在最后,我们将在之前完成模型的基础上,利用模型微调,来进一步提高val_acc。
model.summary()
模型层次
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) (None, 356, 356, 3) 0
__________________________________________________________________________________________________
conv1_pad (ZeroPadding2D) (None, 362, 362, 3) 0 input_1[0][0]
__________________________________________________________________________________________________
conv1 (Conv2D) (None, 178, 178, 64) 9472 conv1_pad[0][0]
__________________________________________________________________________________________________
bn_conv1 (BatchNormalization) (None, 178, 178, 64) 256 conv1[0][0]
__________________________________________________________________________________________________
activation_1 (Activation) (None, 178, 178, 64) 0 bn_conv1[0][0]
__________________________________________________________________________________________________
pool1_pad (ZeroPadding2D) (None, 180, 180, 64) 0 activation_1[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 89, 89, 64) 0 pool1_pad[0][0]
__________________________________________________________________________________________________
res2a_branch2a (Conv2D) (None, 89, 89, 64) 4160 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
bn2a_branch2a (BatchNormalizati (None, 89, 89, 64) 256 res2a_branch2a[0][0]
__________________________________________________________________________________________________
activation_2 (Activation) (None, 89, 89, 64) 0 bn2a_branch2a[0][0]
__________________________________________________________________________________________________
res2a_branch2b (Conv2D) (None, 89, 89, 64) 36928 activation_2[0][0]
__________________________________________________________________________________________________
bn2a_branch2b (BatchNormalizati (None, 89, 89, 64) 256 res2a_branch2b[0][0]
__________________________________________________________________________________________________
activation_3 (Activation) (None, 89, 89, 64) 0 bn2a_branch2b[0][0]
__________________________________________________________________________________________________
res2a_branch2c (Conv2D) (None, 89, 89, 256) 16640 activation_3[0][0]
__________________________________________________________________________________________________
res2a_branch1 (Conv2D) (None, 89, 89, 256) 16640 max_pooling2d_1[0][0]
__________________________________________________________________________________________________
bn2a_branch2c (BatchNormalizati (None, 89, 89, 256) 1024 res2a_branch2c[0][0]
__________________________________________________________________________________________________
bn2a_branch1 (BatchNormalizatio (None, 89, 89, 256) 1024 res2a_branch1[0][0]
__________________________________________________________________________________________________
add_1 (Add) (None, 89, 89, 256) 0 bn2a_branch2c[0][0]
bn2a_branch1[0][0]
__________________________________________________________________________________________________
activation_4 (Activation) (None, 89, 89, 256) 0 add_1[0][0]
__________________________________________________________________________________________________
res2b_branch2a (Conv2D) (None, 89, 89, 64) 16448 activation_4[0][0]
__________________________________________________________________________________________________
bn2b_branch2a (BatchNormalizati (None, 89, 89, 64) 256 res2b_branch2a[0][0]
__________________________________________________________________________________________________
activation_5 (Activation) (None, 89, 89, 64) 0 bn2b_branch2a[0][0]
__________________________________________________________________________________________________
res2b_branch2b (Conv2D) (None, 89, 89, 64) 36928 activation_5[0][0]
__________________________________________________________________________________________________
bn2b_branch2b (BatchNormalizati (None, 89, 89, 64) 256 res2b_branch2b[0][0]
__________________________________________________________________________________________________
activation_6 (Activation) (None, 89, 89, 64) 0 bn2b_branch2b[0][0]
__________________________________________________________________________________________________
res2b_branch2c (Conv2D) (None, 89, 89, 256) 16640 activation_6[0][0]
__________________________________________________________________________________________________
bn2b_branch2c (BatchNormalizati (None, 89, 89, 256) 1024 res2b_branch2c[0][0]
__________________________________________________________________________________________________
add_2 (Add) (None, 89, 89, 256) 0 bn2b_branch2c[0][0]
activation_4[0][0]
__________________________________________________________________________________________________
activation_7 (Activation) (None, 89, 89, 256) 0 add_2[0][0]
