Python VGG16微调

我正在尝试微调VGG16。 但有时我得到的验证精度是恒定的，有时固定为0.0，有时固定为1.0，测试精度也是一样的。 碰巧的是，训练是持续不断的

以下是一些例子：

亚当，bs:64，lr:0.001

train_acc = [0.45828044, 0.4580425, 0.45812184, 0.45820114, 0.45820114, 0.45812184, 0.45820114, 0.45820114, 0.45820114, 0.4580425, 0.45820114, 0.45820114, 0.45812184, 0.45828044, 0.45820114, 0.45828044, 0.45812184, 0.45820114, 0.45812184, 0.45828044, 0.45820114, 0.45820114, 0.45812184, 0.45812184, 0.45820114, 0.45812184, 0.45828044, 0.45820114, 0.45828044, 0.45812184, 0.45820114, 0.45820114, 0.45812184, 0.45820114, 0.45820114, 0.45820114, 0.45828044, 0.45812184, 0.45828044, 0.4580425, 0.4580425, 0.45820114, 0.45820114, 0.45820114, 0.45828044, 0.45820114, 0.45812184, 0.45820114, 0.45820114, 0.45820114]
valid_acc = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
train_loss = [8.31718591143032, 8.35966631966799, 8.358442889857413, 8.357219463677575, 8.357219470939055, 8.358442853550015, 8.357219473359548, 8.357219434631658, 8.357219487882508, 8.359666328139717, 8.357219499984973, 8.357219495143987, 8.35844288017544, 8.355996039918232, 8.357219415267712, 8.355996025395273, 8.358442889857413, 8.357219521769412, 8.358442892277907, 8.355996052020698, 8.35721946609807, 8.357219415267712, 8.35844288017544, 8.358442885016427, 8.357219463677575, 8.358442882595934, 8.355996003610834, 8.357219458836589, 8.355996064123163, 8.357520040521766, 8.357219487882508, 8.357219480621028, 8.358442897118893, 8.357219495143987, 8.357219446734124, 8.35721945157511, 8.355996056861684, 8.358442911641852, 8.355996047179712, 8.359666311196264, 8.359666286991333, 8.35721946609807, 8.357219458836589, 8.35721944431363, 8.355996035077245, 8.357219453995603, 8.358442909221358, 8.357219439472644, 8.357219429790671, 8.357219461257083]
valid_loss = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

test_loss = 0.0
test_acc = 1.0

train_acc = [0.5421161, 0.54179883, 0.54179883, 0.54171956, 0.54171956, 0.5419575, 0.54187816, 0.54179883, 0.54187816, 0.5419575, 0.5419575]
valid_acc = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
train_loss = [6.990036433118249, 7.025707591003573, 7.025707559537161, 7.026923776278036, 7.02692376054483, 7.023275266444017, 7.024491474713166, 7.025707566798641, 7.024491443246754, 7.023275273705497, 7.0232752761259905]
valid_loss = [15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457]

test_loss = 15.33323860168457
test_acc = 0.0

RMSprop，bs:64，lr:0.001

train_acc = [0.45828044, 0.4580425, 0.45812184, 0.45820114, 0.45820114, 0.45812184, 0.45820114, 0.45820114, 0.45820114, 0.4580425, 0.45820114, 0.45820114, 0.45812184, 0.45828044, 0.45820114, 0.45828044, 0.45812184, 0.45820114, 0.45812184, 0.45828044, 0.45820114, 0.45820114, 0.45812184, 0.45812184, 0.45820114, 0.45812184, 0.45828044, 0.45820114, 0.45828044, 0.45812184, 0.45820114, 0.45820114, 0.45812184, 0.45820114, 0.45820114, 0.45820114, 0.45828044, 0.45812184, 0.45828044, 0.4580425, 0.4580425, 0.45820114, 0.45820114, 0.45820114, 0.45828044, 0.45820114, 0.45812184, 0.45820114, 0.45820114, 0.45820114]
valid_acc = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
train_loss = [8.31718591143032, 8.35966631966799, 8.358442889857413, 8.357219463677575, 8.357219470939055, 8.358442853550015, 8.357219473359548, 8.357219434631658, 8.357219487882508, 8.359666328139717, 8.357219499984973, 8.357219495143987, 8.35844288017544, 8.355996039918232, 8.357219415267712, 8.355996025395273, 8.358442889857413, 8.357219521769412, 8.358442892277907, 8.355996052020698, 8.35721946609807, 8.357219415267712, 8.35844288017544, 8.358442885016427, 8.357219463677575, 8.358442882595934, 8.355996003610834, 8.357219458836589, 8.355996064123163, 8.357520040521766, 8.357219487882508, 8.357219480621028, 8.358442897118893, 8.357219495143987, 8.357219446734124, 8.35721945157511, 8.355996056861684, 8.358442911641852, 8.355996047179712, 8.359666311196264, 8.359666286991333, 8.35721946609807, 8.357219458836589, 8.35721944431363, 8.355996035077245, 8.357219453995603, 8.358442909221358, 8.357219439472644, 8.357219429790671, 8.357219461257083]
valid_loss = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

