MobileNet V3について構造の説明と実装のメモ書きです。
ただし、論文すべてを見るわけでなく構造のところを中心に見ていきます。
勉強のメモ書き程度でありあまり正確に実装されていませんので、ご了承ください。
自分の実力不足で読み解けなくなってきています。難しいです。

以下の論文について実装を行っていきます。

タイトル：Searching for MobileNetV3

MobileNet V2は前回扱いました。

SE blockについてはSENetをご参照ください。

MobileNet V3

MobileNet V3ではMobileNet V2で導入したbottleneck構造に,SENetなどで導入したSE(Squeeze and Excitation) blockを組み合わせます。
また、非線形関数としてはh-swishを使用します。
SE blockでは非線形関数としてhard-sigmoidを使い、expansion layerのチャンネル数は入力の1/4とします。

h-swish

swish関数は以下のように定義されます。
$$
swish(x)=x・\sigma(x)
$$

$\sigma(x)$はsigmoid関数を表します。
このsigmoid関数を、hard-sigmoid関数
$$
hard sigmoid(x)=RELU_6(x+3)/6
$$
で置き換えたものをhard-swish関数と呼びます。
$$
hardswish(x)=x\frac{RELU_6(x+3)}{6}
$$

ここで、$RELU_6$はMobileNet V2でも出てきましたが最大値を6としたRELU関数のことです。

構造

構造を見ていきます。
exp_sizeはbottleneck blockの中間層でのチャンネル数、outは出力のチャンネル数、SEはSE blockの有無、NLは使用する非線形関数(HSならh-swish,REならRELU)、sはstrideを表します。
NBNはBatchNormalizationのことです。
2種類紹介されています。
1つ目はMobileNetV3 Large

2つ目はMobileNetV3 smallです。

学習

最適化手法はRMSpropでmomentumが0.9とします。学習率の初期値は0.1で3エポックごとに0.01の減衰率で減衰させます。weight decayは0.00001とします。
確率0.8のdropout層も挿入し、すべての畳み込み層の後にはBatchNormalizationを入れることとします。

実装

MobileNet V2にSE blockの追加をして、非線形関数を変更するだけです。
(シリーズものでは自分の実装を持っておくとこの辺が楽だなと思いました。)

keras

必要なライブラリのインポートをします。

import tensorflow.keras as keras
import tensorflow as tf
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Input, Conv2D, Activation, Dense, GlobalAveragePooling2D, BatchNormalization, DepthwiseConv2D, Add, Multiply, ReLU
from keras.layers.merge import concatenate
from tensorflow.keras import backend as K
from tensorflow.keras.optimizers import RMSprop
from tensorflow.keras.callbacks import LearningRateScheduler
from keras.datasets import cifar10
import numpy as np
import cv2

hard-sigmoidとhard-swishを実装します。

def hard_sigmoid(x):
    return ReLU(6.)(x + 3.) * (1. / 6.)

def hard_swish(x):
    return Multiply()([Activation(hard_sigmoid)(x), x])

SE Blockを実装します。

class SE_Block(Model):
    def __init__(self, in_channels, r=0.25):
        super().__init__()
        self.pool = GlobalAveragePooling2D()
        self.fc1 = Dense(int(in_channels*r))
        self.relu = Activation(hard_sigmoid)
        self.fc2 = Dense(in_channels, activation='sigmoid')

    def call(self, x):
        out = self.pool(x)
        out = self.fc1(out)
        out = self.relu(out)
        out = self.fc2(out)
        out *= x
        return out

mobilenetv3で使うbottleneck blockを実装します。

class mobilenetv3_block(Model):
    def __init__(self, exp_size, out_channels, kernel, SE, NL, s, r=0.25):
        super().__init__()        
        if NL == 'HS':
            activation = hard_swish
        else:
            activation = "relu"

        self.conv1 = Conv2D(filters = exp_size, kernel_size = (1,1), padding = 'same')
        self.bn1 = BatchNormalization()
        self.relu1 = Activation(activation)

        self.conv2 = DepthwiseConv2D(kernel_size = kernel, strides=s, padding = 'same')
        self.bn2 = BatchNormalization()
        self.relu2 = Activation(activation)

