Question

我正在使用tensorflow后端在keras中使用以下模型架构训练CNN，以解决二进制分类问题。我将比例为70：25：5的约41k图像分为训练集，验证集和测试集，得出训练集中的29k张图像，验证中的10k张和测试集中的2k张。

没有类别失衡，pos和neg类别中大约有2万个样本。

model = Sequential()
model.add(Conv2D(32, (7, 7), padding = 'same', input_shape=input_shape))
model.add(Conv2D(32, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.5))

model.add(Conv2D(32, (7, 7), padding = 'same'))
model.add(Conv2D(32, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.6))

model.add(Conv2D(32, (7, 7), padding = 'same'))
model.add(Conv2D(32, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.6))

model.add(Conv2D(64, (7, 7), padding = 'same'))
model.add(Conv2D(64, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.6))

model.add(Conv2D(64, (7, 7), padding = 'same'))
model.add(Conv2D(64, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.6))

model.add(Conv2D(64, (7, 7), padding = 'same'))
model.add(Conv2D(64, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.6))

model.add(Conv2D(128, (7, 7), padding = 'same'))
model.add(Conv2D(128, (7, 7), padding = 'same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.6))

model.add(Flatten())
model.add(Dense(512))
model.add(Activation('relu'))

model.add(Dense(512))
model.add(Activation('relu'))

model.add(Dense(1))
model.add(Activation('sigmoid'))

model.compile(loss='binary_crossentropy',
              optimizer=optimizers.Adam(lr=3e-5),
              metrics=['accuracy'])

checkpoint = ModelCheckpoint(filepath='checkpointORCA_adam-{epoch:02d}-{val_loss:.2f}.h5', monitor='val_loss', verbose=0, save_best_only=True)

reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5,
                              patience=20, min_lr=1e-8)

train_datagen = ImageDataGenerator(rescale=1. / 255,
        shear_range=0.2,
        zoom_range=0.2)

# this is the augmentation configuration we will use for testing:
# only rescaling
test_datagen = ImageDataGenerator(rescale=1. / 255)

# Change the batchsize according to your system RAM
train_batchsize = 32 # changed them to 64 and 128 respectively, but same 
                      results
val_batchsize = 32

train_generator = train_datagen.flow_from_directory(
    train_data_path,
    target_size=(img_width, img_height),
    batch_size=train_batchsize,
    class_mode='binary',
shuffle=True)

# train_generator.reset()
# validation_generator.reset()
validation_generator = test_datagen.flow_from_directory(
    validation_data_path,
    target_size=(img_width, img_height),
    batch_size=val_batchsize,
    class_mode='binary',
shuffle=False)

# validation_generator.reset()

history = model.fit_generator(
    train_generator,
    steps_per_epoch=nb_train_samples // batch_size,
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=nb_validation_samples // batch_size,
    callbacks=[checkpoint, reduce_lr])

