                     (None, 64, 64, 256)  0           bn2a_branch2c[0][0]              
                                                                 bn2a_branch1[0][0]               
__________________________________________________________________________________________________
activation_4 (Activation)       (None, 64, 64, 256)  0           add_1[0][0]                      
__________________________________________________________________________________________________
res2b_branch2a (Conv2D)         (None, 64, 64, 64)   16448       activation_4[0][0]               
__________________________________________________________________________________________________
bn2b_branch2a (BatchNormalizati (None, 64, 64, 64)   256         res2b_branch2a[0][0]             
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 64, 64, 64)   0           bn2b_branch2a[0][0]              
__________________________________________________________________________________________________
res2b_branch2b (Conv2D)         (None, 64, 64, 64)   36928       activation_5[0][0]               
__________________________________________________________________________________________________
bn2b_branch2b (BatchNormalizati (None, 64, 64, 64)   256         res2b_branch2b[0][0]             
__________________________________________________________________________________________________
activation_6 (Activation)       (None, 64, 64, 64)   0           bn2b_branch2b[0][0]              
__________________________________________________________________________________________________
res2b_branch2c (Conv2D)         (None, 64, 64, 256)  16640       activation_6[0][0]               
__________________________________________________________________________________________________
bn2b_branch2c (BatchNormalizati (None, 64, 64, 256)  1024        res2b_branch2c[0][0]             
__________________________________________________________________________________________________
add_2 (Add)                     (None, 64, 64, 256)  0           bn2b_branch2c[0][0]              
                                                                 activation_4[0][0]               
__________________________________________________________________________________________________
activation_7 (Activation)       (None, 64, 64, 256)  0           add_2[0][0]                      
__________________________________________________________________________________________________
res2c_branch2a (Conv2D)         (None, 64, 64, 64)   16448       activation_7[0][0]               
__________________________________________________________________________________________________
bn2c_branch2a (BatchNormalizati (None, 64, 64, 64)   256         res2c_branch2a[0][0]             
__________________________________________________________________________________________________
activation_8 (Activation)       (None, 64, 64, 64)   0           bn2c_branch2a[0][0]              
__________________________________________________________________________________________________
res2c_branch2b (Conv2D)         (None, 64, 64, 64)   36928       activation_8[0][0]               
__________________________________________________________________________________________________
bn2c_branch2b (BatchNormalizati (None, 64, 64, 64)   256         res2c_branch2b[0][0]             
__________________________________________________________________________________________________
activation_9 (Activation)       (None, 64, 64, 64)   0           bn2c_branch2b[0][0]              
__________________________________________________________________________________________________
res2c_branch2c (Conv2D)         (None, 64, 64, 256)  16640       activation_9[0][0]               
__________________________________________________________________________________________________
bn2c_branch2c (BatchNormalizati (None, 64, 64, 256)  1024        res2c_branch2c[0][0]             
__________________________________________________________________________________________________
add_3 (Add)                     (None, 64, 64, 256)  0           bn2c_branch2c[0][0]              
                                                                 activation_7[0][0]               
__________________________________________________________________________________________________
activation_10 (Activation)      (None, 64, 64, 256)  0           add_3[0][0]                      
__________________________________________________________________________________________________
res3a_branch2a (Conv2D)         (None, 32, 32, 128)  32896       activation_10[0][0]              
__________________________________________________________________________________________________
bn3a_branch2a (BatchNormalizati (None, 32, 32, 128)  512         res3a_branch2a[0][0]             
__________________________________________________________________________________________________
activation_11 (Activation)      (None, 32, 32, 128)  0           bn3a_branch2a[0][0]              
__________________________________________________________________________________________________
res3a_branch2b (Conv2D)         (None, 32, 32, 128)  147584      activation_11[0][0]              
__________________________________________________________________________________________________
bn3a_branch2b (BatchNormalizati (None, 32, 32, 128)  512         res3a_branch2b[0][0]             
