2021 ML Dev-Matching | 미술 작품 분류하기 : 출제자가 추천한 우수 코드 C

취업 이야기/데브매칭 문제 해설 2021. 7. 6. 09:32

2021 Dev-Matching: 머신러닝 과제테스트¶

☺️ 문제를 어떤 관점에서 정의하였는지 ☺️

문제는 이미지 분류 과제입니다. 주어진 데이터는 이미지고, 7가지 클래스로 분류하면 됩니다. 이미지는 데이터의 50% 이상이 노이즈이기 때문에, 딥러닝 모델이 유용합니다. 다만 어떤 관점에서 정의했냐고 말하기에는 너무나 명확한 문제라고 생각하기 때문에 문제 해결 과정을 적겠습니다.

8시간이라는 제한된 시간안에 밑바닥부터 코드를 작성해야 했기에, 일단 다른 것은 다 제쳐두고, 이미지 분류라는 과제를 해결하기 위한 베이스라인 코드를 구축하는 것부터 해결하고자 하였습니다. 환경설정, 데이터, 모델, 훈련을 위한 세팅 함수들을 하나씩 작성해나가면서 훈련 프로세스를 구축하였고, 빠른 실험을 위해 config만 수정하면 바로 새로운 실험을 실행할 수 있도록 하였습니다.

저는 문제 해결의 관점으로 다음의 3가지만 생각하였습니다

데이터
모델
앙상블

데이터의 경우 가볍게 EDA(Exploratory Data Analysis, 탐색적 데이터 분석)를 해본 결과 미술관에 전시될 그림들로 구성되어있었습니다. 리얼리즘과 추상화의 중간 즈음에 위치한 그림들로 대부분의 이미지들은 중앙에 위치하고 있고, 몇몇 누락된 이미지들을 제외하면 깔끔한 데이터라고 생각합니다. 무엇보다 불균형이 약간 있지만 크게 문제될 수준은 아니라고 보았습니다.

데이터에서는 특별히 손볼 것이 없다고 생각하여, 사용하는 모델에 맞춰 이미지 사이즈를 조정하는 것과 선형공선성을 막기 위한 normalization 정도만 진행하였습니다. 다양한 색을 augmentation으로 변형하여 넣어준다면 효과적일 수도 있다는 생각이 들었지만 넘어갔습니다.

모델에서는 SOTA(State-of-the-art)모델을 사용하는 것에 중점을 두었습니다. ViT(Vision Transformer)모델의 경우 기존 Convolution filter를 사용하던 모델에 비해 성능이 높고 가볍기 때문에 해당 모델을 중점으로 채용하였고 앙상블을 위해 Convolution filter를 사용한 다른 좋은 성능의 모델들도 사용하였습니다.

빠른 실험을 위해 timm 라이브러리를 사용하였습니다.

앙상블의 경우 빠르고 간단한 KFold 앙상블과, pseudo labeling 앙상블을 고려하였습니다.

정해진 시간 안에 끝내야 하는 만큼 간단하고 효과적인 요소들만 사용하였습니다.

☺️ 문제 해결 위해 만든 모델에 대한 설명과 그 모델을 선택한 이유 ☺️

처음에는 베이스라인 코드를 밑바닥부터 작성한만큼 에러를 잡기 위해 간단한 convnet모델로 출발하였습니다. 아주 작은 모델로 모델 학습이 진행되는지 알아보고자 테스트하였습니다.

그다음으로는 이전 SOTA였던 EfficientNet01을 불러와 작고 가볍지만 pretrained된 모델에서 성능 향상이 일어나는지 확인하였습니다. 성능은 확실히 증가하였고, 전체 훈련 프로세스 과정에 이상이 없다는 것을 확인한 후 바로 timm 모델을 사용하였습니다.

ViT(Vision Transformer)를 사용한 이유는 간단합니다. 가벼우며 성능이 매우 강력하기 때문입니다. 처음부터 자체적으로 훈련시킨다면 inductive bias가 약한 모델인만큼 사용하기 어렵겠지만 timm 라이브러리에서 pretrained된 모델을 얼마든지 가져다 쓸 수 있는 환경이기 때문에 가볍게 가져다가 사용하였습니다.

이외에 앙상블을 위해 convolution filter를 사용한 모델들 중 강력한 모델로 알려진 모델들을 가져다가 훈련시켰습니다.

환경 설정¶

패키지 다운로드
패키지 임포트
Config 설정

In [1]:

import os
import gc
import copy

import numpy as np
import pandas as pd

import easydict
from tqdm import notebook

# 이미지
from PIL import Image

# 시각화
import matplotlib.pyplot as plt

# PyTorch
import torch
from torch.utils.data import Dataset
from torch.utils.data import DataLoader

import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Adam, AdamW

from torch.optim.lr_scheduler import ReduceLROnPlateau

from torchvision import transforms

In [2]:

!pip install timm

import timm

Requirement already satisfied: timm in /usr/local/lib/python3.7/dist-packages (0.4.9) Requirement already satisfied: torch>=1.4 in /usr/local/lib/python3.7/dist-packages (from timm) (1.8.1+cu101) Requirement already satisfied: torchvision in /usr/local/lib/python3.7/dist-packages (from timm) (0.9.1+cu101) Requirement already satisfied: typing-extensions in /usr/local/lib/python3.7/dist-packages (from torch>=1.4->timm) (3.7.4.3) Requirement already satisfied: numpy in /usr/local/lib/python3.7/dist-packages (from torch>=1.4->timm) (1.19.5) Requirement already satisfied: pillow>=4.1.1 in /usr/local/lib/python3.7/dist-packages (from torchvision->timm) (7.1.2)

In [3]:

config = {}

# 설정
config['seed'] = 42
config['device'] = "cuda" if torch.cuda.is_available() else "cpu"

# 데이터
config['num_workers'] = 1
config['pin_memory'] = True

# 모델
config['hidden_dim'] = 128
config['dropout'] = 0.1

# 훈련
config['n_epochs'] = 20
config['batch_size'] = 64
config['lr'] = 0.0001
config['clip_grad'] = 10
config['log_steps'] = 50
config['patience'] = 5

args = easydict.EasyDict(config)

데이터¶

데이터 lookup 테이블 만들기
EDA
Dataset & DataLoader
train / valid 나누기
get_loader

데이터 lookup 테이블 만들기¶

PATH 설정
label2idx, idx2label
train_df
test_df

In [4]:

PROJECT_PATH = "/content/drive/MyDrive/테스트"

# 데이터셋 패스 설정
TRAIN_PATH = os.path.join(PROJECT_PATH, 'train')
TEST_PATH = os.path.join(PROJECT_PATH, 'test')

In [5]:

# label2idx
label2idx = {'dog': 0,
             'elephant': 1,
             'giraffe': 2,
             'guitar': 3,
             'horse': 4,
             'house': 5,
             'person': 6}

idx2label = {v: k for k, v in label2idx.items()}

In [6]:

# train_df 만들기
labels = []
for path in os.walk(TRAIN_PATH):
    label_type = path[0].split('/')[-1]
    if label_type == 'train':
        continue

    base_path = path[0]
    images = path[2]
    
    idx = label2idx[label_type]
    for image in images:
        img_path = os.path.join(base_path, image)
        label = {'img_path': img_path,
                 'label': idx}

        labels.append(label)

train_df = pd.DataFrame(labels)
train_df = train_df.sort_values(['label', 'img_path'])
train_df = train_df.reset_index(drop=True)
train_df.head(5)

Out[6]:

	img_path	label
0	/content/drive/MyDrive/테스트/train/dog/pic_001.jpg	0
1	/content/drive/MyDrive/테스트/train/dog/pic_002.jpg	0
2	/content/drive/MyDrive/테스트/train/dog/pic_003.jpg	0
3	/content/drive/MyDrive/테스트/train/dog/pic_004.jpg	0
4	/content/drive/MyDrive/테스트/train/dog/pic_005.jpg	0

In [7]:

# test_df 만들기
labels = []
for path in os.walk(TEST_PATH):
    base_path = path[0]
    images = path[2]
    
    for image in images:
        img_path = os.path.join(base_path, image)
        label = {'img_path': img_path,
                 'label': -1}

        labels.append(label)

test_df = pd.DataFrame(labels)
test_df = test_df.sort_values(['img_path'])
test_df = test_df.reset_index(drop=True)
test_df.head(5)

Out[7]:

	img_path	label
0	/content/drive/MyDrive/테스트/test/0001.jpg	-1
1	/content/drive/MyDrive/테스트/test/0002.jpg	-1
2	/content/drive/MyDrive/테스트/test/0003.jpg	-1
3	/content/drive/MyDrive/테스트/test/0004.jpg	-1
4	/content/drive/MyDrive/테스트/test/0005.jpg	-1

EDA¶

실제 사진이 아닌, 미술작품들
각 라벨별로 누락된 이미지들이 있음

In [8]:

image = plt.imread(train_df['img_path'][0])
plt.imshow(image);

In [9]:

image = plt.imread(test_df['img_path'][0])
plt.imshow(image);

In [10]:

train_df.groupby('label').count()

Out[10]:

	img_path
label
0	329
1	205
2	235
3	134
4	151
5	245
6	399

In [11]:

test_df.count()

Out[11]:

img_path 350 label 350 dtype: int64

Dataset & DataLoader¶

In [12]:

class ArtDataset(Dataset):
    def __init__(self, df, transform=None):
        self.df = df
        self.transform = transform

    def __getitem__(self, idx):
        data = self.df.iloc[idx]

        # 이미지
        img_path = data['img_path']
        image = Image.open(img_path)

        if self.transform:
            image = self.transform(image)

        # 라벨
        label = data['label']

        return image, label

    def __len__(self):
        return len(self.df)

In [13]:

transform = transforms.Compose([transforms.CenterCrop(224),
                                transforms.ToTensor(),
                                transforms.Normalize([0.485, 0.456, 0.406],
                                                     [0.229, 0.224, 0.225])])

dataset = ArtDataset(train_df, transform)
dataset[7]

Out[13]:

(tensor([[[ 1.6495,  1.6838,  1.7009,  ...,  0.8276,  0.6734,  0.7419],
          [ 1.6838,  1.7865,  1.8722,  ...,  0.8276,  0.6563,  0.6563],
          [ 1.6838,  1.9235,  2.0605,  ...,  0.9474,  0.6563,  0.6049],
          ...,
          [-1.3644, -1.6042, -1.6555,  ..., -1.6213, -1.6727, -1.7069],
          [-1.2959, -1.5699, -1.7754,  ..., -1.7925, -1.8782, -1.8268],
          [-1.4158, -1.7069, -1.7583,  ..., -1.9980, -2.0152, -1.9809]],
 
         [[ 1.7458,  1.7983,  1.8508,  ...,  0.2227,  0.2227,  0.3102],
          [ 1.7808,  1.8859,  2.0259,  ...,  0.0826,  0.0651,  0.0651],
          [ 1.7633,  2.0259,  2.2185,  ...,  0.1527, -0.0399, -0.1099],
          ...,
          [-1.4580, -1.7031, -1.7731,  ..., -1.6856, -1.6155, -1.6506],
          [-1.3704, -1.6681, -1.8782,  ..., -1.8256, -1.8081, -1.7556],
          [-1.4930, -1.8081, -1.8606,  ..., -2.0357, -1.9482, -1.9132]],
 
         [[ 1.4548,  1.4722,  1.5245,  ..., -1.3861, -1.2467, -1.1247],
          [ 1.5245,  1.5942,  1.6988,  ..., -1.5953, -1.4559, -1.4210],
          [ 1.5942,  1.7685,  1.9080,  ..., -1.5081, -1.5256, -1.5430],
          ...,
          [-1.1596, -1.3687, -1.3687,  ..., -1.1596, -1.3339, -1.3687],
          [-1.1247, -1.3339, -1.5081,  ..., -1.4559, -1.6127, -1.5604],
          [-1.2467, -1.4733, -1.4907,  ..., -1.6650, -1.7522, -1.7173]]]), 0)

In [14]:

loader = DataLoader(dataset, shuffle=True, batch_size=32)

for x, y in loader:
    print(x, y)
    break

tensor([[[[ 1.1187,  1.1187,  1.1358,  ...,  0.6734,  0.5878,  0.7419],
          [ 1.1187,  1.1358,  1.1358,  ...,  0.6906,  0.5878,  0.7591],
          [ 1.1358,  1.1358,  1.1358,  ...,  0.6906,  0.5707,  0.7933],
          ...,
          [ 0.2796,  0.2967,  0.3309,  ..., -2.1179, -1.9295, -1.8610],
          [ 0.2967,  0.3309,  0.3309,  ..., -2.0837, -1.8953, -1.8953],
          [-0.2684,  0.0056,  0.4679,  ..., -1.9809, -1.8439, -1.8953]],

         [[ 1.2731,  1.2731,  1.2906,  ...,  0.8179,  0.7304,  0.6779],
          [ 1.2731,  1.2906,  1.2906,  ...,  0.8354,  0.7479,  0.7304],
          [ 1.2906,  1.2906,  1.2906,  ...,  0.8179,  0.7829,  0.7829],
          ...,
          [ 0.3627,  0.3452,  0.3803,  ..., -1.9482, -2.0007, -1.9307],
          [ 0.3102,  0.3102,  0.3803,  ..., -1.9482, -1.9482, -1.9482],
          [-0.2675, -0.0224,  0.5203,  ..., -1.9132, -1.8957, -1.9482]],

         [[ 1.6291,  1.6291,  1.6465,  ...,  1.0714,  0.9494,  0.8971],
          [ 1.6291,  1.6465,  1.6465,  ...,  1.0888,  1.0017,  0.9668],
          [ 1.6465,  1.6465,  1.6465,  ...,  1.1237,  1.0539,  1.0539],
          ...,
          [ 0.4962,  0.4962,  0.5485,  ..., -1.7522, -1.7522, -1.6824],
          [ 0.4265,  0.4439,  0.4962,  ..., -1.5953, -1.7522, -1.7522],
          [-0.1487,  0.1128,  0.6356,  ..., -1.5779, -1.6999, -1.7522]]],


        [[[ 0.6392,  0.8789,  0.9988,  ..., -0.4054, -0.4397, -0.4397],
          [ 0.7933,  0.8447,  0.9646,  ..., -0.4226, -0.4226, -0.4226],
          [ 0.5364,  0.6049,  0.8104,  ..., -0.4054, -0.4054, -0.4054],
          ...,
          [ 1.5125,  1.5468,  1.5468,  ..., -0.4397, -0.4226, -0.4226],
          [ 1.4269,  1.4783,  1.5125,  ..., -0.4397, -0.4226, -0.4226],
          [ 1.3584,  1.3413,  1.3413,  ..., -0.4397, -0.4397, -0.4397]],

         [[-0.2150, -0.0399,  0.0301,  ...,  0.2227,  0.2227,  0.2227],
          [-0.2500, -0.0574,  0.1352,  ...,  0.2402,  0.2402,  0.2402],
          [-0.6702, -0.4076, -0.0399,  ...,  0.2577,  0.2577,  0.2577],
          ...,
          [ 2.0259,  2.0609,  2.0609,  ...,  0.2052,  0.2052,  0.2052],
          [ 1.8333,  1.8859,  1.9209,  ...,  0.2052,  0.2227,  0.2227],
          [ 1.7633,  1.7458,  1.7108,  ...,  0.2052,  0.2052,  0.2052]],

         [[-0.3230, -0.1835, -0.1835,  ...,  1.6640,  1.6117,  1.6117],
          [-0.3578, -0.2532, -0.1138,  ...,  1.6291,  1.6291,  1.6291],
          [-0.7238, -0.6018, -0.3230,  ...,  1.6465,  1.6465,  1.6465],
          ...,
          [ 2.3960,  2.4308,  2.4657,  ...,  1.6814,  1.7511,  1.7511],
          [ 2.2391,  2.2914,  2.3263,  ...,  1.6465,  1.6988,  1.6988],
          [ 2.1694,  2.1520,  2.1346,  ...,  1.6465,  1.6814,  1.6814]]],


        [[[ 0.1254,  0.1083,  0.1254,  ..., -1.3302, -1.3815, -1.4329],
          [ 0.0912,  0.0912,  0.1254,  ..., -1.2788, -1.3302, -1.3815],
          [ 0.0912,  0.0912,  0.0912,  ..., -1.3302, -1.3473, -1.3815],
          ...,
          [-1.5357, -1.6898, -1.5357,  ..., -1.5699, -1.6384, -1.6898],
          [-1.5699, -1.6384, -1.5870,  ..., -1.5185, -1.6213, -1.6727],
          [-1.5357, -1.6042, -1.5699,  ..., -1.4843, -1.5870, -1.6384]],

         [[ 0.8179,  0.8529,  0.8704,  ..., -0.9328, -0.9853, -1.0378],
          [ 0.8179,  0.8179,  0.8704,  ..., -0.9328, -0.9153, -0.9678],
          [ 0.8179,  0.8179,  0.8354,  ..., -1.0203, -0.9503, -0.9853],
          ...,
          [-1.2304, -1.3529, -1.2129,  ..., -1.4580, -1.3704, -1.4230],
          [-1.2654, -1.3354, -1.2829,  ..., -1.4055, -1.3529, -1.4055],
          [-1.2304, -1.3004, -1.2654,  ..., -1.3704, -1.3179, -1.3704]],

         [[ 1.6988,  1.6291,  1.5768,  ..., -0.9853, -0.9330, -0.9853],
          [ 1.6814,  1.6465,  1.6117,  ..., -0.9678, -0.9678, -1.0201],
          [ 1.6988,  1.6814,  1.6117,  ..., -1.0027, -1.0550, -1.0898],
          ...,
          [-1.2467, -1.3513, -1.1421,  ..., -0.9156, -1.1596, -1.2119],
          [-1.2467, -1.3164, -1.2293,  ..., -1.0898, -1.0898, -1.1421],
          [-1.2119, -1.2816, -1.2119,  ..., -1.0550, -1.0550, -1.1073]]],


        ...,


        [[[ 1.1529,  1.1529,  1.1700,  ...,  0.1254,  0.1254,  0.3138],
          [ 1.3070,  1.2728,  1.1187,  ...,  0.3309,  0.4337,  0.5193],
          [ 1.3413,  1.3070,  1.1015,  ...,  0.5536,  0.6563,  0.5364],
          ...,
          [ 2.2489,  2.2489,  2.2489,  ...,  2.0948,  2.0777,  2.0948],
          [ 2.2489,  2.2489,  2.2489,  ...,  2.1975,  2.2147,  2.2147],
          [ 2.2489,  2.2489,  2.2489,  ...,  2.1975,  2.2147,  2.2147]],

         [[ 0.1527,  0.1702,  0.2402,  ..., -0.9503, -0.9503, -0.7577],
          [ 0.2577,  0.2752,  0.1877,  ..., -0.8102, -0.6702, -0.5826],
          [ 0.2927,  0.3102,  0.1176,  ..., -0.6702, -0.4951, -0.6176],
          ...,
          [ 2.4111,  2.4111,  2.4111,  ...,  2.4286,  2.4286,  2.4286],
          [ 2.3936,  2.3936,  2.3936,  ...,  2.3585,  2.3585,  2.3585],
          [ 2.4286,  2.4286,  2.4286,  ...,  2.3936,  2.3761,  2.3761]],

