前言

在第3篇教程里面，我們所編寫的CNN進(jìn)行分類的模型準(zhǔn)確度達(dá)到了80%。對(duì)于一個(gè)分類模型來(lái)說，80%的準(zhǔn)確率不算很低了。但是，在現(xiàn)有的情況下，我們應(yīng)該如何優(yōu)化這個(gè)模型呢？
在從零開始機(jī)器學(xué)習(xí)的系列里面，理論上的優(yōu)化模型可以修改超參數(shù)。同樣，在Keras的這個(gè)CNN程序中，我們可以指定其他的優(yōu)化器（這里用的是ADAM）。修改卷積核大小、步長(zhǎng)、修改激活函數(shù)的類型、加入/取消全連接層、修改每個(gè)層有多少神經(jīng)元也是可行的方法。
面對(duì)這么多可以修改的地方，修改現(xiàn)有的程序是必定需要做的。

修改現(xiàn)有的程序

首先確定要修改的內(nèi)容。全連接的層數(shù)、層的大小和卷積層的層數(shù)。
全連接層可以有0、1、2層；卷積層可以有1、2、3或更多層；然后就是每層的大小可以是32、64或128...
通過以下的代碼來(lái)確定這些排列組合可以得到多少模型：

dense_layers = [0, 1, 2]
layer_sizes = [32, 64, 128]
conv_layers = [1, 2, 3]

for dense_layer in dense_layers:
    for layer_size in layer_sizes:
        for conv_layer in conv_layers:
            NAME = "{}-conv-{}-nodes-{}-dense-{}".format(conv_layer, layer_size, dense_layer, int(time.time()))
            print(NAME)

運(yùn)行上述Python代碼，在控制臺(tái)的輸出如下：

1-conv-32-nodes-0-dense-1540975261
2-conv-32-nodes-0-dense-1540975261
3-conv-32-nodes-0-dense-1540975261
1-conv-64-nodes-0-dense-1540975261
2-conv-64-nodes-0-dense-1540975261
3-conv-64-nodes-0-dense-1540975261
1-conv-128-nodes-0-dense-1540975261
2-conv-128-nodes-0-dense-1540975261
3-conv-128-nodes-0-dense-1540975261
1-conv-32-nodes-1-dense-1540975261
2-conv-32-nodes-1-dense-1540975261
3-conv-32-nodes-1-dense-1540975261
1-conv-64-nodes-1-dense-1540975261
2-conv-64-nodes-1-dense-1540975261
3-conv-64-nodes-1-dense-1540975261
1-conv-128-nodes-1-dense-1540975261
2-conv-128-nodes-1-dense-1540975261
3-conv-128-nodes-1-dense-1540975261
1-conv-32-nodes-2-dense-1540975261
2-conv-32-nodes-2-dense-1540975261
3-conv-32-nodes-2-dense-1540975261
1-conv-64-nodes-2-dense-1540975261
2-conv-64-nodes-2-dense-1540975261
3-conv-64-nodes-2-dense-1540975261
1-conv-128-nodes-2-dense-1540975261
2-conv-128-nodes-2-dense-1540975261
3-conv-128-nodes-2-dense-1540975261

也就是說，通過排列組合可以得到27種模型。
修改第四講的代碼，將上述代碼加入程序，得到如下：

import time
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, MaxPooling2D
import pickle
from tensorflow.keras.callbacks import TensorBoard
import time

gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.33)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))

X = pickle.load(open("X.pickle","rb"))
y = pickle.load(open("y.pickle","rb"))

# normalize data image in grayscale is from 0-255
X = X/255.0

# train multiple models
dense_layers = [0, 1, 2]
layer_sizes = [32, 64, 128]
conv_layers = [1, 2, 3]

for dense_layer in dense_layers:
    for layer_size in layer_sizes:
        for conv_layer in conv_layers:
            model_name = "{}-conv-{}-nodes-{}-dense-{}".format(conv_layer, layer_size, dense_layer, int(time.time()))
            tensorboard = TensorBoard(log_dir='logs/{}'.format(model_name))
            print(model_name)
            model = Sequential()

            model.add(Conv2D(layer_size, (3, 3), input_shape = X.shape[1:]))
            model.add(Activation("relu"))
            model.add(MaxPooling2D(pool_size=(2,2)))

            for l in range(conv_layer-1):
                model.add(Conv2D(layer_size, (3, 3)))
                model.add(Activation("relu"))
                model.add(MaxPooling2D(pool_size=(2,2)))

            model.add(Flatten()) #Conv Layer是2D， DenseLayer是1D的 所以需要將ConvLayer壓平
            for l in range(dense_layer):
                model.add(Dense(layer_size))
                model.add(Activation("relu"))

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

            model.compile(loss="binary_crossentropy",
                         optimizer="adam",
                         metrics=["accuracy"]) # 可以使用categorical_crossentropy作為損失函數(shù)

            model.fit(X, y, batch_size =32, epochs=20, validation_split=0.1, callbacks=[tensorboard])

在控制臺(tái)運(yùn)行，一共需要訓(xùn)練27個(gè)模型，因此在訓(xùn)練之前，最好將logs的目錄下已經(jīng)存在的日志清空。這樣會(huì)方便我們觀察這一講中的不同參數(shù)對(duì)于模型的影響。

使用TensorBoard觀察模型

和上一講一樣，使用tensorboard --logdir=logs/打開TensorBoard。

TensorBoard觀察所有模型

放大折線圖