__________________________________________________________________________________________________
res2c_branch2a (Conv2D) (None, 89, 89, 64) 16448 activation_7[0][0]
__________________________________________________________________________________________________
bn2c_branch2a (BatchNormalizati (None, 89, 89, 64) 256 res2c_branch2a[0][0]
__________________________________________________________________________________________________
activation_8 (Activation) (None, 89, 89, 64) 0 bn2c_branch2a[0][0]
__________________________________________________________________________________________________
res2c_branch2b (Conv2D) (None, 89, 89, 64) 36928 activation_8[0][0]
__________________________________________________________________________________________________
bn2c_branch2b (BatchNormalizati (None, 89, 89, 64) 256 res2c_branch2b[0][0]
__________________________________________________________________________________________________
activation_9 (Activation) (None, 89, 89, 64) 0 bn2c_branch2b[0][0]
__________________________________________________________________________________________________
res2c_branch2c (Conv2D) (None, 89, 89, 256) 16640 activation_9[0][0]
__________________________________________________________________________________________________
bn2c_branch2c (BatchNormalizati (None, 89, 89, 256) 1024 res2c_branch2c[0][0]
__________________________________________________________________________________________________
add_3 (Add) (None, 89, 89, 256) 0 bn2c_branch2c[0][0]
activation_7[0][0]
__________________________________________________________________________________________________
activation_10 (Activation) (None, 89, 89, 256) 0 add_3[0][0]
__________________________________________________________________________________________________
res3a_branch2a (Conv2D) (None, 45, 45, 128) 32896 activation_10[0][0]
__________________________________________________________________________________________________
bn3a_branch2a (BatchNormalizati (None, 45, 45, 128) 512 res3a_branch2a[0][0]
__________________________________________________________________________________________________
activation_11 (Activation) (None, 45, 45, 128) 0 bn3a_branch2a[0][0]
__________________________________________________________________________________________________
res3a_branch2b (Conv2D) (None, 45, 45, 128) 147584 activation_11[0][0]
__________________________________________________________________________________________________
bn3a_branch2b (BatchNormalizati (None, 45, 45, 128) 512 res3a_branch2b[0][0]
__________________________________________________________________________________________________
activation_12 (Activation) (None, 45, 45, 128) 0 bn3a_branch2b[0][0]
__________________________________________________________________________________________________
res3a_branch2c (Conv2D) (None, 45, 45, 512) 66048 activation_12[0][0]
__________________________________________________________________________________________________
res3a_branch1 (Conv2D) (None, 45, 45, 512) 131584 activation_10[0][0]
__________________________________________________________________________________________________
bn3a_branch2c (BatchNormalizati (None, 45, 45, 512) 2048 res3a_branch2c[0][0]
__________________________________________________________________________________________________
bn3a_branch1 (BatchNormalizatio (None, 45, 45, 512) 2048 res3a_branch1[0][0]
__________________________________________________________________________________________________
add_4 (Add) (None, 45, 45, 512) 0 bn3a_branch2c[0][0]
bn3a_branch1[0][0]
__________________________________________________________________________________________________
activation_13 (Activation) (None, 45, 45, 512) 0 add_4[0][0]
__________________________________________________________________________________________________
res3b_branch2a (Conv2D) (None, 45, 45, 128) 65664 activation_13[0][0]
__________________________________________________________________________________________________
bn3b_branch2a (BatchNormalizati (None, 45, 45, 128) 512 res3b_branch2a[0][0]
__________________________________________________________________________________________________
activation_14 (Activation) (None, 45, 45, 128) 0 bn3b_branch2a[0][0]
__________________________________________________________________________________________________
res3b_branch2b (Conv2D) (None, 45, 45, 128) 147584 activation_14[0][0]
__________________________________________________________________________________________________
bn3b_branch2b (BatchNormalizati (None, 45, 45, 128) 512 res3b_branch2b[0][0]
__________________________________________________________________________________________________
activation_15 (Activation) (None, 45, 45, 128) 0 bn3b_branch2b[0][0]