        self.conv3 = Conv2D(filters = out_channels, kernel_size = (1,1), padding = 'same')
        self.bn3 = BatchNormalization()
        self.relu3 = Activation(activation)

        if SE:
            self.se = SE_Block(in_channels=out_channels, r=r)
        else:
            self.se = None

        self.add = Add()

    def call(self, x):
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu1(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu2(out)

        out = self.conv3(out)
        out = self.bn3(out)
        out = self.relu3(out)

        if self.se:
            out = self.se(out)

        if K.int_shape(out) == K.int_shape(x):
            out = self.add([out, x])

        return out

まず、MobileNetV3-Largeの実装をします。

class MobileNetV3_Large(Model):
    def __init__(self):
        super().__init__()
        self.conv1 = Conv2D(16, kernel_size=3, strides=2, padding='same')
        self.bn1 = BatchNormalization()
        self.relu1 = Activation(hard_swish)

        self.bneck1  = mobilenetv3_block(exp_size=16,  out_channels=16,  kernel=3, SE=False, NL='RE', s=1)
        self.bneck2  = mobilenetv3_block(exp_size=64,  out_channels=24,  kernel=3, SE=False, NL='RE', s=2)
        self.bneck3  = mobilenetv3_block(exp_size=72,  out_channels=24,  kernel=3, SE=False, NL='RE', s=1)
        self.bneck4  = mobilenetv3_block(exp_size=72,  out_channels=40,  kernel=5, SE=True,  NL='RE', s=2)
        self.bneck5  = mobilenetv3_block(exp_size=120, out_channels=40,  kernel=5, SE=True,  NL='RE', s=1)
        self.bneck6  = mobilenetv3_block(exp_size=120, out_channels=40,  kernel=5, SE=True,  NL='RE', s=1)
        self.bneck7  = mobilenetv3_block(exp_size=240, out_channels=80,  kernel=3, SE=False, NL='HS', s=2)
        self.bneck8  = mobilenetv3_block(exp_size=200, out_channels=80,  kernel=3, SE=False, NL='HS', s=1)
        self.bneck9  = mobilenetv3_block(exp_size=184, out_channels=80,  kernel=3, SE=False, NL='HS', s=1)
        self.bneck10 = mobilenetv3_block(exp_size=184, out_channels=80,  kernel=3, SE=False, NL='HS', s=1)
        self.bneck11 = mobilenetv3_block(exp_size=480, out_channels=112, kernel=3, SE=True,  NL='HS', s=1)
        self.bneck12 = mobilenetv3_block(exp_size=672, out_channels=112, kernel=3, SE=True,  NL='HS', s=1)
        self.bneck13 = mobilenetv3_block(exp_size=672, out_channels=160, kernel=5, SE=True,  NL='HS', s=2)
        self.bneck14 = mobilenetv3_block(exp_size=960, out_channels=160, kernel=5, SE=True,  NL='HS', s=1)
        self.bneck15 = mobilenetv3_block(exp_size=960, out_channels=160, kernel=5, SE=True,  NL='HS', s=1)

        self.conv2 = Conv2D(960, kernel_size=1, strides=1, padding='same')
        self.bn2 = BatchNormalization()
        self.relu2 = Activation(hard_swish)

        self.pool = GlobalAveragePooling2D()
        self.fc1 = Dense(1280)
        self.relu3 = Activation(hard_swish)
        self.fc2 = Dense(10, activation='softmax')

    def call(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu1(x)

        x = self.bneck1(x)
        x = self.bneck2(x)
        x = self.bneck3(x)
        x = self.bneck4(x)
        x = self.bneck5(x)
        x = self.bneck6(x)
        x = self.bneck7(x)
        x = self.bneck8(x)
        x = self.bneck9(x)
        x = self.bneck10(x)
        x = self.bneck11(x)
        x = self.bneck12(x)
        x = self.bneck13(x)
        x = self.bneck14(x)
        x = self.bneck15(x)

        x = self.conv2(x)
        x = self.bn2(x)
        x = self.relu2(x)

        x = self.pool(x)
        x = self.fc1(x)
        x = self.relu3(x)
        x = self.fc2(x)
        return x

