使用深度学习检测疟疾

我们不会对我们的验证数据集应用任何转换（除非是调整大小，因为这是必须的），因为我们将使用它评估每个纪元的模型性能。对于在传输学习环境中的图像增强的详细解释，请随时查看我上面引用的文章。让我们从一批图像增强转换中查看一些样本结果。

img_id = 0
sample_generator = train_datagen.flow(train_data[img_id:img_id+1], train_labels[img_id:img_id+1],
                                      batch_size=1)
sample = [next(sample_generator) for i in range(0,5)]
fig, ax = plt.subplots(1,5, figsize=(16, 6))
print('Labels:', [item[1][0] for item in sample])
l = [ax[i].imshow(sample[i][0][0]) for i in range(0,5)]

Sample augmented images

你可以清晰的看到与之前的输出的我们图像的轻微变化。我们现在构建我们的学习模型，确保 VGG-19 模型的最后两块是可以训练的。

vgg = tf.keras.applications.vgg19.VGG19(include_top=False, weights='imagenet', 
                                        input_shape=INPUT_SHAPE)
# Freeze the layers
vgg.trainable = True
 
set_trainable = False
for layer in vgg.layers:
    if layer.name in ['block5_conv1', 'block4_conv1']:
        set_trainable = True
    if set_trainable:
        layer.trainable = True
    else:
        layer.trainable = False
    
base_vgg = vgg
base_out = base_vgg.output
pool_out = tf.keras.layers.Flatten()(base_out)
hidden1 = tf.keras.layers.Dense(512, activation='relu')(pool_out)
drop1 = tf.keras.layers.Dropout(rate=0.3)(hidden1)
hidden2 = tf.keras.layers.Dense(512, activation='relu')(drop1)
drop2 = tf.keras.layers.Dropout(rate=0.3)(hidden2)
 
out = tf.keras.layers.Dense(1, activation='sigmoid')(drop2)
 
model = tf.keras.Model(inputs=base_vgg.input, outputs=out)
model.compile(optimizer=tf.keras.optimizers.RMSprop(lr=1e-5),
                loss='binary_crossentropy',
                metrics=['accuracy'])
 
print("Total Layers:", len(model.layers))
print("Total trainable layers:", sum([1 for l in model.layers if l.trainable]))
 
 
# Output
Total Layers: 28
Total trainable layers: 16

在我们的模型中我们降低了学习率，因为我们不想在微调的时候对预训练的层做大的位权更新。模型的训练过程可能有轻微的不同，因为我们使用了数据生成器，因此我们将应用 fit_generator(...) 函数。

tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.5,
                              patience=2, min_lr=0.000001)
 
callbacks = [reduce_lr, tensorboard_callback]
train_steps_per_epoch = train_generator.n // train_generator.batch_size
val_steps_per_epoch = val_generator.n // val_generator.batch_size
history = model.fit_generator(train_generator, steps_per_epoch=train_steps_per_epoch, epochs=EPOCHS,
                              validation_data=val_generator, validation_steps=val_steps_per_epoch, 
                              verbose=1)
 
 
# Output
Epoch 1/25
271/271 [====] - 133s 489ms/step - loss: 0.2267 - accuracy: 0.9117 - val_loss: 0.1414 - val_accuracy: 0.9531
Epoch 2/25
271/271 [====] - 129s 475ms/step - loss: 0.1399 - accuracy: 0.9552 - val_loss: 0.1292 - val_accuracy: 0.9589
...
...
Epoch 24/25
271/271 [====] - 128s 473ms/step - loss: 0.0815 - accuracy: 0.9727 - val_loss: 0.1466 - val_accuracy: 0.9682
Epoch 25/25
271/271 [====] - 128s 473ms/step - loss: 0.0792 - accuracy: 0.9729 - val_loss: 0.1127 - val_accuracy: 0.9641

（编辑：威海站长网）

【声明】本站内容均来自网络，其相关言论仅代表作者个人观点，不代表本站立场。若无意侵犯到您的权利，请及时与联系站长删除相关内容!