这些是培训进度的时期，其中验证准确性以线性方式波动。它先变高然后变低几乎相同的数量。可能是什么原因？

我几乎已经检查出所有其他答案，并且我的数据已经标准化，可以在训练集中适当地进行混洗，lr很小，并且可以很好地检查到其他处于类似问题领域的研究人员所获得的成功。

Found 29124 images belonging to 2 classes.
Found 10401 images belonging to 2 classes.
Epoch 1/60
910/910 [==============================] - 530s 582ms/step - loss: 0.6105 - acc: 0.6161 - val_loss: 0.2298 - val_acc: 0.9548
Epoch 2/60
910/910 [==============================] - 520s 571ms/step - loss: 0.3590 - acc: 0.8480 - val_loss: 0.8340 - val_acc: 0.6604
Epoch 3/60
910/910 [==============================] - 520s 571ms/step - loss: 0.3160 - acc: 0.8695 - val_loss: 0.0983 - val_acc: 0.9558
Epoch 4/60
910/910 [==============================] - 528s 580ms/step - loss: 0.2925 - acc: 0.8830 - val_loss: 0.5063 - val_acc: 0.8385
Epoch 5/60
910/910 [==============================] - 529s 581ms/step - loss: 0.2718 - acc: 0.8895 - val_loss: 0.0541 - val_acc: 0.9745
Epoch 6/60
910/910 [==============================] - 530s 583ms/step - loss: 0.2523 - acc: 0.8982 - val_loss: 0.5849 - val_acc: 0.8060
Epoch 7/60
910/910 [==============================] - 528s 580ms/step - loss: 0.2368 - acc: 0.9076 - val_loss: 0.0682 - val_acc: 0.9695
Epoch 8/60
910/910 [==============================] - 529s 582ms/step - loss: 0.2168 - acc: 0.9160 - val_loss: 0.6503 - val_acc: 0.7660
Epoch 9/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1996 - acc: 0.9213 - val_loss: 0.0339 - val_acc: 0.9850
Epoch 10/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1896 - acc: 0.9258 - val_loss: 0.5710 - val_acc: 0.8033
Epoch 11/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1814 - acc: 0.9285 - val_loss: 0.0391 - val_acc: 0.9834
Epoch 12/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1715 - acc: 0.9342 - val_loss: 0.6787 - val_acc: 0.7792
Epoch 13/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1678 - acc: 0.9361 - val_loss: 0.0451 - val_acc: 0.9796
Epoch 14/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1683 - acc: 0.9356 - val_loss: 0.7874 - val_acc: 0.7306
Epoch 15/60
910/910 [==============================] - 528s 580ms/step - loss: 0.1618 - acc: 0.9387 - val_loss: 0.0483 - val_acc: 0.9761
Epoch 16/60
910/910 [==============================] - 528s 581ms/step - loss: 0.1569 - acc: 0.9398 - val_loss: 0.9105 - val_acc: 0.7060
Epoch 17/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1566 - acc: 0.9397 - val_loss: 0.0380 - val_acc: 0.9853
Epoch 18/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1506 - acc: 0.9416 - val_loss: 0.7649 - val_acc: 0.7435
Epoch 19/60
910/910 [==============================] - 527s 580ms/step - loss: 0.1497 - acc: 0.9429 - val_loss: 0.0507 - val_acc: 0.9778
Epoch 20/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1476 - acc: 0.9439 - val_loss: 0.7189 - val_acc: 0.7665
Epoch 21/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1426 - acc: 0.9447 - val_loss: 0.0377 - val_acc: 0.9873
Epoch 22/60
910/910 [==============================] - 528s 580ms/step - loss: 0.1407 - acc: 0.9463 - val_loss: 0.7066 - val_acc: 0.7817
Epoch 23/60
910/910 [==============================] - 526s 578ms/step - loss: 0.1427 - acc: 0.9444 - val_loss: 0.0376 - val_acc: 0.9877
Epoch 24/60
910/910 [==============================] - 528s 580ms/step - loss: 0.1373 - acc: 0.9467 - val_loss: 0.6619 - val_acc: 0.8023
Epoch 25/60
910/910 [==============================] - 528s 580ms/step - loss: 0.1362 - acc: 0.9466 - val_loss: 0.0457 - val_acc: 0.9844
Epoch 26/60
910/910 [==============================] - 529s 582ms/step - loss: 0.1350 - acc: 0.9474 - val_loss: 0.8683 - val_acc: 0.7046
Epoch 27/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1339 - acc: 0.9492 - val_loss: 0.0411 - val_acc: 0.9855
Epoch 28/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1339 - acc: 0.9499 - val_loss: 0.9552 - val_acc: 0.6762
Epoch 29/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1343 - acc: 0.9488 - val_loss: 0.0446 - val_acc: 0.9859
Epoch 30/60
910/910 [==============================] - 528s 580ms/step - loss: 0.1282 - acc: 0.9513 - val_loss: 0.8127 - val_acc: 0.7298
Epoch 31/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1286 - acc: 0.9504 - val_loss: 0.0484 - val_acc: 0.9857
Epoch 32/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1258 - acc: 0.9506 - val_loss: 0.5007 - val_acc: 0.8479
Epoch 33/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1301 - acc: 0.9495 - val_loss: 0.0467 - val_acc: 0.9859
Epoch 34/60
910/910 [==============================] - 529s 581ms/step - loss: 0.1253 - acc: 0.9516 - val_loss: 0.6061 - val_acc: 0.8056
Epoch 35/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1259 - acc: 0.9521 - val_loss: 0.0469 - val_acc: 0.9873
Epoch 36/60
910/910 [==============================] - 528s 580ms/step - loss: 0.1249 - acc: 0.9511 - val_loss: 0.8658 - val_acc: 0.7121
Epoch 37/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1206 - acc: 0.9548 - val_loss: 0.0459 - val_acc: 0.9869
Epoch 38/60
910/910 [==============================] - 527s 580ms/step - loss: 0.1229 - acc: 0.9512 - val_loss: 0.4516 - val_acc: 0.8646
Epoch 39/60
910/910 [==============================] - 527s 579ms/step - loss: 0.1206 - acc: 0.9528 - val_loss: 0.0469 - val_acc: 0.9861
Epoch 40/60

下图不是针对此问题的，而是与我所询问的情况类似的情况：

Answer 1

我尝试通过以下方法获得曲线：

将模型的容量减少到几层，几乎没有神经元。
降低学习率，并将我的批量大小从32增加到128。

这可能不是唯一的解决方案，但是我提出了一些简单的技巧来使事情变得正确。有用的链接：1和2

我还计划训练更多的时代。

验证准确度/损失随每个连续时期线性上升和下降

1 个答案: