【25】ReLU 家族評比個別使用ReLU LeakyReLU PReLU SELU 來訓練

2021 iThome 鐵人賽

DAY 25

AI & Data

30 天在 Colab 嘗試的 30 個影像分類訓練實驗系列第 25 篇

13th鐵人賽 tensorflow colab

Capillary J

2021-10-09 10:43:45

2129 瀏覽

分享至

Colab連結

今天要來實驗不同的 ReLU 家族的評比，挑戰者有

基本 ReLU
尾巴翹起來的 Leaky ReLU
帶有參數控制尾巴的 PReLU
比較新潮的 SELU

說真的，有那麼多 ReLU，但我自己實務上卻只使用過第一種基本款，所以我也蠻好奇這次實驗會有多大的對比。

而因為這次想跑多次的 epochs 來進行實驗，所以資料集的部分改用 beans 來實驗。

實驗一：基本 ReLU

def bottleneck(net, filters, out_ch, strides, shortcut=True, zero_pad=False):

    padding = 'valid' if zero_pad else 'same'
    shortcut_net = net

    net = tf.keras.layers.Conv2D(filters * 6, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.ReLU(max_value=6)(net)
    if zero_pad:
        net = tf.keras.layers.ZeroPadding2D(padding=((0, 1), (0, 1)))(net)

    net = tf.keras.layers.DepthwiseConv2D(3, strides=strides, use_bias=False, padding=padding)(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.ReLU(max_value=6)(net)

    net = tf.keras.layers.Conv2D(out_ch, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    if shortcut:
        net = tf.keras.layers.Add()([net, shortcut_net])

    return net

def get_mobilenetV2_relu(shape):
    input_node = tf.keras.layers.Input(shape=shape)

    net = tf.keras.layers.Conv2D(32, 3, (2, 2), use_bias=False, padding='same')(input_node)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.ReLU(max_value=6)(net)

    net = tf.keras.layers.DepthwiseConv2D(3, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.ReLU(max_value=6)(net)
    net = tf.keras.layers.Conv2D(16, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    net = bottleneck(net, 16, 24, (2, 2), shortcut=False, zero_pad=True)  # block_1
    net = bottleneck(net, 24, 24, (1, 1), shortcut=True)  # block_2

    net = bottleneck(net, 24, 32, (2, 2), shortcut=False, zero_pad=True)  # block_3
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_4
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_5

    net = bottleneck(net, 32, 64, (2, 2), shortcut=False, zero_pad=True)  # block_6
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_7
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_8
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_9

    net = bottleneck(net, 64, 96, (1, 1), shortcut=False)  # block_10
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_11
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_12

    net = bottleneck(net, 96, 160, (2, 2), shortcut=False, zero_pad=True)  # block_13
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_14
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_15

    net = bottleneck(net, 160, 320, (1, 1), shortcut=False)  # block_16

    net = tf.keras.layers.Conv2D(1280, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.ReLU(max_value=6)(net)

    return input_node, net

input_node, net = get_mobilenetV2_relu((224,224,3))
net = tf.keras.layers.GlobalAveragePooling2D()(net)
net = tf.keras.layers.Dense(NUM_OF_CLASS)(net)

model = tf.keras.Model(inputs=[input_node], outputs=[net])

model.compile(
    optimizer=tf.keras.optimizers.SGD(LR),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()],
)

history = model.fit(
    ds_train,
    epochs=EPOCHS,
    validation_data=ds_test,
    verbose=True)

產出：

Epoch 96/100
loss: 0.0074 - sparse_categorical_accuracy: 0.9961 - val_loss: 0.9123 - val_sparse_categorical_accuracy: 0.8516

實驗二：Leaky ReLU

一般的 ReLU 遇到輸入為負值時，輸出會變成0，很容易讓大半的節點死亡，因此我們在輸入負值時，仍然給它一個 α 值，讓他微分時仍然可以更新權重，在這邊我們用預設的0.3。

def bottleneck(net, filters, out_ch, strides, shortcut=True, zero_pad=False):

    padding = 'valid' if zero_pad else 'same'
    shortcut_net = net

    net = tf.keras.layers.Conv2D(filters * 6, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.LeakyReLU()(net)
    if zero_pad:
        net = tf.keras.layers.ZeroPadding2D(padding=((0, 1), (0, 1)))(net)

    net = tf.keras.layers.DepthwiseConv2D(3, strides=strides, use_bias=False, padding=padding)(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.LeakyReLU()(net)

    net = tf.keras.layers.Conv2D(out_ch, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    if shortcut:
        net = tf.keras.layers.Add()([net, shortcut_net])

    return net

def get_mobilenetV2_leakyrelu(shape):
    input_node = tf.keras.layers.Input(shape=shape)

    net = tf.keras.layers.Conv2D(32, 3, (2, 2), use_bias=False, padding='same')(input_node)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.LeakyReLU()(net)