         [[-0.6367, -0.6193, -0.6018,  ..., -1.4384, -1.4559, -1.2641],
          [-0.5147, -0.5147, -0.6193,  ..., -1.2990, -1.2467, -1.1596],
          [-0.4798, -0.4798, -0.6715,  ..., -1.1247, -1.0898, -1.2119],
          ...,
          [ 2.5529,  2.5529,  2.5529,  ...,  2.6400,  2.6226,  2.6400],
          [ 2.6226,  2.6226,  2.6226,  ...,  2.6400,  2.6400,  2.6400],
          [ 2.6400,  2.6400,  2.6400,  ...,  2.6400,  2.6400,  2.6400]]],


        [[[-2.0837, -2.1179, -2.0837,  ..., -1.8097, -1.4500, -1.1589],
          [-2.0494, -2.1008, -2.1008,  ..., -1.7069, -1.3815, -0.9877],
          [-2.0494, -2.0152, -2.0323,  ..., -1.8439, -1.8097, -1.7069],
          ...,
          [-0.4397, -1.1418, -1.2617,  ..., -2.0494, -1.8782, -1.7754],
          [-0.3883, -0.6965, -0.8507,  ..., -1.9295, -1.9295, -1.8439],
          [-0.9877, -1.3473, -1.4672,  ..., -1.8610, -1.8610, -1.7925]],

         [[-2.0007, -2.0357, -2.0007,  ..., -1.6506, -1.2829, -0.9853],
          [-1.9657, -2.0182, -2.0182,  ..., -1.5455, -1.2129, -0.8102],
          [-1.9657, -1.9307, -1.9482,  ..., -1.6856, -1.6506, -1.5455],
          ...,
          [-0.1625, -0.8803, -1.0028,  ..., -1.6681, -1.7556, -1.6506],
          [-0.1800, -0.5126, -0.6702,  ..., -1.7906, -1.7906, -1.7031],
          [-0.7927, -1.1779, -1.3004,  ..., -1.7206, -1.7206, -1.6506]],

         [[-1.7696, -1.8044, -1.7696,  ..., -1.3687, -1.0376, -0.7413],
          [-1.7347, -1.7870, -1.7870,  ..., -1.2641, -0.9678, -0.5670],
          [-1.7347, -1.6999, -1.7173,  ..., -1.4036, -1.4036, -1.2990],
          ...,
          [-0.0267, -0.7064, -0.8284,  ..., -1.5430, -1.5430, -1.4384],
          [-0.0267, -0.3055, -0.4275,  ..., -1.4733, -1.4733, -1.3861],
          [-0.6367, -0.9678, -1.0550,  ..., -1.4036, -1.4036, -1.3339]]],


        [[[ 0.1597,  0.1426,  0.2796,  ...,  1.0673,  1.0844,  1.1187],
          [-0.3541, -0.1999,  0.0056,  ...,  1.2557,  1.3242,  1.3584],
          [-0.8335, -1.1247, -0.8164,  ...,  1.3584,  1.4098,  1.4440],
          ...,
          [ 1.1700,  1.0159,  0.9303,  ...,  0.6906,  0.3138, -0.6623],
          [ 0.9817,  0.7933,  0.7419,  ...,  1.1187,  0.9817,  0.7077],
          [ 1.0159,  0.9303,  0.8104,  ...,  1.1187,  1.1015,  1.3070]],

         [[ 0.2227,  0.1702,  0.2577,  ...,  1.1856,  1.1856,  1.2206],
          [-0.3550, -0.2500, -0.0749,  ...,  1.3081,  1.3782,  1.4132],
          [-0.9153, -1.2129, -0.9328,  ...,  1.3431,  1.4657,  1.5007],
          ...,
          [ 1.0805,  0.9230,  0.8179,  ...,  0.6429,  0.2577, -0.7402],
          [ 0.8354,  0.6429,  0.5903,  ...,  1.0455,  0.8704,  0.5903],
          [ 0.8704,  0.7829,  0.6604,  ...,  1.0455,  0.9930,  1.2031]],

         [[-0.1661, -0.1661, -0.0615,  ...,  1.1411,  1.1411,  1.1759],
          [-0.5844, -0.4624, -0.2707,  ...,  1.2805,  1.3502,  1.3851],
          [-1.0027, -1.2990, -1.0027,  ...,  1.3851,  1.4374,  1.4722],
          ...,
          [ 1.0714,  0.9494,  0.9145,  ...,  0.8274,  0.5485, -0.4450],
          [ 0.8797,  0.7228,  0.6879,  ...,  1.1237,  1.0017,  0.7228],
          [ 0.9145,  0.8622,  0.7576,  ...,  1.1062,  1.1237,  1.3328]]]]) tensor([6, 0, 5, 6, 6, 6, 3, 2, 0, 1, 5, 0, 6, 4, 5, 0, 6, 0, 3, 3, 0, 2, 3, 3,
        4, 4, 1, 0, 1, 6, 3, 1])

train / valid 나누기¶

pandas DataFrame을 기준으로 stratifiedKFold를 적용한다

In [15]:

from sklearn.model_selection import StratifiedKFold

def split_df(df, kfold_n=5):
    kfold = StratifiedKFold(n_splits=kfold_n)

    # 클래스 비율 고려하여 Fold별로 데이터 나눔
    X = df.img_path.values
    y = df.label.values

    datas = []
    for i, (train_index, valid_index) in enumerate(kfold.split(X, y)):
        train_df = df.iloc[train_index].copy().reset_index(drop=True)
        valid_df = df.iloc[valid_index].copy().reset_index(drop=True)

        datas.append((train_df, valid_df))

    return datas

In [16]:

datas = split_df(train_df)

In [17]:

# 첫번째 데이터셋 train
datas[0][0].groupby('label').count()

Out[17]:

	img_path
label
0	263
1	164
2	188
3	107
4	121
5	196
6	319

In [18]:

# 첫번째 데이터셋 valid
datas[0][1].groupby('label').count()

Out[18]:

	img_path
label
0	66
1	41
2	47
3	27
4	30
5	49
6	80

get_loader¶

In [19]:

def get_loader(train_data, valid_data, transform):
    train_dataset = ArtDataset(train_data, transform)
    valid_dataset = ArtDataset(valid_data, transform)


    train_loader = DataLoader(train_dataset,
                              shuffle=True,
                              num_workers=args.num_workers,
                              batch_size=args.batch_size,
                              pin_memory=args.pin_memory)
    valid_loader = DataLoader(valid_dataset,
                              shuffle=False,
                              num_workers=args.num_workers,
                              batch_size=args.batch_size,
                              pin_memory=args.pin_memory)
    
    return train_loader, valid_loader

모델¶

conv 모델
efficientnet1
timm

In [20]:

for X, y in loader:
    print(X.size(), y.size())
    break

>torch.Size([32, 3, 224, 224]) torch.Size([32])

conv 모델¶

간단한 convolution 모델을 사용해서 테스트하는 용도

In [21]:

class ConvNet(nn.Module):
    def __init__(self, args, class_n=7):
        super().__init__()

        self.model = nn.Sequential(
            # [32, 3, 224, 224] -> [32, 32, 55, 55]
            nn.Conv2d(in_channels=3, out_channels=32, kernel_size=8, stride=4),
            nn.BatchNorm2d(32),
            nn.ReLU(),
            nn.MaxPool2d((2, 2)), # [32, 64, 55, 55] -> [32, 64, 27, 27]

            # [32, 32, 27, 27] -> [32, 64, 11, 11]
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=7, stride=2),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.MaxPool2d((2, 2)), # [32, 64, 11, 11] -> [32, 64, 5, 5]
        )
        self.fc = nn.Linear(1600, class_n)


    def forward(self, x):
        x = self.model(x)
        x = x.view(x.size(0), -1)
        return self.fc(x)

In [22]:

model = ConvNet(args)
model(X)

Out[22]:

tensor([[ 0.9441,  0.3273, -0.6121, -1.1562, -1.1537, -0.2866, -0.3136],
        [ 0.0923,  0.5396, -0.3335, -0.2363, -0.3932, -0.4474,  0.0705],
        [ 1.1090, -0.0963, -0.3213, -0.8175,  0.0193, -0.1474,  0.2303],
        [ 0.2213,  0.5964,  0.1489, -0.2720, -0.7719, -0.3505, -0.3247],
        [ 0.3219,  1.3156,  0.1622, -1.7833, -0.5985, -0.6310, -1.0919],
        [ 0.6533,  0.4952, -0.2269, -1.0678, -0.2662, -0.7674, -0.3342],
        [ 0.5027,  0.1173,  0.1857,  0.0938, -0.4946,  0.8828,  0.1678],
        [ 0.6146,  0.3164,  0.5716, -0.8988, -0.5872, -0.4984, -0.7020],
        [ 0.8043,  0.7840,  0.3966, -0.5947, -0.4601,  0.3243,  0.0259],
        [ 1.0216,  0.4896,  0.2961, -0.5511, -0.4961,  0.1674,  0.0173],
        [ 0.2953,  0.8990, -0.0758, -1.0837, -0.4858, -0.1145,  0.3061],
        [ 0.4921,  0.3758,  0.5821, -0.8128, -1.1178, -0.5434, -0.3811],
        [ 0.9240,  0.8390, -0.2498, -0.8254, -0.8668, -0.4110, -0.0476],
        [ 1.7050,  0.8061,  0.0745, -1.4041, -1.5088,  0.4276,  0.0937],
        [ 0.4990,  0.2662,  0.1519, -1.1867, -0.9486, -1.2554, -0.0370],
        [ 0.9249,  0.7384, -0.2740, -0.4297, -1.1963, -0.0577, -0.1950],
        [ 0.6903,  1.0579, -0.3505, -0.5165, -0.5967, -0.1050, -0.4528],
        [ 0.7049,  0.4872,  0.3703, -0.6034, -0.9301,  0.2915,  0.0649],
        [ 0.7048,  0.7583,  0.3892, -0.7462, -0.4054,  0.1733, -0.2026],
        [ 0.9102, -0.2895,  0.1764, -1.3334, -1.1267, -0.7388, -0.3778],
        [ 0.6860,  0.9166,  0.2442, -0.5162, -0.4703,  0.0927, -0.6246],
        [ 0.5346, -0.3926,  0.6689, -0.8792, -0.7831, -0.3497,  0.0800],
        [ 0.8854,  0.6109, -0.3990, -1.1772, -0.1643, -0.3895, -0.6607],
        [ 1.3579,  0.5562,  0.4058, -0.9862, -1.3069,  0.0973, -0.5180],
        [ 0.5796,  0.8724,  0.6406, -0.8832, -0.5234, -0.4178, -0.2793],
        [ 1.1863,  0.6524,  0.3318, -0.7983, -0.7219, -0.2296, -0.0742],
        [ 0.4154,  0.5497,  0.1223, -0.9258,  0.0560,  0.0630, -0.0349],
        [ 1.0239,  0.2273,  0.2094, -1.0673, -0.4870,  0.0552, -0.2058],
        [ 0.6459,  0.5558,  0.1114, -1.1328, -0.7657,  0.2688, -0.2605],
        [ 0.7271,  0.3053, -0.1695, -0.8216, -0.6764, -0.0481, -0.0794],
        [ 0.2739,  0.3990,  0.3245, -0.8883, -0.3505,  0.0920,  0.1564],
        [ 1.4383,  0.4677,  0.1684, -1.4680, -0.7739,  0.3378, -0.0589]],
       grad_fn=<AddmmBackward>)