次にsmallの実装をします。

class MobileNetV3_Small(Model):
    def __init__(self):
        super().__init__()
        self.conv1 = Conv2D(16, kernel_size=3, strides=2, padding='same')
        self.bn1 = BatchNormalization()
        self.relu1 = Activation(hard_swish)

        self.bneck1  = mobilenetv3_block(exp_size=16,  out_channels=16, kernel=3, SE=True,  NL='RE', s=2)
        self.bneck2  = mobilenetv3_block(exp_size=72,  out_channels=24, kernel=3, SE=False, NL='RE', s=2)
        self.bneck3  = mobilenetv3_block(exp_size=88,  out_channels=24, kernel=3, SE=False, NL='RE', s=1)
        self.bneck4  = mobilenetv3_block(exp_size=96,  out_channels=40, kernel=5, SE=True,  NL='HS', s=2)
        self.bneck5  = mobilenetv3_block(exp_size=240, out_channels=40, kernel=5, SE=True,  NL='HS', s=1)
        self.bneck6  = mobilenetv3_block(exp_size=240, out_channels=40, kernel=5, SE=True,  NL='HS', s=1)
        self.bneck7  = mobilenetv3_block(exp_size=120, out_channels=48, kernel=5, SE=True,  NL='HS', s=1)
        self.bneck8  = mobilenetv3_block(exp_size=144, out_channels=48, kernel=5, SE=True,  NL='HS', s=1)
        self.bneck9  = mobilenetv3_block(exp_size=288, out_channels=96, kernel=5, SE=True,  NL='HS', s=2)
        self.bneck10 = mobilenetv3_block(exp_size=576, out_channels=96, kernel=5, SE=True,  NL='HS', s=1)
        self.bneck11 = mobilenetv3_block(exp_size=576, out_channels=96, kernel=5, SE=True,  NL='HS', s=1)

        self.conv2 = Conv2D(576, kernel_size=1, strides=1, padding='same')
        self.bn2 = BatchNormalization()
        self.relu2 = Activation(hard_swish)

        self.pool = GlobalAveragePooling2D()
        self.fc1 = Dense(1024)
        self.relu3 = Activation(hard_sigmoid)
        self.fc2 = Dense(10, activation='softmax')

    def call(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu1(x)

        x = self.bneck1(x)
        x = self.bneck2(x)
        x = self.bneck3(x)
        x = self.bneck4(x)
        x = self.bneck5(x)
        x = self.bneck6(x)
        x = self.bneck7(x)
        x = self.bneck8(x)
        x = self.bneck9(x)
        x = self.bneck10(x)
        x = self.bneck11(x)

        x = self.conv2(x)
        x = self.bn2(x)
        x = self.relu2(x)

        x = self.pool(x)
        x = self.fc1(x)
        x = self.relu3(x)
        x = self.fc2(x)
        return x

構造を確認します。

model = MobileNetV3_Small()
model.build((None, 224, 224, 3))  # build with input shape.
dummy_input = Input(shape=(224, 224, 3))  # declare without batch demension.
model_summary = Model(inputs=[dummy_input], outputs=model.call(dummy_input))
model_summary.summary()

学習の設定をします。

epochs = 100
initial_lrate = 0.1

def decay(epoch, steps=100):
    initial_lrate = 0.1
    drop = 0.99
    epochs_drop = 3
    lrate = initial_lrate * math.pow(drop, math.floor((1+epoch)/epochs_drop))
    return lrate

sgd = RMSprop(lr=initial_lrate, rho=0.9, epsilon=1.0, decay=0.9)

lr_sc = LearningRateScheduler(decay, verbose=1)

model = MobileNetV1()
model.compile(loss=['categorical_crossentropy'], optimizer=sgd, metrics=['accuracy'])

pytorch

pytorchも同様に書いていきます。
必要なライブラリのインポートをします。

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F

import pytorch_lightning as pl
from torchmetrics import Accuracy as accuracy

hard-sigmoidとhard-swishはもともと用意されているので、それを使います。
SEBlockの実装

class SE_Block(nn.Module):
    def __init__(self, in_channels, r):
        super().__init__()
        self.pool = nn.AdaptiveAvgPool2d((1,1))
        self.fc1 = nn.Linear(in_channels, int(in_channels*r))
        self.relu = nn.ReLU(inplace=False)
        self.fc2 = nn.Linear(int(in_channels*r), in_channels)
        self.sigmoid = nn.Hardsigmoid()