__________________________________________________________________________________________________
activation_12 (Activation)      (None, 32, 32, 128)  0           bn3a_branch2b[0][0]              
__________________________________________________________________________________________________
res3a_branch2c (Conv2D)         (None, 32, 32, 512)  66048       activation_12[0][0]              
__________________________________________________________________________________________________
res3a_branch1 (Conv2D)          (None, 32, 32, 512)  131584      activation_10[0][0]              
__________________________________________________________________________________________________
bn3a_branch2c (BatchNormalizati (None, 32, 32, 512)  2048        res3a_branch2c[0][0]             
__________________________________________________________________________________________________
bn3a_branch1 (BatchNormalizatio (None, 32, 32, 512)  2048        res3a_branch1[0][0]              
__________________________________________________________________________________________________
add_4 (Add)                     (None, 32, 32, 512)  0           bn3a_branch2c[0][0]              
                                                                 bn3a_branch1[0][0]               
__________________________________________________________________________________________________
activation_13 (Activation)      (None, 32, 32, 512)  0           add_4[0][0]                      
__________________________________________________________________________________________________
res3b_branch2a (Conv2D)         (None, 32, 32, 128)  65664       activation_13[0][0]              
__________________________________________________________________________________________________
bn3b_branch2a (BatchNormalizati (None, 32, 32, 128)  512         res3b_branch2a[0][0]             
__________________________________________________________________________________________________
activation_14 (Activation)      (None, 32, 32, 128)  0           bn3b_branch2a[0][0]              
__________________________________________________________________________________________________
res3b_branch2b (Conv2D)         (None, 32, 32, 128)  147584      activation_14[0][0]              
__________________________________________________________________________________________________
bn3b_branch2b (BatchNormalizati (None, 32, 32, 128)  512         res3b_branch2b[0][0]             
__________________________________________________________________________________________________
activation_15 (Activation)      (None, 32, 32, 128)  0           bn3b_branch2b[0][0]              
__________________________________________________________________________________________________
res3b_branch2c (Conv2D)         (None, 32, 32, 512)  66048       activation_15[0][0]              
__________________________________________________________________________________________________
bn3b_branch2c (BatchNormalizati (None, 32, 32, 512)  2048        res3b_branch2c[0][0]             
__________________________________________________________________________________________________
add_5 (Add)                     (None, 32, 32, 512)  0           bn3b_branch2c[0][0]              
                                                                 activation_13[0][0]              
__________________________________________________________________________________________________
activation_16 (Activation)      (None, 32, 32, 512)  0           add_5[0][0]                      
__________________________________________________________________________________________________
res3c_branch2a (Conv2D)         (None, 32, 32, 128)  65664       activation_16[0][0]              
__________________________________________________________________________________________________
bn3c_branch2a (BatchNormalizati (None, 32, 32, 128)  512         res3c_branch2a[0][0]             
__________________________________________________________________________________________________
activation_17 (Activation)      (None, 32, 32, 128)  0           bn3c_branch2a[0][0]              
__________________________________________________________________________________________________
res3c_branch2b (Conv2D)         (None, 32, 32, 128)  147584      activation_17[0][0]              
__________________________________________________________________________________________________
bn3c_branch2b (BatchNormalizati (None, 32, 32, 128)  512         res3c_branch2b[0][0]             
__________________________________________________________________________________________________
activation_18 (Activation)      (None, 32, 32, 128)  0           bn3c_branch2b[0][0]              
__________________________________________________________________________________________________
res3c_branch2c (Conv2D)         (None, 32, 32, 512)  66048       activation_18[0][0]              
__________________________________________________________________________________________________
bn3c_branch2c (BatchNormalizati (None, 32, 32, 512)  2048        res3c_branch2c[0][0]             
__________________________________________________________________________________________________
add_6 (Add)                     (None, 32, 32, 512)  0           bn3c_branch2c[0][0]              