    net = tf.keras.layers.DepthwiseConv2D(3, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.LeakyReLU()(net)
    net = tf.keras.layers.Conv2D(16, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    net = bottleneck(net, 16, 24, (2, 2), shortcut=False, zero_pad=True)  # block_1
    net = bottleneck(net, 24, 24, (1, 1), shortcut=True)  # block_2

    net = bottleneck(net, 24, 32, (2, 2), shortcut=False, zero_pad=True)  # block_3
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_4
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_5

    net = bottleneck(net, 32, 64, (2, 2), shortcut=False, zero_pad=True)  # block_6
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_7
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_8
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_9

    net = bottleneck(net, 64, 96, (1, 1), shortcut=False)  # block_10
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_11
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_12

    net = bottleneck(net, 96, 160, (2, 2), shortcut=False, zero_pad=True)  # block_13
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_14
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_15

    net = bottleneck(net, 160, 320, (1, 1), shortcut=False)  # block_16

    net = tf.keras.layers.Conv2D(1280, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.LeakyReLU()(net)

    return input_node, net

input_node, net = get_mobilenetV2_leakyrelu((224,224,3))
net = tf.keras.layers.GlobalAveragePooling2D()(net)
net = tf.keras.layers.Dense(NUM_OF_CLASS)(net)

model = tf.keras.Model(inputs=[input_node], outputs=[net])

model.compile(
    optimizer=tf.keras.optimizers.SGD(LR),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()],
)

history = model.fit(
    ds_train,
    epochs=EPOCHS,
    validation_data=ds_test,
    verbose=True)

產出:

Epoch 96/100
loss: 0.0074 - sparse_categorical_accuracy: 0.9981 - val_loss: 1.2042 - val_sparse_categorical_accuracy: 0.7969

實驗三：PReLU

有了 LeakyReLU 後，就會有人覺得給予固定的 α 值似乎讓模型缺乏了彈性，因此不如讓這個值也變成模型學習的參數，所以就有了 PReLU，在這邊我們讓他的初始值為0.25。

def bottleneck(net, filters, out_ch, strides, shortcut=True, zero_pad=False):

    padding = 'valid' if zero_pad else 'same'
    shortcut_net = net

    net = tf.keras.layers.Conv2D(filters * 6, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25))(net)
    if zero_pad:
        net = tf.keras.layers.ZeroPadding2D(padding=((0, 1), (0, 1)))(net)

    net = tf.keras.layers.DepthwiseConv2D(3, strides=strides, use_bias=False, padding=padding)(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25))(net)

    net = tf.keras.layers.Conv2D(out_ch, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    if shortcut:
        net = tf.keras.layers.Add()([net, shortcut_net])

    return net

def get_mobilenetV2_prelu(shape):
    input_node = tf.keras.layers.Input(shape=shape)

    net = tf.keras.layers.Conv2D(32, 3, (2, 2), use_bias=False, padding='same')(input_node)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25))(net)

    net = tf.keras.layers.DepthwiseConv2D(3, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25))(net)
    net = tf.keras.layers.Conv2D(16, 1, use_bias=False, padding='same')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    net = bottleneck(net, 16, 24, (2, 2), shortcut=False, zero_pad=True)  # block_1
    net = bottleneck(net, 24, 24, (1, 1), shortcut=True)  # block_2

    net = bottleneck(net, 24, 32, (2, 2), shortcut=False, zero_pad=True)  # block_3
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_4
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_5

    net = bottleneck(net, 32, 64, (2, 2), shortcut=False, zero_pad=True)  # block_6
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_7
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_8
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_9

    net = bottleneck(net, 64, 96, (1, 1), shortcut=False)  # block_10
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_11
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_12

    net = bottleneck(net, 96, 160, (2, 2), shortcut=False, zero_pad=True)  # block_13
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_14
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_15

    net = bottleneck(net, 160, 320, (1, 1), shortcut=False)  # block_16

    net = tf.keras.layers.Conv2D(1280, 1, use_bias=False, padding='same',)(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25))(net)

    return input_node, net

input_node, net = get_mobilenetV2_prelu((224,224,3))
net = tf.keras.layers.GlobalAveragePooling2D()(net)
net = tf.keras.layers.Dense(NUM_OF_CLASS)(net)

model = tf.keras.Model(inputs=[input_node], outputs=[net])

model.compile(
    optimizer=tf.keras.optimizers.SGD(LR),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()],
)

history = model.fit(
    ds_train,
    epochs=EPOCHS,
    validation_data=ds_test,
    verbose=True)

產出:

Epoch 99/100
loss: 0.0119 - sparse_categorical_accuracy: 0.9961 - val_loss: 1.3487 - val_sparse_categorical_accuracy: 0.8125