__________________________________________________________________________________________________
res3b_branch2c (Conv2D) (None, 45, 45, 512) 66048 activation_15[0][0]
__________________________________________________________________________________________________
bn3b_branch2c (BatchNormalizati (None, 45, 45, 512) 2048 res3b_branch2c[0][0]
__________________________________________________________________________________________________
add_5 (Add) (None, 45, 45, 512) 0 bn3b_branch2c[0][0]
activation_13[0][0]
__________________________________________________________________________________________________
activation_16 (Activation) (None, 45, 45, 512) 0 add_5[0][0]
__________________________________________________________________________________________________
res3c_branch2a (Conv2D) (None, 45, 45, 128) 65664 activation_16[0][0]
__________________________________________________________________________________________________
bn3c_branch2a (BatchNormalizati (None, 45, 45, 128) 512 res3c_branch2a[0][0]
__________________________________________________________________________________________________
activation_17 (Activation) (None, 45, 45, 128) 0 bn3c_branch2a[0][0]
__________________________________________________________________________________________________
res3c_branch2b (Conv2D) (None, 45, 45, 128) 147584 activation_17[0][0]
__________________________________________________________________________________________________
bn3c_branch2b (BatchNormalizati (None, 45, 45, 128) 512 res3c_branch2b[0][0]
__________________________________________________________________________________________________
activation_18 (Activation) (None, 45, 45, 128) 0 bn3c_branch2b[0][0]
__________________________________________________________________________________________________
res3c_branch2c (Conv2D) (None, 45, 45, 512) 66048 activation_18[0][0]
__________________________________________________________________________________________________
bn3c_branch2c (BatchNormalizati (None, 45, 45, 512) 2048 res3c_branch2c[0][0]
__________________________________________________________________________________________________
add_6 (Add) (None, 45, 45, 512) 0 bn3c_branch2c[0][0]
activation_16[0][0]
__________________________________________________________________________________________________
activation_19 (Activation) (None, 45, 45, 512) 0 add_6[0][0]
__________________________________________________________________________________________________
res3d_branch2a (Conv2D) (None, 45, 45, 128) 65664 activation_19[0][0]
__________________________________________________________________________________________________
bn3d_branch2a (BatchNormalizati (None, 45, 45, 128) 512 res3d_branch2a[0][0]
__________________________________________________________________________________________________
activation_20 (Activation) (None, 45, 45, 128) 0 bn3d_branch2a[0][0]
__________________________________________________________________________________________________
res3d_branch2b (Conv2D) (None, 45, 45, 128) 147584 activation_20[0][0]
__________________________________________________________________________________________________
bn3d_branch2b (BatchNormalizati (None, 45, 45, 128) 512 res3d_branch2b[0][0]
__________________________________________________________________________________________________
activation_21 (Activation) (None, 45, 45, 128) 0 bn3d_branch2b[0][0]
__________________________________________________________________________________________________
res3d_branch2c (Conv2D) (None, 45, 45, 512) 66048 activation_21[0][0]
__________________________________________________________________________________________________
bn3d_branch2c (BatchNormalizati (None, 45, 45, 512) 2048 res3d_branch2c[0][0]
__________________________________________________________________________________________________
add_7 (Add) (None, 45, 45, 512) 0 bn3d_branch2c[0][0]
activation_19[0][0]
__________________________________________________________________________________________________
activation_22 (Activation) (None, 45, 45, 512) 0 add_7[0][0]
__________________________________________________________________________________________________
res4a_branch2a (Conv2D) (None, 23, 23, 256) 131328 activation_22[0][0]
__________________________________________________________________________________________________
bn4a_branch2a (BatchNormalizati (None, 23, 23, 256) 1024 res4a_branch2a[0][0]
__________________________________________________________________________________________________
activation_23 (Activation) (None, 23, 23, 256) 0 bn4a_branch2a[0][0]