                                                                 activation_16[0][0]              
__________________________________________________________________________________________________
activation_19 (Activation)      (None, 32, 32, 512)  0           add_6[0][0]                      
__________________________________________________________________________________________________
res3d_branch2a (Conv2D)         (None, 32, 32, 128)  65664       activation_19[0][0]              
__________________________________________________________________________________________________
bn3d_branch2a (BatchNormalizati (None, 32, 32, 128)  512         res3d_branch2a[0][0]             
__________________________________________________________________________________________________
activation_20 (Activation)      (None, 32, 32, 128)  0           bn3d_branch2a[0][0]              
__________________________________________________________________________________________________
res3d_branch2b (Conv2D)         (None, 32, 32, 128)  147584      activation_20[0][0]              
__________________________________________________________________________________________________
bn3d_branch2b (BatchNormalizati (None, 32, 32, 128)  512         res3d_branch2b[0][0]             
__________________________________________________________________________________________________
activation_21 (Activation)      (None, 32, 32, 128)  0           bn3d_branch2b[0][0]              
__________________________________________________________________________________________________
res3d_branch2c (Conv2D)         (None, 32, 32, 512)  66048       activation_21[0][0]              
__________________________________________________________________________________________________
bn3d_branch2c (BatchNormalizati (None, 32, 32, 512)  2048        res3d_branch2c[0][0]             
__________________________________________________________________________________________________
add_7 (Add)                     (None, 32, 32, 512)  0           bn3d_branch2c[0][0]              
                                                                 activation_19[0][0]              
__________________________________________________________________________________________________
activation_22 (Activation)      (None, 32, 32, 512)  0           add_7[0][0]                      
__________________________________________________________________________________________________
res4a_branch2a (Conv2D)         (None, 16, 16, 256)  131328      activation_22[0][0]              
__________________________________________________________________________________________________
bn4a_branch2a (BatchNormalizati (None, 16, 16, 256)  1024        res4a_branch2a[0][0]             
__________________________________________________________________________________________________
activation_23 (Activation)      (None, 16, 16, 256)  0           bn4a_branch2a[0][0]              
__________________________________________________________________________________________________
res4a_branch2b (Conv2D)         (None, 16, 16, 256)  590080      activation_23[0][0]              
__________________________________________________________________________________________________
bn4a_branch2b (BatchNormalizati (None, 16, 16, 256)  1024        res4a_branch2b[0][0]             
__________________________________________________________________________________________________
activation_24 (Activation)      (None, 16, 16, 256)  0           bn4a_branch2b[0][0]              
__________________________________________________________________________________________________
res4a_branch2c (Conv2D)         (None, 16, 16, 1024) 263168      activation_24[0][0]              
__________________________________________________________________________________________________
res4a_branch1 (Conv2D)          (None, 16, 16, 1024) 525312      activation_22[0][0]              
__________________________________________________________________________________________________
bn4a_branch2c (BatchNormalizati (None, 16, 16, 1024) 4096        res4a_branch2c[0][0]             
__________________________________________________________________________________________________
bn4a_branch1 (BatchNormalizatio (None, 16, 16, 1024) 4096        res4a_branch1[0][0]              
__________________________________________________________________________________________________
add_8 (Add)                     (None, 16, 16, 1024) 0           bn4a_branch2c[0][0]              
                                                                 bn4a_branch1[0][0]               
__________________________________________________________________________________________________
activation_25 (Activation)      (None, 16, 16, 1024) 0           add_8[0][0]                      
__________________________________________________________________________________________________
res4b_branch2a (Conv2D)         (None, 16, 16, 256)  262400      activation_25[0][0]              
__________________________________________________________________________________________________
bn4b_branch2a (BatchNormalizati (None, 16, 16, 256)  1024        res4b_branch2a[0][0]             
__________________________________________________________________________________________________
activation_26 (Activation)      (None, 16, 16, 256)  0           bn4b_branch2a[0][0]              
