前面的部分與之前的程式碼相同，解析 csv 的資訊：

import pandas as pd
import numpy as np
import os
from PIL import Image

# 匯入的時候出錯，pandas.parser.CParserError: Error tokenizing data. C error
fasion_df = pd.read_csv("./fashion_product_images_small/myntradataset/styles.csv", on_bad_lines='skip')
# 輸出頭尾資料
fasion_df['image'] = fasion_df.apply(lambda row: str(row['id']) + ".jpg", axis=1)
fasion_df = fasion_df.reset_index(drop=True)

此處我們將無法讀取到的圖片和不是3維的圖片移除：

image_name =  fasion_df['image'].to_numpy()
image_path = [os.path.join("./fashion_product_images_small/myntradataset/images",i) for i in image_name ]

img_pil =  Image.open(image_path[0])
# 用來存有問題的圖片
fail_img_list = []
for i in range(len(image_path)):
    if os.path.isfile(image_path[i]):
        img_pil =  Image.open(image_path[i])
        if len(img_pil.getbands()) != 3:
            # print("image channels not RGB: ", image_path[i])
            print(os.path.basename(image_path[i]))
            fail_img_list.append(os.path.basename(image_path[i]))

    else:
        print("image path not exist: ", image_path[i])
        fail_img_list.append(os.path.basename(image_path[i]))
# 將例外的圖片去除
for fail_img_name in fail_img_list:
    fasion_df = fasion_df.drop(fasion_df[fasion_df["image"]==fail_img_name].index)

master_category = fasion_df["masterCategory"]
master_category_np = master_category.to_numpy()
print(len(master_category_np))

master_category_set = set(master_category_np)

master_category_list = list(master_category_set)
dic = {}
# 建立字典將每個類別對應到一個數值
for i in range(len(master_category_set)):
    dic[master_category_list[i]] = i

image_name =  fasion_df['image'].to_numpy()

image_name =  fasion_df['image'].to_numpy()
print("圖片總數: ", len(image_name))
image_path = [os.path.join("./fashion_product_images_small/myntradataset/images",i) for i in image_name ]
# 切割成訓練和測試集
image_path_train = image_path[:35000]
image_path_test = image_path[35000:]
master_category = fasion_df["masterCategory"].to_numpy()
category_label = []
# 將 label 文字轉為數字
for i in range(len(master_category)):
    category_label.append(dic[master_category[i]])
category_label = np.array(category_label)
category_label_train = category_label[:35000]
category_label_test = category_label[35000:]

因為使用 mobileNet，我們先中央裁減接著將圖片 resize 到 224*224：

from torch.utils.data import Dataset, DataLoader
from torchvision import transforms, utils
from matplotlib import pyplot as plt

normalize = transforms.Normalize(
    mean=[0.485, 0.456, 0.406],
    std=[0.229, 0.224, 0.225]
)
transform = transforms.Compose([
    transforms.CenterCrop(60),
    transforms.Resize(224),
    transforms.RandomHorizontalFlip(), # 圖像水平翻轉
    transforms.ToTensor(),
    normalize
])

def default_loader(path):
    img_pil =  Image.open(path)
    img_tensor = transform(img_pil)
    return img_tensor

class trainset(Dataset):
    def __init__(self, image_path, label, loader=default_loader):
        # 定義好 image 的路徑
        self.images = image_path
        self.label = label
        self.loader = loader

    def __getitem__(self, index):
        fn = self.images[index]
        img = self.loader(fn)
        label = self.label[index]
        return img, label

    def __len__(self):
        return len(self.images)
    
class testset(Dataset):
    def __init__(self, image_path, label, loader=default_loader):
        # 定義好 image 的路徑
        self.images = image_path
        self.label = label
        self.loader = loader

    def __getitem__(self, index):
        fn = self.images[index]
        img = self.loader(fn)
        label = self.label[index]
        return img, label

    def __len__(self):
        return len(self.images)
    
train_dataset  = trainset(image_path=image_path_train, label=category_label_train)
train_dataloader  = DataLoader(train_dataset, batch_size=64)
test_dataset  = testset(image_path=image_path_test, label=category_label_train)
test_loader = DataLoader(test_dataset, batch_size=64)

抓到要使用的裝置：

from torch import nn
import torch
device = (
    "cuda"
    if torch.cuda.is_available()
    else "mps"
    if torch.backends.mps.is_available()
    else "cpu"
)

載入 mobilenet_v3_model 架構和使用 pretrained 好的權重，把輸出層改成我們的類別數：

import torchvision.models as models
import torchvision


mobilenet_v3_model = models.mobilenet_v3_small(pretrained=True)
# print(mobilenet_v3_model)
num_classes = 7
mobilenet_v3_model.classifier = nn.Sequential(
    nn.Linear(576, 1024), 
    nn.ReLU(), 
    nn.Linear(1024, num_classes)
)
mobilenet_v3_model.to(device)

learning_rate = 1e-2
batch_size = 64
epochs = 10
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(mobilenet_v3_model.parameters(), lr=learning_rate)

def train_loop(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    model.train()
    for batch, (X, y) in enumerate(dataloader):
        y = torch.tensor(y, dtype=torch.long)
        # 將資料讀取到GPU中
        X, y = X.to(device), y.to(device)
        # 運算出結果並計算loss
        pred = model(X)
        loss = loss_fn(pred, y)

        # 反向傳播
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

        if batch % 1000 == 0:
            loss, current = loss.item(), (batch + 1) * len(X)
            print(f"loss: {loss:>7f}  [{current:>5d}/{size:>5d}]")

def test_loop(dataloader, model, loss_fn):
    model.eval()
    size = len(dataloader.dataset)
    num_batches = len(dataloader)
    test_loss, correct = 0, 0
    # 驗證或測試時記得加入 torch.no_grad() 讓神經網路不要更新
    with torch.no_grad():
        for X, y in dataloader:
            y = torch.tensor(y, dtype=torch.long)
            X, y = X.to(device), y.to(device)
            pred = model(X)
            test_loss += loss_fn(pred, y).item()
            correct += (pred.argmax(1) == y).type(torch.float).sum().item()

    test_loss /= num_batches
    correct /= size
    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")

for t in range(epochs):
    print(f"Epoch {t+1}\n-------------------------------")
    train_loop(train_dataloader, mobilenet_v3_model, loss_fn, optimizer)
    test_loop(test_loader, mobilenet_v3_model, loss_fn)
print("Done!")

訓練完大概有80%的準確率：

結語

今天大概把時尚資料集用到一個段落，明天會把我鐵人賽的系列文章整理成一個大綱，方便大家閱讀。