使用 ResNet50 进行二进制分类的恒定验证精度答案

【问题标题】：Constant validation accuracy in binary classification using ResNet50使用 ResNet50 进行二进制分类的恒定验证精度
【发布时间】：2020-09-18 11:39:52
【问题描述】：

我正在训练一个深度学习网络来对甲状腺结节进行分类（二进制分类为 0 或 1）。因此，我意识到我的模型具有恒定的验证准确性。我尝试了一切，更改了学习率，如果损失没有增加，则进行一些回调。所以，我需要帮助，一些想法来处理它。我会给你一个 10 个 epoch 的例子，但是，即使我改为 100 个 epoch，它仍然具有相同的行为。这是架构的网络代码：

callbacks = [
    EarlyStopping(patience=20, verbose=1),
    ReduceLROnPlateau(factor=0.1, patience=5, min_lr=0.0001, verbose=1),
]

epochs=10
model = Sequential()
model.add(tf.keras.applications.ResNet50( include_top=False, pooling='avg', weights='imagenet'))
model.add(BatchNormalization())
model.add(Dropout(0.4))
model.add(Dense(2, activation='sigmoid'))
model.layers[0].trainable = False
weights=class_weight.compute_class_weight('balanced', np.unique(y_train), y_train)
opt =Adam(learning_rate=0.001)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
tf.random.set_seed(2)
rede=model.fit(X_train, y_train_2d, epochs=epochs, batch_size=18, class_weight=weights, callbacks=callbacks,\
               validation_data=(X_valid, y_valid_2d))

Train on 360 samples, validate on 46 samples
Epoch 1/10
342/360 [===========================>..] - ETA: 1s - loss: 0.7582 - accuracy: 0.5673
Epoch 00001: val_loss did not improve from 0.16301
360/360 [==============================] - 44s 123ms/sample - loss: 0.7706 - accuracy: 0.5583 - val_loss: 0.5528 - val_accuracy: 0.8043
Epoch 2/10
342/360 [===========================>..] - ETA: 1s - loss: 0.5604 - accuracy: 0.7135
Epoch 00002: val_loss did not improve from 0.16301
360/360 [==============================] - 41s 114ms/sample - loss: 0.5525 - accuracy: 0.7222 - val_loss: 0.5047 - val_accuracy: 0.8043
Epoch 3/10
342/360 [===========================>..] - ETA: 1s - loss: 0.5056 - accuracy: 0.7485
Epoch 00003: val_loss did not improve from 0.16301
360/360 [==============================] - 41s 114ms/sample - loss: 0.4993 - accuracy: 0.7500 - val_loss: 0.4946 - val_accuracy: 0.8043
Epoch 4/10
342/360 [===========================>..] - ETA: 1s - loss: 0.5004 - accuracy: 0.7661
Epoch 00004: val_loss did not improve from 0.16301
360/360 [==============================] - 41s 114ms/sample - loss: 0.5019 - accuracy: 0.7667 - val_loss: 0.4942 - val_accuracy: 0.8043
Epoch 5/10
342/360 [===========================>..] - ETA: 1s - loss: 0.4967 - accuracy: 0.8070
Epoch 00005: val_loss did not improve from 0.16301
360/360 [==============================] - 41s 114ms/sample - loss: 0.4897 - accuracy: 0.8083 - val_loss: 0.4952 - val_accuracy: 0.8043
Epoch 6/10
342/360 [===========================>..] - ETA: 1s - loss: 0.4615 - accuracy: 0.7427
Epoch 00006: val_loss did not improve from 0.16301
360/360 [==============================] - 41s 114ms/sample - loss: 0.4750 - accuracy: 0.7361 - val_loss: 0.4985 - val_accuracy: 0.8043
Epoch 7/10
342/360 [===========================>..] - ETA: 1s - loss: 0.4119 - accuracy: 0.7924
Epoch 00007: val_loss did not improve from 0.16301
360/360 [==============================] - 40s 111ms/sample - loss: 0.4164 - accuracy: 0.7944 - val_loss: 0.4953 - val_accuracy: 0.8043
Epoch 8/10
342/360 [===========================>..] - ETA: 0s - loss: 0.4182 - accuracy: 0.8158
Epoch 00008: val_loss did not improve from 0.16301
360/360 [==============================] - 20s 55ms/sample - loss: 0.4155 - accuracy: 0.8194 - val_loss: 0.4957 - val_accuracy: 0.8043
Epoch 9/10
342/360 [===========================>..] - ETA: 0s - loss: 0.3892 - accuracy: 0.8304
Epoch 00009: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.