    def forward(self, x):
        out = self.pool(x)
        out = out.view(out.shape[0], -1)
        out = self.fc1(out)
        out = self.relu(out)
        out = self.fc2(out)
        out = self.sigmoid(out)
        out = out.view(out.shape[0], out.shape[1], 1,1).expand_as(x)
        return out * x

mobilenetv3のbottleneck　blockを実装します。

class mobilenetv3_block(nn.Module):
    def __init__(self, in_channels, exp_size, out_channels, SE, NL, s, r=0.25):
        super().__init__()        
        if NL == 'HS':
            activation = nn.Hardswish
        else:
            activation = nn.ReLU

        self.conv1 = nn.Conv2d(in_channels=in_channels, out_channels=exp_size, kernel_size = 1)
        self.bn1 = nn.BatchNorm2d(exp_size)
        self.relu1 = activation(False)

        self.conv2 = nn.Conv2d(in_channels=exp_size, out_channels=exp_size, kernel_size = 3, stride=s, groups=exp_size, padding = 1)
        self.bn2 = nn.BatchNorm2d(exp_size)
        self.relu2 = activation(False)

        self.conv3 = nn.Conv2d(in_channels=exp_size, out_channels=out_channels, kernel_size = 1)
        self.bn3 = nn.BatchNorm2d(out_channels)
        self.relu3 = activation(False)

        if SE:
            self.se = SE_Block(in_channels=out_channels, r=r)
        else:
            self.se = None

    def forward(self, x):
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu1(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu2(out)

        out = self.conv3(out)
        out = self.bn3(out)
        out = self.relu3(out)

        if self.se:
            out = self.se(out)

        if out.shape == x.shape:
            out = out + x

        return out

まず、MobileNetV3-Largeの実装をします。

class MobileNetV3_Large(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=2, padding=1)
        self.bn1 = nn.BatchNorm2d(16)
        self.relu1 = nn.Hardswish()

        self.bneck1  = mobilenetv3_block(in_channels=16,  exp_size=16,  out_channels=16,  SE=False, NL='RE', s=1)
        self.bneck2  = mobilenetv3_block(in_channels=16,  exp_size=64,  out_channels=24,  SE=False, NL='RE', s=2)
        self.bneck3  = mobilenetv3_block(in_channels=24,  exp_size=72,  out_channels=24,  SE=False, NL='RE', s=1)
        self.bneck4  = mobilenetv3_block(in_channels=24,  exp_size=72,  out_channels=40,  SE=True,  NL='RE', s=2)
        self.bneck5  = mobilenetv3_block(in_channels=40,  exp_size=120, out_channels=40,  SE=True,  NL='RE', s=1)
        self.bneck6  = mobilenetv3_block(in_channels=40,  exp_size=120, out_channels=40,  SE=True,  NL='RE', s=1)
        self.bneck7  = mobilenetv3_block(in_channels=40,  exp_size=240, out_channels=80,  SE=False, NL='HS', s=2)
        self.bneck8  = mobilenetv3_block(in_channels=80,  exp_size=200, out_channels=80,  SE=False, NL='HS', s=1)
        self.bneck9  = mobilenetv3_block(in_channels=80,  exp_size=184, out_channels=80,  SE=False, NL='HS', s=1)
        self.bneck10 = mobilenetv3_block(in_channels=80,  exp_size=184, out_channels=80,  SE=False, NL='HS', s=1)
        self.bneck11 = mobilenetv3_block(in_channels=80,  exp_size=480, out_channels=112, SE=True,  NL='HS', s=1)
        self.bneck12 = mobilenetv3_block(in_channels=112, exp_size=672, out_channels=112, SE=True,  NL='HS', s=1)
        self.bneck13 = mobilenetv3_block(in_channels=112, exp_size=672, out_channels=160, SE=True,  NL='HS', s=2)
        self.bneck14 = mobilenetv3_block(in_channels=160, exp_size=960, out_channels=160, SE=True,  NL='HS', s=1)
        self.bneck15 = mobilenetv3_block(in_channels=160, exp_size=960, out_channels=160, SE=True,  NL='HS', s=1)

