VGG16微调

Question

我正在尝试微调 VGG16。 但有时我得到一个恒定的验证准确度，有时它固定为 0.0，有时它固定为 1.0，在测试准确度上也是如此。 也碰巧训练是不断的。

这些是一些例子：

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

RMSprop，bs：64，lr：0.001

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

SDG，bs：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

SDG，bs：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，但训练准确度通常会增加，验证准确度会以非常混乱的方式表现，从一个时期到另一个时期变化很大。 请问您有什么建议吗？

Answer 1

训练准确率提高和验证准确率波动非常明显：model 正在尝试学习如何“记忆”训练集，因此我们有验证集以防止它过度拟合。

从结果中也可以看出，您的 model 似乎学得这么低。 尝试调整超参数。

我注意到的一件事（但无法确认）：如果您使用迁移学习并且学习率如此之大，它可能会破坏预训练的 model（在这里，VGG）的所有辛勤工作。 我从 Google 的笔记本中找到了这个学习率调度程序，尝试使用它：

start_lr = 0.00001
min_lr = 0.00001
max_lr = 0.00005 * tpu_strategy.num_replicas_in_sync
rampup_epochs = 5
sustain_epochs = 0
exp_decay = .8

def lrfn(epoch):
  if epoch < rampup_epochs:
    return (max_lr - start_lr)/rampup_epochs * epoch + start_lr
  elif epoch < rampup_epochs + sustain_epochs:
    return max_lr
  else:
    return (max_lr - min_lr) * exp_decay**(epoch-rampup_epochs-sustain_epochs) + min_lr
    
lr_callback = tf.keras.callbacks.LearningRateScheduler(lambda epoch: lrfn(epoch), verbose=True)
...
model.fit(..., callbacks=[lr_callback])

这个想法是在第一个时期设置一个低学习率，然后增加它并慢慢降低它。