__________________________________________________________________________________________________
res4a_branch2b (Conv2D) (None, 23, 23, 256) 590080 activation_23[0][0]
__________________________________________________________________________________________________
bn4a_branch2b (BatchNormalizati (None, 23, 23, 256) 1024 res4a_branch2b[0][0]
__________________________________________________________________________________________________
activation_24 (Activation) (None, 23, 23, 256) 0 bn4a_branch2b[0][0]
__________________________________________________________________________________________________
res4a_branch2c (Conv2D) (None, 23, 23, 1024) 263168 activation_24[0][0]
__________________________________________________________________________________________________
res4a_branch1 (Conv2D) (None, 23, 23, 1024) 525312 activation_22[0][0]
__________________________________________________________________________________________________
bn4a_branch2c (BatchNormalizati (None, 23, 23, 1024) 4096 res4a_branch2c[0][0]
__________________________________________________________________________________________________
bn4a_branch1 (BatchNormalizatio (None, 23, 23, 1024) 4096 res4a_branch1[0][0]
__________________________________________________________________________________________________
add_8 (Add) (None, 23, 23, 1024) 0 bn4a_branch2c[0][0]
bn4a_branch1[0][0]
__________________________________________________________________________________________________
activation_25 (Activation) (None, 23, 23, 1024) 0 add_8[0][0]
__________________________________________________________________________________________________
res4b_branch2a (Conv2D) (None, 23, 23, 256) 262400 activation_25[0][0]
__________________________________________________________________________________________________
bn4b_branch2a (BatchNormalizati (None, 23, 23, 256) 1024 res4b_branch2a[0][0]
__________________________________________________________________________________________________
activation_26 (Activation) (None, 23, 23, 256) 0 bn4b_branch2a[0][0]
__________________________________________________________________________________________________
res4b_branch2b (Conv2D) (None, 23, 23, 256) 590080 activation_26[0][0]
__________________________________________________________________________________________________
bn4b_branch2b (BatchNormalizati (None, 23, 23, 256) 1024 res4b_branch2b[0][0]
__________________________________________________________________________________________________
activation_27 (Activation) (None, 23, 23, 256) 0 bn4b_branch2b[0][0]
__________________________________________________________________________________________________
res4b_branch2c (Conv2D) (None, 23, 23, 1024) 263168 activation_27[0][0]
__________________________________________________________________________________________________
bn4b_branch2c (BatchNormalizati (None, 23, 23, 1024) 4096 res4b_branch2c[0][0]
__________________________________________________________________________________________________
add_9 (Add) (None, 23, 23, 1024) 0 bn4b_branch2c[0][0]
activation_25[0][0]
__________________________________________________________________________________________________
activation_28 (Activation) (None, 23, 23, 1024) 0 add_9[0][0]
__________________________________________________________________________________________________
res4c_branch2a (Conv2D) (None, 23, 23, 256) 262400 activation_28[0][0]
__________________________________________________________________________________________________
bn4c_branch2a (BatchNormalizati (None, 23, 23, 256) 1024 res4c_branch2a[0][0]
__________________________________________________________________________________________________
activation_29 (Activation) (None, 23, 23, 256) 0 bn4c_branch2a[0][0]
__________________________________________________________________________________________________
res4c_branch2b (Conv2D) (None, 23, 23, 256) 590080 activation_29[0][0]
__________________________________________________________________________________________________
bn4c_branch2b (BatchNormalizati (None, 23, 23, 256) 1024 res4c_branch2b[0][0]
__________________________________________________________________________________________________
activation_30 (Activation) (None, 23, 23, 256) 0 bn4c_branch2b[0][0]
__________________________________________________________________________________________________
res4c_branch2c (Conv2D) (None, 23, 23, 1024) 263168 activation_30[0][0]
__________________________________________________________________________________________________
bn4c_branch2c (BatchNormalizati (None, 23, 23, 1024) 4096 res4c_branch2c[0][0]
__________________________________________________________________________________________________
add_10 (Add) (None, 23, 23, 1024) 0 bn4c_branch2c[0][0]