test_loss = 0.0
test_acc = 1.0

train_acc = [0.5421161, 0.54179883, 0.54179883, 0.54171956, 0.54171956, 0.5419575, 0.54187816, 0.54179883, 0.54187816, 0.5419575, 0.5419575]
valid_acc = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
train_loss = [6.990036433118249, 7.025707591003573, 7.025707559537161, 7.026923776278036, 7.02692376054483, 7.023275266444017, 7.024491474713166, 7.025707566798641, 7.024491443246754, 7.023275273705497, 7.0232752761259905]
valid_loss = [15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457]

test_loss = 15.33323860168457
test_acc = 0.0

可持续发展目标，英国标准：64，lr:0.01，动量：0.2

train_acc = [0.5406091, 0.5419575, 0.54187816, 0.54179883, 0.54187816, 0.54187816, 0.54187816, 0.54187816, 0.54179883, 0.54171956, 0.54179883]
valid_acc = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
train_loss = [6.990036433118249, 7.025707591003573, 7.025707559537161, 7.026923776278036, 7.02692376054483, 7.023275266444017, 7.024491474713166, 7.025707566798641, 7.024491443246754, 7.023275273705497, 7.0232752761259905]
valid_loss = [15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457, 15.33323860168457]

test_loss = 15.33323860168457
test_acc = 0.0

可持续发展目标，英国标准：64，lr:0.01，动量：0.4

train_acc = [0.45740798, 0.45828044, 0.45820114, 0.45828044, 0.45820114, 0.4580425, 0.45820114, 0.45820114, 0.45820114, 0.45820114, 0.45820114]
valid_acc = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
train_loss = [8.329831461313413, 8.355996044759218, 8.357219475780042, 8.355996035077245, 8.357219502405467, 8.35966631603725, 8.357219461257083, 8.357219461257083, 8.357219456416097, 8.357219441893138, 8.357219478200534]
valid_loss = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

test_loss = 0.0
test_acc = 1.0

对于微调，我使用了以下顶层：

model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1, activation='sigmoid'))

你知道为什么会这样吗

---

无论如何，我仍在尝试训练网络，但训练精度通常会提高，验证精度表现得非常混乱，从一个时代到另一个时代变化很大。你有什么建议吗？

训练准确度增加，验证准确度波动非常明显：模型试图学习如何“记忆”训练集，因此我们有验证集以防止其过度拟合

同样从结果来看，你的模型似乎学得很低。尝试调整超参数

有一件事我注意到了（但无法证实）：如果你使用迁移学习，而且学习率如此之高，它可能会破坏预训练模型（在这里，VGG）的所有艰苦工作。我从谷歌的笔记本中找到了这个学习率计划程序，请尝试使用：

start\u lr=0.00001
最小值=0.00001
max_lr=0.00005*tpu_strategy.num_replications_in_sync
加速期=5
持续时间=0
exp_衰减=.8
def lrfn（历元）：
如果历元<加速\历元：
返回（最大持续时间-开始持续时间）/加速持续时间*持续时间+开始持续时间
elif历元<加速历元+持续历元：
返回最大值
其他：
返回（最大值-最小值）*经验衰减**（历元-加速-历元-维持历元）+最小值
lr_callback=tf.keras.callbacks.LearningRateScheduler（lambda epoch:lrfn（epoch），verbose=True）
...
model.fit（…，回调=[lr\u回调]）

我们的想法是在第一个阶段设置一个较低的学习率，然后提高，然后慢慢降低