keras中model.fit（加载模型无训练后）和model.predict的不同输出

Question

def get_vgg_twoeyes() 是我的模型的定义。 我已经加载了一个预训练模型，它是在同一台计算机上训练的，然后我想对模型进行微调。在重新训练模型之前，我将 model.trainable 设置为 false 以确保模型的权重是固定的。在我训练之前，权重与保存的权重相同。我发现 model.fit 的输出与 model.predict 的输出不同。 因为我假设与 model.predict 具有相同权重的 model.fit 应该输出相同的结果，因为 model.trainable 为 false，这意味着 model.fit 的行为符合 model.predict。

def get_vgg_twoeyes(optimizer='adam', model_type='VGG16',fc1_size=1024, fc2_size=512, fc3_size=256)：

kern_init = initializers.glorot_normal()
img_input = Input(shape=(36, 60, 3), name='img_input')
headpose_input = Input(shape=(2,), name='headpose_input')

# create the base pre-trained model
if model_type == 'VGG19':
    base_model = VGG19(input_tensor=img_input, weights='imagenet', include_top=False)
elif model_type == 'VGG16':
    base_model = VGG16(input_tensor=img_input, weights='imagenet', include_top=False)
else:
    raise Exception('Unknown model type in get_vgg_twoeyes')

# add a global spatial average pooling layer
x = base_model.output
x = GlobalAveragePooling2D()(x)

# let's add a fully-connected layer
x = Dense(fc1_size, kernel_initializer=kern_init)(x)
x = concatenate([x, headpose_input])
x = BatchNormalization()(x)
x = Activation('relu')(x)

x = Dense(fc2_size, kernel_initializer=kern_init)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)

gaze_predictions = Dense(2, kernel_initializer=kern_init, name='pred_gaze')(x)

# this is the model we will train
model = Model(inputs=[img_input, headpose_input], outputs=gaze_predictions)
model.compile(optimizer=optimizer, loss=angle_loss, metrics=['accuracy', accuracy_angle])
return model

# fine-tune the model
models=load_model(model_path + "15Fold" + prefix + ''+str(i) + 
    suffix + ".h5",custom_objects={'accuracy_angle':accuracy_angle, 
     'angle_loss': angle_loss}))

model.trainable=False
adam = Adam(lr=0.0001, beta_1=0.9, beta_2=0.95)
model.compile(optimizer=adam, loss=angle_loss, metrics= ['accuracy', accuracy_angle]) 
model.fit({'img_input':cal_images,'headpose_input':cal_headposes},
cal_gazes,shuffle=False,batch_size=32,epochs=1,callbacks= 
  [losshistory()])

predgaze=model.predict({'img_input': cal_images, 'headpose_input': 
   cal_headposes},  batch_size=2,verbose=1)

Answer 1

在设置model.trainable=False 之后，您可能必须再次编译模型。否则，您可以像

一样手动单独冻结图层

for l in model.layers: l.trainable=False