實驗四：SELU

介紹 SELU 之前，要先知道他是 ELU 的一種，ELU在輸入小於0時，會是一個 exponential 形狀，所以在計算量上都會比較大，ELU 的公式如下：

來源

而 SELU 公式和 ELU 差不多，只是原作者找到了一個確切的 α 數值和一個固定的 λ 值來做縮放，而且在使用這個 act func 時，前面的權重初始化要使用 lecun_normal，如果有用過 dropout 則是要替換成另一種 AlphaDropout。

def bottleneck(net, filters, out_ch, strides, shortcut=True, zero_pad=False):

    padding = 'valid' if zero_pad else 'same'
    shortcut_net = net

    net = tf.keras.layers.Conv2D(filters * 6, 1, use_bias=False, padding='same', kernel_initializer='lecun_normal')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.activations.selu(net)
    if zero_pad:
        net = tf.keras.layers.ZeroPadding2D(padding=((0, 1), (0, 1)))(net)

    net = tf.keras.layers.DepthwiseConv2D(3, strides=strides, use_bias=False, padding=padding, kernel_initializer='lecun_normal')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.activations.selu(net)

    net = tf.keras.layers.Conv2D(out_ch, 1, use_bias=False, padding='same', kernel_initializer='lecun_normal')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    if shortcut:
        net = tf.keras.layers.Add()([net, shortcut_net])

    return net

def get_mobilenetV2_selu(shape):
    input_node = tf.keras.layers.Input(shape=shape)

    net = tf.keras.layers.Conv2D(32, 3, (2, 2), use_bias=False, padding='same', kernel_initializer='lecun_normal')(input_node)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.activations.selu(net)

    net = tf.keras.layers.DepthwiseConv2D(3, use_bias=False, padding='same', kernel_initializer='lecun_normal')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.activations.selu(net)
    net = tf.keras.layers.Conv2D(16, 1, use_bias=False, padding='same', kernel_initializer='lecun_normal')(net)
    net = tf.keras.layers.BatchNormalization()(net)

    net = bottleneck(net, 16, 24, (2, 2), shortcut=False, zero_pad=True)  # block_1
    net = bottleneck(net, 24, 24, (1, 1), shortcut=True)  # block_2

    net = bottleneck(net, 24, 32, (2, 2), shortcut=False, zero_pad=True)  # block_3
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_4
    net = bottleneck(net, 32, 32, (1, 1), shortcut=True)  # block_5

    net = bottleneck(net, 32, 64, (2, 2), shortcut=False, zero_pad=True)  # block_6
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_7
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_8
    net = bottleneck(net, 64, 64, (1, 1), shortcut=True)  # block_9

    net = bottleneck(net, 64, 96, (1, 1), shortcut=False)  # block_10
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_11
    net = bottleneck(net, 96, 96, (1, 1), shortcut=True)  # block_12

    net = bottleneck(net, 96, 160, (2, 2), shortcut=False, zero_pad=True)  # block_13
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_14
    net = bottleneck(net, 160, 160, (1, 1), shortcut=True)  # block_15

    net = bottleneck(net, 160, 320, (1, 1), shortcut=False)  # block_16

    net = tf.keras.layers.Conv2D(1280, 1, use_bias=False, padding='same', kernel_initializer='lecun_normal')(net)
    net = tf.keras.layers.BatchNormalization()(net)
    net = tf.keras.activations.selu(net)

    return input_node, net

input_node, net = get_mobilenetV2_selu((224,224,3))
net = tf.keras.layers.GlobalAveragePooling2D()(net)
net = tf.keras.layers.Dense(NUM_OF_CLASS)(net)

model = tf.keras.Model(inputs=[input_node], outputs=[net])

model.compile(
    optimizer=tf.keras.optimizers.SGD(LR),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.SparseCategoricalAccuracy()],
)

history = model.fit(
    ds_train,
    epochs=EPOCHS,
    validation_data=ds_test,
    verbose=True)

產出：

Epoch 85/100
loss: 0.0018 - sparse_categorical_accuracy: 1.0000 - val_loss: 0.6651 - val_sparse_categorical_accuracy: 0.8672

以上就是今天針對這四個不同的 ReLU 所做的實驗，大致上都有得到不錯的效果，今天如果你要處理的任務比較急或者是第一次訓練的話，或許還是基本款的 ReLU 就好，如果是相比 Kaggle 這種必須計較分數的比賽，那就很推薦嘗試一些新的手法來訓練。

【24】如果把 Dropout 放在 CNN 之後會發生什麼事

【26】你都把 Batch Normalization 放在 ReLU 前面還是後面

系列文

30 天在 Colab 嘗試的 30 個影像分類訓練實驗共 31 篇

RSS系列文訂閱系列文

8 人訂閱

完整目錄

直播研討會

{{ item.channelVendor }} {{ item.webinarstarted }} |

直播中

尚未有邦友留言

立即登入留言

參賽組數

1064 組

團體組數

40 組

累計文章數

22199 篇

完賽人數

602 人

15th鐵人賽 16th鐵人賽 13th鐵人賽 14th鐵人賽 12th鐵人賽 11th鐵人賽鐵人賽 2019鐵人賽 javascript 2018鐵人賽 python 2017鐵人賽 windows php c# windows server linux css react vue.js

IT邦幫忙

30 天在 Colab 嘗試的 30 個影像分類訓練實驗系列 第 25 篇