activation_28[0][0]
__________________________________________________________________________________________________
activation_31 (Activation) (None, 23, 23, 1024) 0 add_10[0][0]
__________________________________________________________________________________________________
res4d_branch2a (Conv2D) (None, 23, 23, 256) 262400 activation_31[0][0]
__________________________________________________________________________________________________
bn4d_branch2a (BatchNormalizati (None, 23, 23, 256) 1024 res4d_branch2a[0][0]
__________________________________________________________________________________________________
activation_32 (Activation) (None, 23, 23, 256) 0 bn4d_branch2a[0][0]
__________________________________________________________________________________________________
res4d_branch2b (Conv2D) (None, 23, 23, 256) 590080 activation_32[0][0]
__________________________________________________________________________________________________
bn4d_branch2b (BatchNormalizati (None, 23, 23, 256) 1024 res4d_branch2b[0][0]
__________________________________________________________________________________________________
activation_33 (Activation) (None, 23, 23, 256) 0 bn4d_branch2b[0][0]
__________________________________________________________________________________________________
res4d_branch2c (Conv2D) (None, 23, 23, 1024) 263168 activation_33[0][0]
__________________________________________________________________________________________________
bn4d_branch2c (BatchNormalizati (None, 23, 23, 1024) 4096 res4d_branch2c[0][0]
__________________________________________________________________________________________________
add_11 (Add) (None, 23, 23, 1024) 0 bn4d_branch2c[0][0]
activation_31[0][0]
__________________________________________________________________________________________________
activation_34 (Activation) (None, 23, 23, 1024) 0 add_11[0][0]
__________________________________________________________________________________________________
res4e_branch2a (Conv2D) (None, 23, 23, 256) 262400 activation_34[0][0]
__________________________________________________________________________________________________
bn4e_branch2a (BatchNormalizati (None, 23, 23, 256) 1024 res4e_branch2a[0][0]
__________________________________________________________________________________________________
activation_35 (Activation) (None, 23, 23, 256) 0 bn4e_branch2a[0][0]
__________________________________________________________________________________________________
res4e_branch2b (Conv2D) (None, 23, 23, 256) 590080 activation_35[0][0]
__________________________________________________________________________________________________
bn4e_branch2b (BatchNormalizati (None, 23, 23, 256) 1024 res4e_branch2b[0][0]
__________________________________________________________________________________________________
activation_36 (Activation) (None, 23, 23, 256) 0 bn4e_branch2b[0][0]
__________________________________________________________________________________________________
res4e_branch2c (Conv2D) (None, 23, 23, 1024) 263168 activation_36[0][0]
__________________________________________________________________________________________________
bn4e_branch2c (BatchNormalizati (None, 23, 23, 1024) 4096 res4e_branch2c[0][0]
__________________________________________________________________________________________________
add_12 (Add) (None, 23, 23, 1024) 0 bn4e_branch2c[0][0]
activation_34[0][0]
__________________________________________________________________________________________________
activation_37 (Activation) (None, 23, 23, 1024) 0 add_12[0][0]
__________________________________________________________________________________________________
res4f_branch2a (Conv2D) (None, 23, 23, 256) 262400 activation_37[0][0]
__________________________________________________________________________________________________
bn4f_branch2a (BatchNormalizati (None, 23, 23, 256) 1024 res4f_branch2a[0][0]
__________________________________________________________________________________________________
activation_38 (Activation) (None, 23, 23, 256) 0 bn4f_branch2a[0][0]
__________________________________________________________________________________________________
res4f_branch2b (Conv2D) (None, 23, 23, 256) 590080 activation_38[0][0]
__________________________________________________________________________________________________
bn4f_branch2b (BatchNormalizati (None, 23, 23, 256) 1024 res4f_branch2b[0][0]
__________________________________________________________________________________________________
activation_39 (Activation) (None, 23, 23, 256) 0 bn4f_branch2b[0][0]