Epoch 00009: val_loss did not improve from 0.16301
360/360 [==============================] - 21s 57ms/sample - loss: 0.3853 - accuracy: 0.8278 - val_loss: 0.4997 - val_accuracy: 0.8043
Epoch 10/10
342/360 [===========================>..] - ETA: 0s - loss: 0.3395 - accuracy: 0.8333
Epoch 00010: val_loss did not improve from 0.16301
360/360 [==============================] - 21s 58ms/sample - loss: 0.3375 - accuracy: 0.8333 - val_loss: 0.5003 - val_accuracy: 0.8043

【问题讨论】：

标签： tensorflow keras deep-learning resnet

【解决方案1】：

对于二元分类，您应该将输出密集层更改为 1 个单元，并将损失更改为 binary_crossentropy。

model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer=opt, metrics=['accuracy'])

还要确保您的目标数组 (y_train,y_valid) 具有正确的形状和数据。它们应该包含二进制标签，0 或 1

【讨论】：

验证损失似乎不是严格恒定的——从 0.55 下降到 ~0.5 的高原是显着的。我认为如果 OP 对您描述的层进行更改，他可能会看到改进
好的，但是对于 y_train 和 y_valid，哪个形状是正确的？我更改了 y_train_2d 和 y_valid_2d，条目如下：良性：[0,1] 恶性：[1,0]
应该是良性的：[0]，恶性的：[1]，反之亦然

【解决方案2】：

@AlexM4 是对的，对于二进制分类，您应该使用带有 sigmoid 分类的单个输出单元。

还值得指出的是，for multi-class classification, you shouldn't use a sigmoid activation 而是 softmax，它产生了可能类的概率分布（sigmoid 理论上可以为所有类产生 1.0 的概率，这是没有意义的）。这可能会阻碍您的网络在此处的性能。

解决更大的问题（移动到单个输出单元）将消除解决我刚才描述的问题的需要，但在未来的实验中值得牢记。

【讨论】：

【解决方案3】：

我尝试将输出密集度更改为 1，而不是 2。准确度仍然具有相同的行为。

epochs=10
model = Sequential()
model.add(tf.keras.applications.ResNet50( include_top=False, pooling='avg', weights='imagenet'))
model.add(BatchNormalization())
model.add(Dropout(0.4))
model.add(Dense(1, activation='sigmoid'))
model.layers[0].trainable = False
weights=class_weight.compute_class_weight('balanced', np.unique(y_train), y_train)
opt =Adam(learning_rate=0.001)
model.compile(loss='binary_crossentropy', optimizer=opt, metrics=['accuracy'])
tf.random.set_seed(2)
rede=model.fit(X_train, y_train, epochs=epochs, batch_size=18, class_weight=weights, callbacks=callbacks,\
               validation_data=(X_valid, y_valid))
Train on 360 samples, validate on 46 samples
Epoch 1/10
360/360 [==============================] - 17s 48ms/sample - loss: 0.8778 - accuracy: 0.5639 - val_loss: 0.3967 - val_accuracy: 0.8696
Epoch 2/10
360/360 [==============================] - 9s 24ms/sample - loss: 0.6527 - accuracy: 0.7028 - val_loss: 0.3973 - val_accuracy: 0.8696
Epoch 3/10
360/360 [==============================] - 9s 26ms/sample - loss: 0.6054 - accuracy: 0.7222 - val_loss: 0.4171 - val_accuracy: 0.8696
Epoch 4/10
360/360 [==============================] - 10s 28ms/sample - loss: 0.5388 - accuracy: 0.7500 - val_loss: 0.4211 - val_accuracy: 0.8696
Epoch 5/10
360/360 [==============================] - 11s 30ms/sample - loss: 0.4830 - accuracy: 0.7694 - val_loss: 0.4367 - val_accuracy: 0.8696
Epoch 6/10
342/360 [===========================>..] - ETA: 0s - loss: 0.4986 - accuracy: 0.8070
Epoch 00006: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
360/360 [==============================] - 11s 29ms/sample - loss: 0.4952 - accuracy: 0.8056 - val_loss: 0.4576 - val_accuracy: 0.8696
Epoch 7/10
360/360 [==============================] - 11s 30ms/sample - loss: 0.4257 - accuracy: 0.8306 - val_loss: 0.4661 - val_accuracy: 0.8696
Epoch 8/10
360/360 [==============================] - 17s 46ms/sample - loss: 0.4054 - accuracy: 0.8250 - val_loss: 0.4722 - val_accuracy: 0.8696
Epoch 9/10
360/360 [==============================] - 17s 48ms/sample - loss: 0.4484 - accuracy: 0.7917 - val_loss: 0.4761 - val_accuracy: 0.8696
Epoch 10/10
360/360 [==============================] - 16s 44ms/sample - loss: 0.3844 - accuracy: 0.8472 - val_loss: 0.4785 - val_accuracy: 0.8696

【讨论】：

第一层是什么——你能打印一个模型摘要吗？也许您冻结的模型比您预期的要多。我从来没有在 Keras 中使用过 ResNet 骨干网，所以这只是我的一个疯狂的预感。