efficientnet1¶

In [23]:

class Eff01(nn.Module):
    def __init__(self, args, class_n=7):
        super().__init__()
        self.model = timm.create_model('efficientnet_b1', pretrained=True, num_classes=class_n)

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

In [24]:

model = Eff01(args)
model(X)

Out[24]:

tensor([[ 3.9316e-01, -1.8075e+00, -6.8078e-01,  1.3850e+00, -1.6314e+00,
          4.2127e+00,  8.1528e-01],
        [ 1.7965e+00,  7.2678e+00, -4.4178e+00,  3.0586e-01,  1.9594e+00,
          1.6846e+00, -6.7419e-01],
        [-2.7659e+00, -1.9599e+00, -2.3481e+00, -4.4121e+00,  8.8137e-02,
          3.4413e+00, -3.1732e+00],
        [-2.1057e+00, -2.9621e+00, -2.7104e+00, -3.4373e+00,  2.9861e+00,
          2.3211e+00, -4.8303e+00],
        [ 5.2630e+00,  8.9739e+00,  3.2290e+00,  5.1347e+00, -2.9204e+00,
          5.2858e+00, -4.6337e+00],
        [ 2.9802e+00,  6.9154e+00, -4.1486e+00, -4.6411e+00, -3.9898e+00,
          4.4854e+00,  3.2991e+00],
        [ 2.5207e+00,  1.6186e+00,  2.0065e+00,  3.0777e+00,  1.5605e+00,
         -3.3963e-01, -8.9962e-01],
        [-2.4768e+00,  4.9098e+00, -7.2241e+00, -1.5993e+00, -2.9143e+00,
         -5.8529e+00, -5.0937e+00],
        [-6.0758e+00,  7.5426e-01, -2.3609e+00, -2.4397e+00, -5.7454e+00,
          2.9179e+00, -2.9950e+00],
        [ 1.4790e+00,  6.0400e+00,  1.0130e+00,  1.6859e+00,  1.0481e+00,
         -2.8229e+00,  2.9617e-02],
        [ 5.7227e-01, -2.2472e+00, -2.3249e+00, -3.1104e+00, -1.1003e+00,
         -2.7433e+00, -3.7159e+00],
        [-2.2360e-01,  2.4394e-01,  4.0713e+00, -9.7840e-01, -1.8921e+00,
          1.4798e+00, -1.5911e+00],
        [-1.4358e+00,  1.8061e+00,  1.8506e+00, -2.6608e+00,  4.8493e+00,
          9.6075e-01, -5.7549e-01],
        [ 1.3821e+00,  4.6102e+00, -5.7297e-01, -2.3651e+00,  3.2042e+00,
          2.1695e-01, -2.7694e+00],
        [-2.1986e+00, -1.3774e+00, -2.0758e+00, -2.2942e+00, -6.2637e+00,
          3.4803e+00, -9.1589e+00],
        [-1.0898e+00,  3.8918e+00, -5.0518e+00, -7.5061e+00,  1.0734e+00,
          3.0160e+00, -2.0883e+00],
        [-3.8563e+00,  4.9663e+00, -1.8223e-01,  3.4113e+00,  6.0263e-01,
          2.2603e+00,  4.2581e+00],
        [ 5.6363e+00, -2.7442e-02, -1.2771e+00,  2.7350e+00,  5.1099e-01,
          4.8014e+00,  2.6841e+00],
        [ 1.5867e+00,  8.0267e+00,  3.5557e+00, -2.1183e-03, -2.8267e-01,
         -2.8626e+00,  2.4946e+00],
        [ 6.7608e-01,  3.2706e+00, -1.2610e+00,  3.0974e+00, -2.9917e+00,
          3.7218e+00, -6.0823e-01],
        [ 1.8446e+00,  3.5921e+00,  3.1441e+00, -2.8459e+00, -1.0907e+00,
         -1.2436e+00, -3.9847e+00],
        [ 2.0054e+00,  4.8740e+00,  4.1562e+00,  7.2944e-01,  4.3491e-01,
          1.6272e+00,  3.8274e-01],
        [-3.8758e+00,  2.1684e+00,  1.8707e+00, -6.0805e+00, -1.8065e+00,
         -1.1074e+00, -4.8875e+00],
        [ 5.3469e-01,  3.1542e+00,  1.5800e+00,  3.5849e-01, -5.1072e-01,
          2.5812e+00,  3.3352e-01],
        [-1.2886e+00,  6.4772e+00,  8.9048e-01, -3.9631e-01, -1.3061e+00,
          8.1103e+00, -3.2352e-01],
        [-1.2911e-01,  9.9117e-01,  2.9715e-02,  5.2929e+00,  1.9702e+00,
          2.3241e+00,  7.3989e+00],
        [-4.0124e+00, -4.2550e+00,  8.7637e-01, -8.5220e+00,  1.9189e+00,
         -8.4398e-02, -3.6078e+00],
        [ 2.8761e+00,  1.7192e+00,  6.2730e+00, -3.0518e-01, -6.0232e-01,
         -1.1055e+00, -3.1051e+00],
        [-2.1232e+00,  4.9213e-01,  4.0293e+00,  1.0153e+00,  3.6330e+00,
          3.2296e+00, -3.1552e+00],
        [-4.8670e+00,  2.8554e+00,  3.5081e+00, -4.5551e+00, -5.0727e-01,
          3.7878e+00, -4.3766e+00],
        [-2.5460e-01, -1.3759e+00, -1.8093e+00,  8.9533e-01,  3.3150e-03,
          2.9745e+00,  2.5689e+00],
        [ 7.1304e-01,  1.9664e+00,  2.8399e-01, -8.2121e-01, -3.6928e+00,
         -7.3601e+00, -5.2162e+00]], grad_fn=<AddmmBackward>)

timm¶

timm이 제공하는 다양한 모델을 마음대로 가져다쓰기 위한 모델

In [25]:

class Timm(nn.Module):
    def __init__(self, args, class_n=7):
        super().__init__()
        self.model = timm.create_model(args.timm_model, pretrained=True, num_classes=class_n)

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

훈련¶

초기 세팅¶

In [90]:

config = {}

# 설정
config['seed'] = 42
config['device'] = "cuda" if torch.cuda.is_available() else "cpu"
config['model_path'] = os.path.join(PROJECT_PATH, 'model')

# 데이터
config['num_workers'] = 1
config['pin_memory'] = False

# 모델
config['dropout'] = 0.1

# 훈련
config['n_epochs'] = 20
config['lr'] = 0.0001
config['batch_size'] = 32
config['clip_grad'] = 10
config['log_steps'] = 1
config['patience'] = 10

config['model'] = 'timm'
config['timm_model'] = 'vit_deit_base_distilled_patch16_224'

config['optimizer'] = 'adam'
config['scheduler'] = 'plateau'

config['load'] = False
config['model_name'] = ""

config['kfold'] = 1

config['pseudo_labeling'] = False
config['pseudo_iter_n'] = 2

args = easydict.EasyDict(config)

criterion¶

In [27]:

criterion = nn.CrossEntropyLoss()

get_model¶

In [28]:

def get_model(args):
    """
    Load model and move tensors to a given devices.
    """
    if args.model == 'convnet': model = ConvNet(args)
    if args.model == 'eff1': model = Eff01(args)
    if args.model == 'eff3': model = Eff03(args)
    if args.model == 'timm': model = Timm(args)

    model.to(args.device)

    return model

get_optimizer¶

In [29]:

def get_optimizer(model, args):
    if args.optimizer == 'adam':
        optimizer = Adam(model.parameters(), lr=args.lr, weight_decay=0.0)
    