__________________________________________________________________________________________________
res4f_branch2c (Conv2D) (None, 23, 23, 1024) 263168 activation_39[0][0]
__________________________________________________________________________________________________
bn4f_branch2c (BatchNormalizati (None, 23, 23, 1024) 4096 res4f_branch2c[0][0]
__________________________________________________________________________________________________
add_13 (Add) (None, 23, 23, 1024) 0 bn4f_branch2c[0][0]
activation_37[0][0]
__________________________________________________________________________________________________
activation_40 (Activation) (None, 23, 23, 1024) 0 add_13[0][0]
__________________________________________________________________________________________________
res5a_branch2a (Conv2D) (None, 12, 12, 512) 524800 activation_40[0][0]
__________________________________________________________________________________________________
bn5a_branch2a (BatchNormalizati (None, 12, 12, 512) 2048 res5a_branch2a[0][0]
__________________________________________________________________________________________________
activation_41 (Activation) (None, 12, 12, 512) 0 bn5a_branch2a[0][0]
__________________________________________________________________________________________________
res5a_branch2b (Conv2D) (None, 12, 12, 512) 2359808 activation_41[0][0]
__________________________________________________________________________________________________
bn5a_branch2b (BatchNormalizati (None, 12, 12, 512) 2048 res5a_branch2b[0][0]
__________________________________________________________________________________________________
activation_42 (Activation) (None, 12, 12, 512) 0 bn5a_branch2b[0][0]
__________________________________________________________________________________________________
res5a_branch2c (Conv2D) (None, 12, 12, 2048) 1050624 activation_42[0][0]
__________________________________________________________________________________________________
res5a_branch1 (Conv2D) (None, 12, 12, 2048) 2099200 activation_40[0][0]
__________________________________________________________________________________________________
bn5a_branch2c (BatchNormalizati (None, 12, 12, 2048) 8192 res5a_branch2c[0][0]
__________________________________________________________________________________________________
bn5a_branch1 (BatchNormalizatio (None, 12, 12, 2048) 8192 res5a_branch1[0][0]
__________________________________________________________________________________________________
add_14 (Add) (None, 12, 12, 2048) 0 bn5a_branch2c[0][0]
bn5a_branch1[0][0]
__________________________________________________________________________________________________
activation_43 (Activation) (None, 12, 12, 2048) 0 add_14[0][0]
__________________________________________________________________________________________________
res5b_branch2a (Conv2D) (None, 12, 12, 512) 1049088 activation_43[0][0]
__________________________________________________________________________________________________
bn5b_branch2a (BatchNormalizati (None, 12, 12, 512) 2048 res5b_branch2a[0][0]
__________________________________________________________________________________________________
activation_44 (Activation) (None, 12, 12, 512) 0 bn5b_branch2a[0][0]
__________________________________________________________________________________________________
res5b_branch2b (Conv2D) (None, 12, 12, 512) 2359808 activation_44[0][0]
__________________________________________________________________________________________________
bn5b_branch2b (BatchNormalizati (None, 12, 12, 512) 2048 res5b_branch2b[0][0]
__________________________________________________________________________________________________
activation_45 (Activation) (None, 12, 12, 512) 0 bn5b_branch2b[0][0]
__________________________________________________________________________________________________
res5b_branch2c (Conv2D) (None, 12, 12, 2048) 1050624 activation_45[0][0]
__________________________________________________________________________________________________
bn5b_branch2c (BatchNormalizati (None, 12, 12, 2048) 8192 res5b_branch2c[0][0]
__________________________________________________________________________________________________
add_15 (Add) (None, 12, 12, 2048) 0 bn5b_branch2c[0][0]
activation_43[0][0]
__________________________________________________________________________________________________
activation_46 (Activation) (None, 12, 12, 2048) 0 add_15[0][0]
__________________________________________________________________________________________________
res5c_branch2a (Conv2D) (None, 12, 12, 512) 1049088 activation_46[0][0]
__________________________________________________________________________________________________
bn5c_branch2a (BatchNormalizati (None, 12, 12, 512) 2048 res5c_branch2a[0][0]
