Rasmlarni tasniflash

Nazariya yetarli - endi haqiqiy to'plamda ishlaymiz. Fashion-MNIST kiyim rasmlaridan iborat: 70 000 ta kulrang 28x28 rasm, 10 ta sinf.

Ma'lumotni yuklash #

torchvision mashhur to'plamlarni o'zi yuklab beradi:

import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torchvision import datasets, transforms

torch.manual_seed(42)

o_ds = datasets.FashionMNIST("./data", train=True,  download=True, transform=transforms.ToTensor())
t_ds = datasets.FashionMNIST("./data", train=False, download=True, transform=transforms.ToTensor())

sinflar = ['Futbolka', 'Shim', 'Sviter', 'Libos', 'Palto',
           'Sandal', "Ko'ylak", 'Krossovka', 'Sumka', 'Botinka']

print("o'quv:", len(o_ds), "| test:", len(t_ds), "| sinflar:", len(sinflar))
x, y = o_ds[0]
print("rasm:", tuple(x.shape), "min/max:", round(x.min().item(), 2), round(x.max().item(), 2))
print("sinf taqsimoti:", torch.bincount(o_ds.targets).tolist())

o'quv: 60000 | test: 10000 | sinflar: 10
rasm: (1, 28, 28) min/max: 0.0 1.0
sinf taqsimoti: [6000, 6000, 6000, 6000, 6000, 6000, 6000, 6000, 6000, 6000]

Har sinfdan aynan 6000 tadan - to'plam mukammal muvozanatli. Bu qulay: aniqlikni baholashda hech qanday tuzatish kerak emas.

transforms.ToTensor() ikkita ishni bajaradi: rasmni tenzorga aylantiradi va qiymatlarni 0-255 dan 0.0-1.0 ga keltiradi.

Ikki modelni o'qitamiz #

o_dl = DataLoader(o_ds, batch_size=64, shuffle=True)
t_dl = DataLoader(t_ds, batch_size=256)

class CNN(nn.Module):
    def __init__(self):
        super().__init__()
        self.xususiyat = nn.Sequential(
            nn.Conv2d(1, 16, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
            nn.Conv2d(16, 32, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2))
        self.tasnif = nn.Sequential(nn.Flatten(), nn.Linear(32 * 7 * 7, 10))

    def forward(self, x):
        return self.tasnif(self.xususiyat(x))

O'quv sikli 8-bo'limdagi bilan bir xil - faqat ma'lumot boshqacha.

--- Zich tarmoq (101,770 parametr) ---
Davr | O'quv yo'q. | Test yo'q. | Test aniqlik | Vaqt
   1 |     0.5459 |     0.4845 |      83.33% |   6.1s
   2 |     0.3987 |     0.4080 |      85.69% |   5.8s
   3 |     0.3611 |     0.3952 |      85.87% |   5.8s
   4 |     0.3318 |     0.3918 |      86.02% |   5.7s
   5 |     0.3132 |     0.3538 |      87.33% |   5.7s

--- CNN (20,490 parametr) ---
Davr | O'quv yo'q. | Test yo'q. | Test aniqlik | Vaqt
   1 |     0.5235 |     0.3838 |      86.64% |  12.2s
   2 |     0.3447 |     0.3407 |      87.63% |  12.4s
   3 |     0.3060 |     0.3133 |      88.73% |  12.5s
   4 |     0.2825 |     0.3311 |      88.35% |  12.6s
   5 |     0.2656 |     0.2976 |      89.34% |  12.4s

Natijani o'qiymiz #

CNN 5 barobar kam parametr bilan 2 foiz ko'proq aniqlik berdi. Buning sababi 13-bo'limda aytilgan edi: konvolyutsiya piksellarning qo'shnichiligini saqlaydi.

Diqqat qiling: 4-davrda CNN ning test yo'qotishi 0.3133 dan 0.3311 ga o'sdi, keyin yana tushdi. Bunday kichik tebranish normal - 11-bo'limdagi barqaror o'sish esa muammo belgisi edi.

Qaysi sinf qiyin #

Umumiy aniqlik 89.34% - lekin u har sinfda bir xil emas:

cnn.eval()
togri = torch.zeros(10); jami = torch.zeros(10)
chalkashlik = torch.zeros(10, 10, dtype=torch.long)

with torch.no_grad():
    for xb, yb in t_dl:
        p = cnn(xb).argmax(1)
        for haq, bash in zip(yb, p):
            jami[haq] += 1
            chalkashlik[haq, bash] += 1
            if haq == bash:
                togri[haq] += 1

for i in range(10):
    print(f"{sinflar[i]:10s} | {togri[i]/jami[i]:6.1%}")

Sinf       | Aniqlik
-----------|--------
Futbolka   |  80.9%
Shim       |  97.7%
Sviter     |  83.1%
Libos      |  95.6%
Palto      |  83.7%
Sandal     |  97.0%
Ko'ylak    |  65.7%
Krossovka  |  94.8%
Sumka      |  97.8%
Botinka    |  97.1%

umumiy: 89.34%

Farq juda katta: Sumka 97.8%, Ko'ylak esa atigi 65.7%. Umumiy son buni butunlay yashirgan edi.

Nima bilan chalkashtiryapti #

juft = [(chalkashlik[i,j].item(), i, j) for i in range(10) for j in range(10) if i != j]
for n, i, j in sorted(juft, reverse=True)[:5]:
    print(f"  {sinflar[i]:10s} -> {sinflar[j]:10s} : {n} marta")

  Ko'ylak    -> Futbolka   : 114 marta
  Ko'ylak    -> Palto      : 92 marta
  Futbolka   -> Ko'ylak    : 88 marta
  Sviter     -> Palto      : 84 marta
  Futbolka   -> Libos      : 65 marta

Xatolar tasodifiy emas. Hammasi ustki kiyimlar orasida: ko'ylak, futbolka, sviter, palto. 28x28 kulrang rasmda ularning shakli deyarli bir xil - mato va tugmalar ko'rinmaydi.

Poyabzal va sumka esa hech qachon ular bilan chalkashmadi.

Xulosa #

• torchvision.datasets mashhur to'plamlarni bitta qator bilan yuklaydi.
• transforms.ToTensor() rasmni tenzorga aylantiradi va 0-1 oralig'iga keltiradi.
• Fashion-MNIST muvozanatli: har sinfdan 6000 tadan.
• CNN 20 490 parametr bilan 89.34%, zich tarmoq 101 770 parametr bilan 87.33% berdi.
• Ya'ni CNN 5 barobar kam parametrda aniqroq ishladi.
• CNN CPU da sekinroq, chunki konvolyutsiya ko'p hisob talab qiladi - GPU da bu farq yo'qoladi.
• Test yo'qotishining kichik tebranishi normal; barqaror o'sish esa ortiqcha moslashish belgisi.
• Umumiy aniqlik sinflar orasidagi katta farqni yashiradi.
• Bizning modelda Sumka 97.8%, Ko'ylak esa atigi 65.7% aniqlik berdi.
• Xatolar tasodifiy emas - ular o'zaro o'xshash sinflar orasida to'planadi.

Keyingi bo'limda ma'lumotni normallash va sun'iy kengaytirish bilan natijani yaxshilaymiz.