        self.conv2 = nn.Conv2d(160, 960, kernel_size=1, stride=1)
        self.bn2 = nn.BatchNorm2d(960)
        self.relu2 = nn.Hardswish()

        self.pool = nn.AdaptiveAvgPool2d((1,1))
        self.fc1 = nn.Linear(960, 1280)
        self.relu3 = nn.Hardswish()
        self.fc2 = nn.Linear(1280, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu1(x)

        x = self.bneck1(x)
        x = self.bneck2(x)
        x = self.bneck3(x)
        x = self.bneck4(x)
        x = self.bneck5(x)
        x = self.bneck6(x)
        x = self.bneck7(x)
        x = self.bneck8(x)
        x = self.bneck9(x)
        x = self.bneck10(x)
        x = self.bneck11(x)
        x = self.bneck12(x)
        x = self.bneck13(x)
        x = self.bneck14(x)
        x = self.bneck15(x)

        x = self.conv2(x)
        x = self.bn2(x)
        x = self.relu2(x)

        x = self.pool(x)
        x = x.view(x.shape[0], -1)
        x = self.fc1(x)
        x = self.relu3(x)
        x = self.fc2(x)
        return x

次にMobileNetV3-Smallの実装をします。

class MobileNetV3_Small(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=2, padding=1)
        self.bn1 = nn.BatchNorm2d(16)
        self.relu1 = nn.Hardswish()

        self.bneck1  = mobilenetv3_block(in_channels=16, exp_size=16,  out_channels=16, SE=True,  NL='RE', s=2)
        self.bneck2  = mobilenetv3_block(in_channels=16, exp_size=72,  out_channels=24, SE=False, NL='RE', s=2)
        self.bneck3  = mobilenetv3_block(in_channels=24, exp_size=88,  out_channels=24, SE=False, NL='RE', s=1)
        self.bneck4  = mobilenetv3_block(in_channels=24, exp_size=96,  out_channels=40, SE=True,  NL='HS', s=2)
        self.bneck5  = mobilenetv3_block(in_channels=40, exp_size=240, out_channels=40, SE=True,  NL='HS', s=1)
        self.bneck6  = mobilenetv3_block(in_channels=40, exp_size=240, out_channels=40, SE=True,  NL='HS', s=1)
        self.bneck7  = mobilenetv3_block(in_channels=40, exp_size=120, out_channels=48, SE=True,  NL='HS', s=1)
        self.bneck8  = mobilenetv3_block(in_channels=48, exp_size=144, out_channels=48, SE=True,  NL='HS', s=1)
        self.bneck9  = mobilenetv3_block(in_channels=48, exp_size=288, out_channels=96, SE=True,  NL='HS', s=2)
        self.bneck10 = mobilenetv3_block(in_channels=96, exp_size=576, out_channels=96, SE=True,  NL='HS', s=1)
        self.bneck11 = mobilenetv3_block(in_channels=96, exp_size=576, out_channels=96, SE=True,  NL='HS', s=1)

        self.conv2 = nn.Conv2d(96, 576, kernel_size=1, stride=1)
        self.bn2 = nn.BatchNorm2d(576)
        self.relu2 = nn.Hardswish()

        self.pool = nn.AdaptiveAvgPool2d((1,1))
        self.fc1 = nn.Linear(576, 1024)
        self.relu3 = nn.Hardswish()
        self.fc2 = nn.Linear(1024, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu1(x)

        x = self.bneck1(x)
        x = self.bneck2(x)
        x = self.bneck3(x)
        x = self.bneck4(x)
        x = self.bneck5(x)
        x = self.bneck6(x)
        x = self.bneck7(x)
        x = self.bneck8(x)
        x = self.bneck9(x)
        x = self.bneck10(x)
        x = self.bneck11(x)

        x = self.conv2(x)
        x = self.bn2(x)
        x = self.relu2(x)

        x = self.pool(x)
        x = x.view(x.shape[0], -1)
        x = self.fc1(x)
        x = self.relu3(x)
        x = self.fc2(x)
        return x