__________________________________________________________________________________________________
activation_47 (Activation) (None, 12, 12, 512) 0 bn5c_branch2a[0][0]
__________________________________________________________________________________________________
res5c_branch2b (Conv2D) (None, 12, 12, 512) 2359808 activation_47[0][0]
__________________________________________________________________________________________________
bn5c_branch2b (BatchNormalizati (None, 12, 12, 512) 2048 res5c_branch2b[0][0]
__________________________________________________________________________________________________
activation_48 (Activation) (None, 12, 12, 512) 0 bn5c_branch2b[0][0]
__________________________________________________________________________________________________
res5c_branch2c (Conv2D) (None, 12, 12, 2048) 1050624 activation_48[0][0]
__________________________________________________________________________________________________
bn5c_branch2c (BatchNormalizati (None, 12, 12, 2048) 8192 res5c_branch2c[0][0]
__________________________________________________________________________________________________
add_16 (Add) (None, 12, 12, 2048) 0 bn5c_branch2c[0][0]
activation_46[0][0]
__________________________________________________________________________________________________
activation_49 (Activation) (None, 12, 12, 2048) 0 add_16[0][0]
__________________________________________________________________________________________________
average_pool (GlobalAveragePool (None, 2048) 0 activation_49[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 168) 344232 average_pool[0][0]
==================================================================================================
有了这些,我们便可以开始模型微调之旅。
base_model.trainable = False
这是为了保证,卷积基中庞大的参数不会因为新的数据的传入而发生变化。这种方法虽然简单但是很难保证预训练网络的参数正好是最符合我们新传入的数据的,因此,我们可以一点一点的解冻顶部的层,让它们也参与到模型的训练中去。
解冻代码
ResNet50_LAYERS_TO_FREEZE = 'res5a_branch2c'
# 微调模型
base_model.trainable = True
set_trainable = False
for layer in base_model.layers:
if layer.name == ResNet50_LAYERS_TO_FREEZE:
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
最终,经过多次试验,我们最终选择解冻至res5a_branch2c层。
训练模型后,验证集的准确度达到86.16%
optimizer=optimizers.RMSprop(lr=1e-3)
训练模型后,验证集的准确度达到87.59%
def preprocess(image):
mean = [R_MEAN, G_MEAN, B_MEAN]
image[..., 0] -= mean[0]
image[..., 1] -= mean[1]
image[..., 2] -= mean[2]
return image
# 可以对训练集进行数据增强处理
train_datagen = ImageDataGenerator(preprocessing_function=preprocess,
# rotation_range=20,
# width_shift_range=0.1,
# height_shift_range=0.1,
# zoom_range=0.1,
horizontal_flip=True,
# fill_mode='constant'
)
# 测试集不许动,去均值中心化完了之后不许动
validation_datagen = ImageDataGenerator(preprocessing_function=preprocess)
# 利用python生成器不断的生成训练样本
train_generator = train_datagen.flow_from_directory(
# 训练集地址
train_dir,
# 缩放到356*356
target_size=(img_width, img_height),
# 每个批量包含batch_size个样本
batch_size=batch_size,
# 因为是单标签,多分类问题,最后损失函数要用catagorical_crossentropy,所以此处用catagorical
class_mode='categorical')
validation_generator = validation_datagen.flow_from_directory(
validation_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode='categorical')
# 回调函数用来监控val_loss,一旦超过3次迭代没有下降,降低学习率;
# 且在每轮过后保存当前模型权重,监控val_acc,如果没有改善,那么不需要覆盖模型文件
callback_list = [
keras.callbacks.ReduceLROnPlateau(
monitor='val_loss',
factor=0.1,
patience=3,
verbose=1,
),
ModelCheckpoint(
filepath='non_Bird_' + str(img_width) + '_ResNet50.h5',
monitor='val_acc',
save_best_only=True,
verbose=1
)
]
# 训练模型
history = model.fit_generator(
train_generator,
steps_per_epoch=num_train_sample // batch_size,
epochs=epochs,
validation_data=validation_generator,
validation_steps=num_validation_sample // batch_size,
callbacks=callback_list,
verbose=1)
训练模型后,验证集的准确度达到88.97%
还差1%!!!!
到底哪里还有问题???
K.set_learning_phase(0)
base_model = ResNet50(weights='imagenet',
include_top=False,
input_shape=(img_width, img_height, 3))
# 微调模型
base_model.trainable = True
set_trainable = False
for layer in base_model.layers:
if layer.name == ResNet50_LAYERS_TO_FREEZE:
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
K.set_learning_phase(1)
x = base_model.output
x = GlobalAveragePooling2D(name='average_pool')(x)
predictions = Dense(class_num, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
# 编译模型
model.compile(loss='categorical_crossentropy',
optimizer=optimizers.RMSprop(lr=1e-3),
metrics=['acc'])
训练模型后,验证集的准确度达到91.71%
任务完成
多亏了这位前辈的指导
https://blog.csdn.net/aojue1109/article/details/88181927
一些关于优化器的选择
https://blog.csdn.net/u013249853/article/details/89148990
一些关于val_acc不能正常提升的经验之谈
https://blog.csdn.net/u013249853/article/details/89393982
迁徙学习的入门和小技巧
https://blog.csdn.net/TTdreamloong/article/details/80380736
本次任务的所有代码
https://github.com/chenhao-ZZU/non_Bird