    # 모든 parameter들의 grad값을 0으로 초기화
    optimizer.zero_grad()
    
    return optimizer

get_scheduler¶

In [73]:

def get_scheduler(optimizer, args):
    if args.scheduler == 'plateau':
        scheduler = ReduceLROnPlateau(optimizer, patience=3, factor=0.5, mode='max', verbose=True)

    return scheduler

train¶

In [31]:

def train(args, model, train_loader, optimizer):
    model.train()

    corrects = 0

    for step, (images, labels) in enumerate(train_loader):
        images = images.to(args.device)
        labels = labels.to(args.device)

        outputs = model(images)
        if args.timm_model == 'vit_deit_base_distilled_patch16_224':
            outputs = outputs[0]
        
        loss = criterion(outputs, labels)
        loss.backward()

        optimizer.step()
        optimizer.zero_grad()

        if step % args.log_steps == 0:
            print(f"Training steps: {step} Loss: {str(loss.item())}")

        _, preds = torch.max(outputs, 1)

        corrects += torch.sum(preds == labels.data)

    acc = corrects / args.train_len

    return acc

validate¶

In [32]:

def validate(args, model, valid_loader):
    model.eval()

    corrects = 0

    for images, labels in valid_loader:
        images = images.to(args.device)
        labels = labels.to(args.device)
        

        outputs = model(images)

        _, preds = torch.max(outputs, 1)
        corrects += torch.sum(preds == labels.data)

    acc = corrects / args.valid_len

    print(f'VALID ACC : {acc}\n')

    return acc, outputs

pseudo_labeling¶

In [103]:

def pseudo_labeling(args, model, train_data):
    # train_data에 새로운 df
    new_train_data = copy.deepcopy(test_df)

    test_dataset = ArtDataset(test_df, transform)
    test_loader = DataLoader(test_dataset,
                            shuffle=False,
                            num_workers=args.num_workers,
                            batch_size=args.batch_size,
                            pin_memory=args.pin_memory)
    model.eval()
    answers = []
    for images, labels in test_loader:
        images = images.to(args.device)
        labels = labels.to(args.device)

        outputs = model(images)

        _, preds = torch.max(outputs, 1)

        answers.extend(list(preds.cpu().numpy()))

    new_train_data.label = answers

    new_train_data = pd.concat([train_data, new_train_data]).reset_index(drop=True)

    return new_train_data

run¶

In [123]:

def run(args, train_data, valid_data):
    """데이터셋을 주면 처음부터 끝까지 훈련 사이클을 실행"""

    # 캐시 메모리 비우기 및 가비지 컬렉터 가동!
    torch.cuda.empty_cache()
    gc.collect()
    
    # 데이터 로드
    args.train_len = len(train_data) 
    args.valid_len = len(valid_data) 
    transform = transforms.Compose([transforms.CenterCrop(224),
                                    transforms.ToTensor(),
                                    transforms.Normalize([0.485, 0.456, 0.406],
                                                         [0.229, 0.224, 0.225])])
    train_loader, valid_loader = get_loader(train_data, valid_data, transform)

    if args.load:
        model = torch.load(os.path.join(args.model_path, args.model_name))
    else:
        model = get_model(args)
    optimizer = get_optimizer(model, args)
    scheduler = get_scheduler(optimizer, args)

    early_stopping_counter = 0
    best_acc = -1

    if args.pseudo_labeling:
        iter_n = args.pseudo_iter_n
    else:
        iter_n = 1

    # pseudo labeling을 위한 iteration
    for iter_i in range(iter_n):
        for epoch in notebook.tqdm(range(args.n_epochs)):
            print(f'Epoch {epoch + 1}/{args.n_epochs}')

            ### TRAIN
            train_acc = train(args, model, train_loader, optimizer)

            ### VALID
            valid_acc, outputs = validate(args, model, valid_loader)

            ### model save and early stopping
            if valid_acc > best_acc:
                best_acc = valid_acc
                early_stopping_counter = 0
                
                # 모델 저장!
                if args.model == 'timm':
                    save_name = f"{args.timm_model}_fold{args.kfold}_{str(best_acc.item())[:4]}"
                else:
                    save_name = f"{args.model}_fold{args.kfold}_{str(best_acc.item())[:4]}"

                torch.save(model, os.path.join(args.model_path, save_name))
                print(f'model saved! {save_name}')
            else:
                early_stopping_counter += 1
                if early_stopping_counter >= args.patience:
                    print(f'EarlyStopping counter: {early_stopping_counter} out of {args.patience}')
                    break     

            # scheduler
            if args.scheduler == 'plateau':
                scheduler.step(best_acc)

        # Pseudo Labeling
        if args.pseudo_labeling:
            print('pseudo labeing ongoing!')
            train_data = pseudo_labeling(args, model, train_data)

            # make a loader with new train data
            train_loader, _ = get_loader(train_data, valid_data, transform)

데이터 불러오기¶

In [105]:

datas = split_df(train_df)

# 하나의 fold만 사용
train_data = datas[args.kfold - 1][0]
valid_data = datas[args.kfold - 1][1]

훈련 진행¶

In [107]:

args.lr = 0.0001
args.load = False
args.model_name = ''

args.batch_size = 16
args.log_steps = 20

args.patience = 10

In [ ]:

timm.list_models(pretrained=True)

In [109]:

args.timm_model = 'vit_deit_base_distilled_patch16_224'
# args.timm_model = 'ecaresnet50t'
# args.timm_model = 'seresnext50_32x4d'
# args.timm_model = 'tf_efficientnet_b3_ns'
# args.timm_model = 'eca_nfnet_l0'
# args.timm_model = 'resnest50d'
# args.timm_model = 'regnety_016'

In [121]:

# pseudo labeling
args.pseudo_labeling = True
args.pseudo_iter_n = 3

args.n_epochs = 3

In [124]:

run(args, train_data, valid_data)

Epoch 1/3
Training steps: 0 Loss: 1.9610543251037598
Training steps: 20 Loss: 0.11243961751461029
Training steps: 40 Loss: 0.31864628195762634
Training steps: 60 Loss: 0.04142267629504204
Training steps: 80 Loss: 0.014966970309615135
VALID ACC : 0.9764705896377563

model saved! vit_deit_base_distilled_patch16_224_fold1_0.97
Epoch 2/3
Training steps: 0 Loss: 0.006128528621047735
Training steps: 20 Loss: 0.04373385012149811
Training steps: 40 Loss: 0.010203133337199688
Training steps: 60 Loss: 0.28457173705101013
Training steps: 80 Loss: 0.24638208746910095
VALID ACC : 0.979411780834198

model saved! vit_deit_base_distilled_patch16_224_fold1_0.97
Epoch 3/3
Training steps: 0 Loss: 0.11676128208637238
Training steps: 20 Loss: 0.0013939981581643224
Training steps: 40 Loss: 0.003479683306068182
Training steps: 60 Loss: 0.007280834950506687
Training steps: 80 Loss: 0.0033565755002200603
VALID ACC : 0.958823561668396


pseudo labeing ongoing!



Epoch 1/3
Training steps: 0 Loss: 0.04027740657329559
Training steps: 20 Loss: 0.12000448256731033
Training steps: 40 Loss: 0.017177144065499306
Training steps: 60 Loss: 0.2898399531841278
Training steps: 80 Loss: 0.005810225382447243
Training steps: 100 Loss: 0.12840282917022705
VALID ACC : 0.9323529601097107

Epoch 2/3
Training steps: 0 Loss: 0.1331356167793274
Training steps: 20 Loss: 0.08409298211336136
Training steps: 40 Loss: 0.018287230283021927
Training steps: 60 Loss: 0.0019370263908058405
Training steps: 80 Loss: 0.0035161140840500593
Training steps: 100 Loss: 0.0007211645133793354
VALID ACC : 0.9382352828979492

Epoch 3/3
Training steps: 0 Loss: 0.0049949875101447105
Training steps: 20 Loss: 0.0009458191343583167
Training steps: 40 Loss: 0.00010684832523111254
Training steps: 60 Loss: 0.0020058471709489822
Training steps: 80 Loss: 0.0013503937516361475
Training steps: 100 Loss: 0.0018512691603973508
VALID ACC : 0.9352941513061523

Epoch     6: reducing learning rate of group 0 to 5.0000e-05.

pseudo labeing ongoing!



Epoch 1/3
Training steps: 0 Loss: 0.0003172778233420104
Training steps: 20 Loss: 0.0002910386538133025
Training steps: 40 Loss: 0.0008431083406321704
Training steps: 60 Loss: 0.0006055183475837111
Training steps: 80 Loss: 0.00013664942525792867
Training steps: 100 Loss: 0.0006959352176636457
Training steps: 120 Loss: 9.717841749079525e-05
VALID ACC : 0.979411780834198

Epoch 2/3
Training steps: 0 Loss: 0.00014311331324279308
Training steps: 20 Loss: 0.0005445697461254895
Training steps: 40 Loss: 0.000317348021781072
Training steps: 60 Loss: 0.00019007959053851664
Training steps: 80 Loss: 0.0006655334727838635
Training steps: 100 Loss: 0.00010388605733169243
Training steps: 120 Loss: 0.00014714842836838216
VALID ACC : 0.979411780834198

Epoch 3/3
Training steps: 0 Loss: 0.0006675210315734148
Training steps: 20 Loss: 0.0002456515794619918
Training steps: 40 Loss: 0.0003927541838493198
Training steps: 60 Loss: 8.302312926389277e-05
Training steps: 80 Loss: 8.846087439451367e-05
Training steps: 100 Loss: 0.00022786625777371228
Training steps: 120 Loss: 7.981226372066885e-05
VALID ACC : 0.979411780834198


pseudo labeing ongoing!