__________________________________________________________________________________________________
res4b_branch2b (Conv2D)         (None, 16, 16, 256)  590080      activation_26[0][0]              
__________________________________________________________________________________________________
bn4b_branch2b (BatchNormalizati (None, 16, 16, 256)  1024        res4b_branch2b[0][0]             
__________________________________________________________________________________________________
activation_27 (Activation)      (None, 16, 16, 256)  0           bn4b_branch2b[0][0]              
__________________________________________________________________________________________________
res4b_branch2c (Conv2D)         (None, 16, 16, 1024) 263168      activation_27[0][0]              
__________________________________________________________________________________________________
bn4b_branch2c (BatchNormalizati (None, 16, 16, 1024) 4096        res4b_branch2c[0][0]             
__________________________________________________________________________________________________
add_9 (Add)                     (None, 16, 16, 1024) 0           bn4b_branch2c[0][0]              
                                                                 activation_25[0][0]              
__________________________________________________________________________________________________
activation_28 (Activation)      (None, 16, 16, 1024) 0           add_9[0][0]                      
__________________________________________________________________________________________________
res4c_branch2a (Conv2D)         (None, 16, 16, 256)  262400      activation_28[0][0]              
__________________________________________________________________________________________________
bn4c_branch2a (BatchNormalizati (None, 16, 16, 256)  1024        res4c_branch2a[0][0]             
__________________________________________________________________________________________________
activation_29 (Activation)      (None, 16, 16, 256)  0           bn4c_branch2a[0][0]              
__________________________________________________________________________________________________
res4c_branch2b (Conv2D)         (None, 16, 16, 256)  590080      activation_29[0][0]              
__________________________________________________________________________________________________
bn4c_branch2b (BatchNormalizati (None, 16, 16, 256)  1024        res4c_branch2b[0][0]             
__________________________________________________________________________________________________
activation_30 (Activation)      (None, 16, 16, 256)  0           bn4c_branch2b[0][0]              
__________________________________________________________________________________________________
res4c_branch2c (Conv2D)         (None, 16, 16, 1024) 263168      activation_30[0][0]              
__________________________________________________________________________________________________
bn4c_branch2c (BatchNormalizati (None, 16, 16, 1024) 4096        res4c_branch2c[0][0]             
__________________________________________________________________________________________________
add_10 (Add)                    (None, 16, 16, 1024) 0           bn4c_branch2c[0][0]              
                                                                 activation_28[0][0]              
__________________________________________________________________________________________________
activation_31 (Activation)      (None, 16, 16, 1024) 0           add_10[0][0]                     
__________________________________________________________________________________________________
res4d_branch2a (Conv2D)         (None, 16, 16, 256)  262400      activation_31[0][0]              
__________________________________________________________________________________________________
... omitted ... 
Total params: 23,593,859
Trainable params: 23,540,739
Non-trainable params: 53,120
__________________________________________________________________________________________________

Answer 1

Those pretrained models contain BatchNormalization layers.

It's expected that they perform differently between train and test (this is also true for Dropout layers, but the differences would not be so drastic).

A BatchNormalization during training will use the mean and variance from the current batch to do normalization, it will also apply some statistic compensation for the fact that one batch may not be representative of the full dataset.

But during evaluation, the BatchNormalization will use the adjusted values that were gathered during training for mean and variation. (In this case, gathered during "pretraining" not your training)

For BatchNormalization to work correctly, you need that the inputs to the pretrained model are in the same range as the model's original training data. Otherwise you have to leave the BatchNormalization layers trainable so the mean and variance adjust for your data.

But your training needs significant batch sizes as well as real data to train properly.

---

Hints for training images.

In the same module where you import the pretrained model, you can import the preprocess_input function. Give it some images loaded with keras.preprocessing.images.load_img and see what is the model's expected range.

When using ImageDataGenerator, you can pass this preprocess_input function so the generator gives you expected data.