構造の確認をします。

from torchsummary import summary

summary(MobileNetV3_Small(), (3,224,224))

学習の設定をします。

class MobileNetV1Trainer(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.model = MobileNetV3_Small()

    def forward(self, x):
        x = self.model(x)
        return x

    def training_step(self, batch, batch_idx):
        x, y = batch 
        #x, y = x.to(device), y.to(device)
        y_hat = self.forward(x)
        loss = nn.CrossEntropyLoss()(y_hat, y)
        return {'loss': loss, 'y_hat':y_hat, 'y':y, 'batch_loss': loss.item()*x.size(0)}

    def validation_step(self, batch, batch_idx):
        x, y = batch
        #x, y = x.to(device), y.to(device)
        y_hat = self.forward(x)
        loss = nn.CrossEntropyLoss()(y_hat, y)
        return {'y_hat':y_hat, 'y':y, 'batch_loss': loss.item()*x.size(0)}

    def test_step(self, batch, batch_nb):
        x, y = batch
        #x, y = x.to(device), y.to(device)
        y_hat = self.forward(x)
        loss = nn.CrossEntropyLoss()(y_hat, y)
        y_label = torch.argmax(y_hat, dim=1)
        acc = accuracy()(y_label, y)
        return {'test_loss': loss, 'test_acc': acc}

    def training_epoch_end(self, train_step_output):
        y_hat = torch.cat([val['y_hat'] for val in train_step_outputs], dim=0)
        y = torch.cat([val['y'] for val in train_step_outputs], dim=0)
        epoch_loss = sum([val['batch_loss'] for val in train_step_outputs]) / y_hat.size(0)
        preds = torch.argmax(y_hat, dim=1)
        acc = accuracy()(preds, y)
        self.log('train_loss', epoch_loss, prog_bar=True, on_epoch=True)
        self.log('train_acc', acc, prog_bar=True, on_epoch=True)

        print('---------- Current Epoch {} ----------'.format(self.current_epoch + 1))
        print('train Loss: {:.4f} train Acc: {:.4f}'.format(epoch_loass, acc))

    def validation_epoch_end(self, val_step_outputs):
        y_hat = torch.cat([val['y_hat'] for val in val_step_outputs], dim=0)
        y = torch.cat([val['y'] for val in val_step_outputs], dim=0)
        epoch_loss = sum([val['batch_loss'] for val in val_step_outputs]) / y_hat.size(0)
        preds = torch.argmax(y_hat, dim=1)
        acc = accuracy()(preds, y)
        self.log('val_loss', epoch_loss, prog_bar=True, on_epoch=True)
        self.log('val_acc', acc, prog_bar=True, on_epoch=True)

        print('valid Loss: {:.4f} valid Acc: {:.4f}'.format(epoch_loss, acc))

    # New: テストデータに対するエポックごとの処理
    def test_epoch_end(self, test_step_outputs):
        y_hat = torch.cat([val['y_hat'] for val in test_step_outputs], dim=0)
        y = torch.cat([val['y'] for val in test_step_outputs], dim=0)
        epoch_loss = sum([val['batch_loss'] for val in test_step_outputs]) / y_hat.size(0)
        preds = torch.argmax(y_hat, dim=1)
        acc = accuracy()(preds, y)
        self.log('test_loss', epoch_loss, prog_bar=True, on_epoch=True)
        self.log('test_acc', acc, prog_bar=True, on_epoch=True)

        print('test Loss: {:.4f} test Acc: {:.4f}'.format(epoch_loss, acc))

    def configure_optimizers(self):
        optimizer = optim.RMSprop(self.parameters(), lr=0.1, eps=1.0, momentum=0.9)
        scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.01)
        return {'optimizer': optimizer, 'lr_scheduler': scheduler}

転移学習(pytoch)

torchvisionで用意されています。

from torchvision.models import mobilenet_v3_large
from torchvision.models import mobilenet_v3_small
model = mobilenet_v3_small(pretrained=True)