추론¶

In [39]:

test_dataset = ArtDataset(test_df, transform)
test_loader = DataLoader(test_dataset,
                         shuffle=False,
                         num_workers=args.num_workers,
                         batch_size=args.batch_size,
                         pin_memory=args.pin_memory)

In [125]:

torch.cuda.empty_cache()
gc.collect()

Out[125]:

In [126]:

infer_model_name = 'timm_0.99'
infer_model_name = 'ecaresnet50t_0.98'
infer_model_name = 'vit_deit_base_distilled_patch16_224_0.97'


model = torch.load(os.path.join(args.model_path, infer_model_name))

In [127]:

model.eval()
answers = []
for images, labels in test_loader:
    images = images.to(args.device)
    labels = labels.to(args.device)

    outputs = model(images)

    _, preds = torch.max(outputs, 1)

    answers.extend(list(preds.cpu().numpy()))

In [129]:

print(answers)

>[2, 3, 3, 3, 3, 3, 4, 4, 3, 1, 6, 2, 0, 2, 3, 1, 2, 0, 6, 3, 3, 5, 2, 3, 0, 5, 1, 2, 0, 5, 4, 5, 6, 2, 0, 5, 5, 4, 2, 1, 4, 0, 2, 3, 1, 3, 0, 5, 5, 2, 6, 5, 4, 1, 5, 0, 4, 5, 1, 1, 5, 0, 6, 1, 1, 2, 4, 1, 1, 3, 1, 3, 0, 1, 1, 6, 2, 0, 3, 4, 1, 6, 1, 6, 6, 4, 3, 6, 1, 2, 2, 5, 1, 0, 5, 6, 3, 3, 1, 6, 5, 6, 5, 0, 3, 2, 5, 3, 0, 4, 4, 6, 2, 5, 4, 2, 0, 0, 5, 6, 4, 2, 2, 6, 4, 1, 5, 6, 0, 4, 1, 1, 6, 4, 4, 2, 5, 3, 5, 0, 4, 3, 1, 5, 1, 5, 4, 0, 2, 5, 6, 1, 6, 3, 2, 2, 0, 1, 4, 5, 2, 0, 6, 1, 3, 4, 1, 5, 6, 2, 1, 5, 3, 0, 0, 3, 2, 5, 3, 4, 2, 0, 3, 6, 0, 3, 3, 2, 0, 4, 4, 2, 2, 4, 4, 6, 6, 6, 3, 2, 5, 5, 4, 6, 2, 1, 3, 6, 0, 2, 3, 1, 1, 3, 1, 5, 2, 2, 0, 0, 0, 1, 2, 2, 6, 4, 3, 1, 5, 4, 3, 5, 0, 0, 4, 3, 0, 1, 4, 1, 6, 1, 2, 1, 6, 0, 5, 4, 6, 3, 4, 2, 5, 3, 1, 4, 2, 3, 4, 0, 6, 0, 2, 0, 1, 0, 1, 4, 0, 1, 1, 5, 6, 4, 6, 2, 4, 4, 3, 3, 6, 4, 0, 3, 1, 3, 0, 0, 5, 2, 5, 5, 2, 5, 0, 6, 5, 1, 5, 1, 5, 3, 0, 3, 3, 1, 6, 5, 0, 4, 0, 4, 3, 4, 4, 2, 5, 1, 2, 0, 6, 6, 1, 3, 3, 0, 0, 5, 2, 4, 6, 2, 6, 3, 6, 0, 4, 6, 2, 0, 2, 0, 5, 5, 1, 6, 3, 3, 5, 1]

In [130]:

label2idx

Out[130]:

{'dog': 0,
 'elephant': 1,
 'giraffe': 2,
 'guitar': 3,
 'horse': 4,
 'house': 5,
 'person': 6}

In [131]:

submission_path = os.path.join(PROJECT_PATH, 'submission.csv')
submission_df = pd.DataFrame({'answer_value': answers})
submission_df.to_csv(submission_path)

In [132]:

submission_df.value_counts()

Out[132]:

answer_value
3               53
1               53
0               52
5               50
2               50
4               47
6               45
dtype: int64

앙상블¶

KFold 앙상블

kfold 훈련¶

In [68]:

args.timm_model = 'vit_deit_base_distilled_patch16_224'
# args.timm_model = 'ecaresnet50t'
# args.timm_model = 'seresnext50_32x4d'
# args.timm_model = 'tf_efficientnet_b3_ns'
# args.timm_model = 'eca_nfnet_l0'
# args.timm_model = 'resnest50d'
# args.timm_model = 'regnety_016'

In [72]:

args.lr = 0.0001
args.load = False
args.model_name = ''

args.batch_size = 16
args.log_steps = 20

args.patience = 5

In [74]:

kfold_n = 5
datas = split_df(train_df)
for fold in range(1, kfold_n + 1):
    args.kfold = fold

    # 하나의 fold만 사용
    train_data = datas[args.kfold - 1][0]
    valid_data = datas[args.kfold - 1][1]
    run(args, train_data, valid_data)

Epoch 1/20
Training steps: 0 Loss: 2.070225954055786
Training steps: 20 Loss: 0.09332413226366043
Training steps: 40 Loss: 0.22841119766235352
Training steps: 60 Loss: 0.04848042130470276
Training steps: 80 Loss: 0.17917747795581818
VALID ACC : 0.979411780834198

model saved! vit_deit_base_distilled_patch16_224_fold1_0.97
Epoch 2/20
Training steps: 0 Loss: 0.013894079253077507
Training steps: 20 Loss: 0.11277658492326736
Training steps: 40 Loss: 0.06558889150619507
Training steps: 60 Loss: 0.12228739261627197
Training steps: 80 Loss: 0.015994392335414886
VALID ACC : 0.9764705896377563

Epoch 3/20
Training steps: 0 Loss: 0.007395995780825615
Training steps: 20 Loss: 0.0025502904318273067
Training steps: 40 Loss: 0.0007685854798182845
Training steps: 60 Loss: 0.0006357259699143469
Training steps: 80 Loss: 0.0006690919981338084
VALID ACC : 0.979411780834198

Epoch 4/20
Training steps: 0 Loss: 0.0008087260066531599
Training steps: 20 Loss: 0.0012696878984570503
Training steps: 40 Loss: 0.0009208667906932533
Training steps: 60 Loss: 0.0008509294711984694
Training steps: 80 Loss: 0.0008439231896772981
VALID ACC : 0.9823529720306396

model saved! vit_deit_base_distilled_patch16_224_fold1_0.98
Epoch 5/20
Training steps: 0 Loss: 0.0004001033375971019
Training steps: 20 Loss: 0.00033618463203310966
Training steps: 40 Loss: 0.0004378109297249466
Training steps: 60 Loss: 0.00022693183564115316
Training steps: 80 Loss: 0.0003972008707933128
VALID ACC : 0.9852941632270813

model saved! vit_deit_base_distilled_patch16_224_fold1_0.98
Epoch 6/20
Training steps: 0 Loss: 0.0001803989871405065
Training steps: 20 Loss: 0.00019188910664524883
Training steps: 40 Loss: 0.00015008499030955136
Training steps: 60 Loss: 0.00031750480411574244
Training steps: 80 Loss: 0.00015961111057549715
VALID ACC : 0.9852941632270813

Epoch 7/20
Training steps: 0 Loss: 0.0003038535942323506
Training steps: 20 Loss: 0.00013863165804650635
Training steps: 40 Loss: 0.0002890844480134547
Training steps: 60 Loss: 0.00029595577507279813
Training steps: 80 Loss: 0.00019118939235340804
VALID ACC : 0.9852941632270813

Epoch 8/20
Training steps: 0 Loss: 0.00021021650172770023
Training steps: 20 Loss: 0.00018152184202335775
Training steps: 40 Loss: 0.00023899308871477842
Training steps: 60 Loss: 0.000142873395816423
Training steps: 80 Loss: 0.00011664496560115367
VALID ACC : 0.9852941632270813

Epoch 9/20
Training steps: 0 Loss: 0.00021138088777661324
Training steps: 20 Loss: 0.00028158232453279197
Training steps: 40 Loss: 0.00019580998923629522
Training steps: 60 Loss: 0.00010884227231144905
Training steps: 80 Loss: 0.00013760558795183897
VALID ACC : 0.9852941632270813

Epoch     9: reducing learning rate of group 0 to 5.0000e-05.
Epoch 10/20
Training steps: 0 Loss: 0.00022388184152077883
Training steps: 20 Loss: 0.00010977809870382771
Training steps: 40 Loss: 0.00017114584625232965
Training steps: 60 Loss: 0.00014162009756546468
Training steps: 80 Loss: 7.580553938169032e-05
VALID ACC : 0.9852941632270813

EarlyStopping counter: 5 out of 5




Epoch 1/20
Training steps: 0 Loss: 2.125271797180176
Training steps: 20 Loss: 0.1512129306793213
Training steps: 40 Loss: 0.021708954125642776
Training steps: 60 Loss: 0.11550218611955643
Training steps: 80 Loss: 0.08638344705104828
VALID ACC : 0.9558823704719543

model saved! vit_deit_base_distilled_patch16_224_fold2_0.95
Epoch 2/20
Training steps: 0 Loss: 0.011126736178994179
Training steps: 20 Loss: 0.004127616062760353
Training steps: 40 Loss: 0.008746129460632801
Training steps: 60 Loss: 0.002552709076553583
Training steps: 80 Loss: 0.005797188729047775
VALID ACC : 0.9676470756530762

model saved! vit_deit_base_distilled_patch16_224_fold2_0.96
Epoch 3/20
Training steps: 0 Loss: 0.002060434315353632
Training steps: 20 Loss: 0.0007710144273005426
Training steps: 40 Loss: 0.0009948632214218378
Training steps: 60 Loss: 0.0002511692000553012
Training steps: 80 Loss: 0.0007774809491820633
VALID ACC : 0.9735294580459595

model saved! vit_deit_base_distilled_patch16_224_fold2_0.97
Epoch 4/20
Training steps: 0 Loss: 0.0007255766540765762
Training steps: 20 Loss: 0.0005823898245580494
Training steps: 40 Loss: 0.0002615299599710852
Training steps: 60 Loss: 0.0008422219543717802
Training steps: 80 Loss: 0.00038621696876361966
VALID ACC : 0.9764705896377563

model saved! vit_deit_base_distilled_patch16_224_fold2_0.97
Epoch 5/20
Training steps: 0 Loss: 0.0003897449350915849
Training steps: 20 Loss: 0.00023776116722729057
Training steps: 40 Loss: 0.00044705800246447325
Training steps: 60 Loss: 0.0005612970562651753
Training steps: 80 Loss: 0.00017410620057489723
VALID ACC : 0.9852941632270813

model saved! vit_deit_base_distilled_patch16_224_fold2_0.98
Epoch 6/20
Training steps: 0 Loss: 0.0002749292179942131
Training steps: 20 Loss: 0.0002847451833076775
Training steps: 40 Loss: 0.00047713532694615424
Training steps: 60 Loss: 0.0004889280535280704
Training steps: 80 Loss: 0.00030827586306259036
VALID ACC : 0.9852941632270813

Epoch 7/20
Training steps: 0 Loss: 0.00025418432778678834
Training steps: 20 Loss: 0.00016597700596321374
Training steps: 40 Loss: 0.00019834867271129042
Training steps: 60 Loss: 0.0003435465332586318
Training steps: 80 Loss: 0.0001903756201500073
VALID ACC : 0.9852941632270813

Epoch 8/20
Training steps: 0 Loss: 0.00014426129928324372
Training steps: 20 Loss: 0.00022852919937577099
Training steps: 40 Loss: 0.0002765594981610775
Training steps: 60 Loss: 0.000177494584931992
Training steps: 80 Loss: 0.0002137339033652097
VALID ACC : 0.9852941632270813

Epoch 9/20
Training steps: 0 Loss: 0.00018333044135943055
Training steps: 20 Loss: 0.00010870480036828667
Training steps: 40 Loss: 0.00012791194603778422
Training steps: 60 Loss: 0.00012232620792929083
Training steps: 80 Loss: 9.94913061731495e-05
VALID ACC : 0.9852941632270813

Epoch     9: reducing learning rate of group 0 to 5.0000e-05.
Epoch 10/20
Training steps: 0 Loss: 0.00013787909119855613
Training steps: 20 Loss: 0.00010700257553253323
Training steps: 40 Loss: 5.6122637033695355e-05
Training steps: 60 Loss: 0.00016332794621121138
Training steps: 80 Loss: 6.30063223070465e-05
VALID ACC : 0.9882352948188782

model saved! vit_deit_base_distilled_patch16_224_fold2_0.98
Epoch 11/20
Training steps: 0 Loss: 9.694581967778504e-05
Training steps: 20 Loss: 0.00010940924403257668
Training steps: 40 Loss: 0.00017602772277314216
Training steps: 60 Loss: 0.00010274062515236437
Training steps: 80 Loss: 8.837206405587494e-05
VALID ACC : 0.9882352948188782

Epoch 12/20
Training steps: 0 Loss: 0.0001780791935743764
Training steps: 20 Loss: 0.00010090007708640769
Training steps: 40 Loss: 0.00012891730875708163
Training steps: 60 Loss: 7.917111361166462e-05
Training steps: 80 Loss: 0.0002364671090617776
VALID ACC : 0.9882352948188782

Epoch 13/20
Training steps: 0 Loss: 8.280579640995711e-05
Training steps: 20 Loss: 7.735982217127457e-05
Training steps: 40 Loss: 9.142953786067665e-05
Training steps: 60 Loss: 0.00010475851740920916
Training steps: 80 Loss: 5.940067785559222e-05
VALID ACC : 0.9882352948188782

Epoch 14/20
Training steps: 0 Loss: 9.359944669995457e-05
Training steps: 20 Loss: 6.828831828897819e-05
Training steps: 40 Loss: 7.491827273042873e-05
Training steps: 60 Loss: 9.29912130231969e-05
Training steps: 80 Loss: 0.0001112468889914453
VALID ACC : 0.9882352948188782

Epoch    14: reducing learning rate of group 0 to 2.5000e-05.
Epoch 15/20
Training steps: 0 Loss: 0.0001017541071632877
Training steps: 20 Loss: 0.00010287428449373692
Training steps: 40 Loss: 7.4692492489703e-05
Training steps: 60 Loss: 5.7015880884137005e-05
Training steps: 80 Loss: 0.0001095659754355438
VALID ACC : 0.9882352948188782

EarlyStopping counter: 5 out of 5




Epoch 1/20
Training steps: 0 Loss: 1.9553475379943848
Training steps: 20 Loss: 0.360125869512558
Training steps: 40 Loss: 0.11961036920547485
Training steps: 60 Loss: 0.013302672654390335
Training steps: 80 Loss: 0.2760090231895447
VALID ACC : 0.9647058844566345

model saved! vit_deit_base_distilled_patch16_224_fold3_0.96
Epoch 2/20
Training steps: 0 Loss: 0.005790058057755232
Training steps: 20 Loss: 0.011808047071099281
Training steps: 40 Loss: 0.012922742404043674
Training steps: 60 Loss: 0.0027748234570026398
Training steps: 80 Loss: 0.0008948363829404116
VALID ACC : 0.9764705896377563

model saved! vit_deit_base_distilled_patch16_224_fold3_0.97
Epoch 3/20
Training steps: 0 Loss: 0.0015448693884536624
Training steps: 20 Loss: 0.0006571502890437841
Training steps: 40 Loss: 0.0027235643938183784
Training steps: 60 Loss: 0.014082379639148712
Training steps: 80 Loss: 0.11243818700313568
VALID ACC : 0.9617647528648376

Epoch 4/20
Training steps: 0 Loss: 0.0008669137023389339
Training steps: 20 Loss: 0.15325620770454407
Training steps: 40 Loss: 0.024426553398370743
Training steps: 60 Loss: 0.003827356733381748
Training steps: 80 Loss: 0.022955361753702164
VALID ACC : 0.9441176652908325

Epoch 5/20
Training steps: 0 Loss: 0.0011019408702850342
Training steps: 20 Loss: 0.001850143657065928
Training steps: 40 Loss: 0.000597894424572587
Training steps: 60 Loss: 0.000499193905852735
Training steps: 80 Loss: 0.0005587120540440083
VALID ACC : 0.9676470756530762

Epoch 6/20
Training steps: 0 Loss: 0.0008818659116514027
Training steps: 20 Loss: 0.0005890672327950597
Training steps: 40 Loss: 0.000380369572667405
Training steps: 60 Loss: 0.00021075723634567112
Training steps: 80 Loss: 0.0003150541160721332
VALID ACC : 0.9735294580459595

Epoch     6: reducing learning rate of group 0 to 5.0000e-05.
Epoch 7/20
Training steps: 0 Loss: 0.00035146938171237707
Training steps: 20 Loss: 0.0002934508374892175
Training steps: 40 Loss: 0.000293821154627949
Training steps: 60 Loss: 0.00024732472957111895
Training steps: 80 Loss: 0.0001991338504012674
VALID ACC : 0.9735294580459595

EarlyStopping counter: 5 out of 5



Epoch 1/20
Training steps: 0 Loss: 2.0614306926727295
Training steps: 20 Loss: 0.21165452897548676
Training steps: 40 Loss: 0.15650087594985962
Training steps: 60 Loss: 0.09856440871953964
Training steps: 80 Loss: 0.25801852345466614
VALID ACC : 0.9528023600578308

model saved! vit_deit_base_distilled_patch16_224_fold4_0.95
Epoch 2/20
Training steps: 0 Loss: 0.003911319188773632
Training steps: 20 Loss: 0.013249287381768227
Training steps: 40 Loss: 0.0009106044308282435
Training steps: 60 Loss: 0.0026690915692597628
Training steps: 80 Loss: 0.01531395222991705
VALID ACC : 0.9498525261878967

Epoch 3/20
Training steps: 0 Loss: 0.0017143264412879944
Training steps: 20 Loss: 0.002851467579603195
Training steps: 40 Loss: 0.005741201341152191
Training steps: 60 Loss: 0.011219356209039688
Training steps: 80 Loss: 0.002695375354960561
VALID ACC : 0.9646017551422119

model saved! vit_deit_base_distilled_patch16_224_fold4_0.96
Epoch 4/20
Training steps: 0 Loss: 0.000511382007971406
Training steps: 20 Loss: 0.0013728523626923561
Training steps: 40 Loss: 0.0010254302760586143
Training steps: 60 Loss: 0.0003245208063162863
Training steps: 80 Loss: 0.0008100293343886733
VALID ACC : 0.9793510437011719

model saved! vit_deit_base_distilled_patch16_224_fold4_0.97
Epoch 5/20
Training steps: 0 Loss: 0.0006320927059277892
Training steps: 20 Loss: 0.0003737296792678535
Training steps: 40 Loss: 0.00028914542053826153
Training steps: 60 Loss: 0.0009214243618771434
Training steps: 80 Loss: 0.0002785417309496552
VALID ACC : 0.976401150226593

Epoch 6/20
Training steps: 0 Loss: 0.000439854251453653
Training steps: 20 Loss: 0.00021529578953050077
Training steps: 40 Loss: 0.00015027981135062873
Training steps: 60 Loss: 0.00014621201262343675
Training steps: 80 Loss: 0.0001887156249722466
VALID ACC : 0.976401150226593

Epoch 7/20
Training steps: 0 Loss: 0.00013472522550728172
Training steps: 20 Loss: 0.00019639963284134865
Training steps: 40 Loss: 0.00017765158554539084
Training steps: 60 Loss: 0.0002927388995885849
Training steps: 80 Loss: 0.0001379675231873989
VALID ACC : 0.976401150226593