最後の出力層のみ書き換えて学習させます。

class MobileNetV2Trainer(pl.LightningModule):
    def __init__(self):
        super(VGGTrainer, self).__init__()
        model = mobilenet_v3_small(pretrained=True)

        model.classifier[3] = nn.Linear(in_features=1024, out_features=10)
        self.model = model

        update_param_names = ['classifier.3.weight', 'classifier.3.bias']

        for name, param in self.model.named_parameters():
            if name in update_param_names:
                param.requires_grad = True
            else:
                param.requires_grad = False

    def forward(self, x):
        x = self.model(x)
        return x

    def training_step(self, batch, batch_idx):
        x, y = batch 
        #x, y = x.to(device), y.to(device)
        y_hat = self.forward(x)
        loss = nn.CrossEntropyLoss()(y_hat, y)
        return {'loss': loss, 'y_hat':y_hat, 'y':y, 'batch_loss': loss.item()*x.size(0)}

    def validation_step(self, batch, batch_idx):
        x, y = batch
        #x, y = x.to(device), y.to(device)
        y_hat = self.forward(x)
        loss = nn.CrossEntropyLoss()(y_hat, y)
        return {'y_hat':y_hat, 'y':y, 'batch_loss': loss.item()*x.size(0)}

    def test_step(self, batch, batch_nb):
        x, y = batch
        #x, y = x.to(device), y.to(device)
        y_hat = self.forward(x)
        loss = nn.CrossEntropyLoss()(y_hat, y)
        y_label = torch.argmax(y_hat, dim=1)
        acc = accuracy()(y_label, y)
        return {'test_loss': loss, 'test_acc': acc}

    def training_epoch_end(self, train_step_output):
        y_hat = torch.cat([val['y_hat'] for val in train_step_outputs], dim=0)
        y = torch.cat([val['y'] for val in train_step_outputs], dim=0)
        epoch_loss = sum([val['batch_loss'] for val in train_step_outputs]) / y_hat.size(0)
        preds = torch.argmax(y_hat, dim=1)
        acc = accuracy()(preds, y)
        self.log('train_loss', epoch_loss, prog_bar=True, on_epoch=True)
        self.log('train_acc', acc, prog_bar=True, on_epoch=True)

        print('---------- Current Epoch {} ----------'.format(self.current_epoch + 1))
        print('train Loss: {:.4f} train Acc: {:.4f}'.format(epoch_loass, acc))

    def validation_epoch_end(self, val_step_outputs):
        y_hat = torch.cat([val['y_hat'] for val in val_step_outputs], dim=0)
        y = torch.cat([val['y'] for val in val_step_outputs], dim=0)
        epoch_loss = sum([val['batch_loss'] for val in val_step_outputs]) / y_hat.size(0)
        preds = torch.argmax(y_hat, dim=1)
        acc = accuracy()(preds, y)
        self.log('val_loss', epoch_loss, prog_bar=True, on_epoch=True)
        self.log('val_acc', acc, prog_bar=True, on_epoch=True)

        print('valid Loss: {:.4f} valid Acc: {:.4f}'.format(epoch_loss, acc))

    # New: テストデータに対するエポックごとの処理
    def test_epoch_end(self, test_step_outputs):
        y_hat = torch.cat([val['y_hat'] for val in test_step_outputs], dim=0)
        y = torch.cat([val['y'] for val in test_step_outputs], dim=0)
        epoch_loss = sum([val['batch_loss'] for val in test_step_outputs]) / y_hat.size(0)
        preds = torch.argmax(y_hat, dim=1)
        acc = accuracy()(preds, y)
        self.log('test_loss', epoch_loss, prog_bar=True, on_epoch=True)
        self.log('test_acc', acc, prog_bar=True, on_epoch=True)

        print('test Loss: {:.4f} test Acc: {:.4f}'.format(epoch_loss, acc))

    def configure_optimizers(self):
        optimizer = optim.SGD(self.parameters(), lr=0.001, momentum=0.9, weight_decay=5e-4)
        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=10, threshold=0.0001)
        #return [optimizer,], [scheduler,]
        return {'optimizer': optimizer, 'lr_scheduler': scheduler, 'monitor': 'val_loss'}

以上で実装を終わります。