Epoch 8/20
Training steps: 0 Loss: 0.00033308877027593553
Training steps: 20 Loss: 0.00018385719158686697
Training steps: 40 Loss: 0.0002205742202932015
Training steps: 60 Loss: 0.0002880125539377332
Training steps: 80 Loss: 0.00019521107606124133
VALID ACC : 0.976401150226593

Epoch     8: reducing learning rate of group 0 to 5.0000e-05.
Epoch 9/20
Training steps: 0 Loss: 0.0002322245854884386
Training steps: 20 Loss: 0.00014505763829220086
Training steps: 40 Loss: 0.00020668211800511926
Training steps: 60 Loss: 0.00011255430581513792
Training steps: 80 Loss: 0.00016966232215054333
VALID ACC : 0.976401150226593

EarlyStopping counter: 5 out of 5



Epoch 1/20
Training steps: 0 Loss: 2.1110196113586426
Training steps: 20 Loss: 0.1107415184378624
Training steps: 40 Loss: 0.09970464557409286
Training steps: 60 Loss: 0.012141451239585876
Training steps: 80 Loss: 0.011343749240040779
VALID ACC : 0.9262536764144897

model saved! vit_deit_base_distilled_patch16_224_fold5_0.92
Epoch 2/20
Training steps: 0 Loss: 0.01333638560026884
Training steps: 20 Loss: 0.005270378198474646
Training steps: 40 Loss: 0.0025757476687431335
Training steps: 60 Loss: 0.007112572900950909
Training steps: 80 Loss: 0.0005468337913043797
VALID ACC : 0.9587020874023438

model saved! vit_deit_base_distilled_patch16_224_fold5_0.95
Epoch 3/20
Training steps: 0 Loss: 0.000730900326743722
Training steps: 20 Loss: 0.0010238775284960866
Training steps: 40 Loss: 0.000511404185090214
Training steps: 60 Loss: 0.00064817228121683
Training steps: 80 Loss: 0.00030412187334150076
VALID ACC : 0.9557521939277649

Epoch 4/20
Training steps: 0 Loss: 0.0005783180240541697
Training steps: 20 Loss: 0.00042364178807474673
Training steps: 40 Loss: 0.0002950250345747918
Training steps: 60 Loss: 0.0002105626481352374
Training steps: 80 Loss: 0.00021895382087677717
VALID ACC : 0.9557521939277649

Epoch 5/20
Training steps: 0 Loss: 0.0002488949103280902
Training steps: 20 Loss: 0.00030456739477813244
Training steps: 40 Loss: 0.0004265912575647235
Training steps: 60 Loss: 0.0003578931500669569
Training steps: 80 Loss: 0.0007239663973450661
VALID ACC : 0.9557521939277649

Epoch 6/20
Training steps: 0 Loss: 0.0002544424496591091
Training steps: 20 Loss: 0.0001761214662110433
Training steps: 40 Loss: 0.00019317405531182885
Training steps: 60 Loss: 0.00023175563546828926
Training steps: 80 Loss: 0.00015963845362421125
VALID ACC : 0.9557521939277649

Epoch     6: reducing learning rate of group 0 to 5.0000e-05.
Epoch 7/20
Training steps: 0 Loss: 0.00016027921810746193
Training steps: 20 Loss: 0.00013730306818615645
Training steps: 40 Loss: 9.590276749804616e-05
Training steps: 60 Loss: 0.00013750368088949472
Training steps: 80 Loss: 7.908276893431321e-05
VALID ACC : 0.9557521939277649

EarlyStopping counter: 5 out of 5

soft voting¶

In [133]:

torch.cuda.empty_cache()
gc.collect()

Out[133]:

In [165]:

models = [
        #   "vit_deit_base_distilled_patch16_224_fold1_pseudo_0.97",
          "vit_deit_base_distilled_patch16_224_0.97",
          "vit_deit_base_distilled_patch16_224_0.99",
          "vit_deit_base_distilled_patch16_224_fold1_0.98",
          "vit_deit_base_distilled_patch16_224_fold2_0.98",
          "vit_deit_base_distilled_patch16_224_fold3_0.97",
          "vit_deit_base_distilled_patch16_224_fold4_0.97",
          "vit_deit_base_distilled_patch16_224_fold5_0.95"]

models = [torch.load(os.path.join(args.model_path, model)).eval() for model in models]

In [166]:

args.batch_size = 2
test_dataset = ArtDataset(test_df, transform)
test_loader = DataLoader(test_dataset,
                         shuffle=False,
                         num_workers=args.num_workers,
                         batch_size=args.batch_size,
                         pin_memory=args.pin_memory)

In [167]:

answers = []
for images, labels in test_loader:
    images = images.to(args.device)
    labels = labels.to(args.device)

    predicts = torch.zeros(images.size(0), 7)
    for model in models:
        outputs = model(images)
        outputs = F.softmax(outputs.cpu(), dim=1)
        predicts += outputs

    # prediction들의 average 계산
    predict_avg = predicts / len(models)

    _, preds = torch.max(predict_avg, 1)

    answers.extend(list(preds.cpu().numpy()))

In [168]:

print(answers)

[2, 3, 3, 3, 3, 3, 4, 4, 3, 1, 6, 2, 6, 2, 3, 1, 2, 0, 6, 3, 3, 5, 2, 3, 0, 5, 1, 2, 0, 5, 1, 5, 6, 2, 0, 5, 5, 4, 2, 1, 4, 0, 2, 3, 1, 3, 0, 5, 5, 2, 6, 5, 4, 1, 5, 0, 4, 5, 1, 0, 5, 0, 6, 1, 1, 2, 4, 1, 1, 3, 1, 3, 0, 1, 1, 6, 2, 0, 3, 4, 1, 6, 1, 6, 6, 4, 3, 6, 1, 2, 2, 5, 1, 0, 5, 6, 3, 3, 1, 6, 5, 6, 6, 0, 3, 2, 5, 3, 0, 0, 4, 6, 2, 5, 4, 2, 0, 0, 5, 6, 4, 2, 2, 6, 4, 1, 5, 6, 0, 4, 1, 1, 6, 4, 4, 2, 5, 6, 5, 0, 4, 3, 1, 5, 1, 5, 4, 0, 2, 5, 6, 1, 6, 3, 2, 2, 0, 1, 4, 5, 2, 4, 6, 2, 3, 4, 1, 5, 6, 2, 1, 5, 3, 4, 0, 3, 2, 5, 3, 4, 2, 0, 3, 6, 0, 3, 3, 2, 0, 4, 4, 2, 2, 4, 4, 6, 6, 6, 3, 2, 5, 5, 4, 6, 2, 1, 3, 6, 0, 2, 3, 1, 1, 3, 1, 5, 2, 2, 0, 0, 0, 1, 2, 2, 6, 6, 3, 2, 5, 4, 3, 5, 0, 0, 4, 5, 0, 6, 4, 1, 6, 1, 2, 1, 6, 0, 5, 4, 6, 3, 4, 2, 5, 3, 1, 4, 2, 3, 4, 0, 6, 0, 2, 0, 1, 0, 1, 4, 0, 1, 1, 5, 6, 4, 6, 2, 4, 4, 3, 6, 6, 4, 0, 3, 1, 3, 0, 0, 5, 2, 5, 5, 2, 5, 0, 6, 5, 1, 5, 1, 5, 3, 0, 3, 3, 1, 6, 5, 0, 4, 0, 4, 3, 4, 4, 2, 5, 1, 2, 0, 6, 6, 1, 3, 3, 0, 0, 5, 2, 4, 6, 6, 6, 3, 6, 0, 4, 6, 4, 0, 2, 0, 5, 5, 1, 6, 3, 3, 5, 1]

In [169]:

submission_path = os.path.join(PROJECT_PATH, 'submission.csv')
submission_df = pd.DataFrame({'answer_value': answers})
submission_df.to_csv(submission_path)

In [170]:

submission_df.value_counts()

Out[170]:

answer_value
6               52
0               51
5               50
3               50
2               50
1               50
4               47
dtype: int64

In [149]:

submission_df.value_counts()

Out[149]:

answer_value
6               52
2               51
5               50
3               50
1               50
0               50
4               47
dtype: int64

'취업 이야기 > 데브매칭 문제 해설' 카테고리의 다른 글

2021 앱 Dev-Matching \| K-MOOC 강좌정보 서비스 (0)	2021.07.20
2021 ML Dev-Matching \| 미술 작품 분류하기 : 우수 코드 공개 (0)	2021.07.06
2021 ML Dev-Matching \| 미술 작품 분류하기 : 출제자가 추천한 우수 코드 B (0)	2021.07.06
2021 ML Dev-Matching \| 미술 작품 분류하기 : 출제자가 추천한 우수 코드 A (0)	2021.07.06
'2021 Dev-Matching: 웹 백엔드 개발자(상반기)' 기출 문제 해설 (0)	2021.04.26

ABOUT ME

프로그래머스 공식 블로그

목차

2021 Dev-Matching: 머신러닝 과제테스트¶

환경 설정¶

데이터¶

데이터 lookup 테이블 만들기¶

EDA¶

Dataset & DataLoader¶

train / valid 나누기¶

get_loader¶

모델¶

conv 모델¶

efficientnet1¶

timm¶

훈련¶

초기 세팅¶

criterion¶

get_model¶

get_optimizer¶

get_scheduler¶

train¶

validate¶

pseudo_labeling¶

run¶

데이터 불러오기¶

훈련 진행¶

추론¶

앙상블¶

kfold 훈련¶

soft voting¶

'취업 이야기 > 데브매칭 문제 해설' 카테고리의 다른 글

티스토리툴바

ABOUT ME

목차

2021 Dev-Matching: 머신러닝 과제테스트¶

환경 설정¶

데이터¶

데이터 lookup 테이블 만들기¶

EDA¶

Dataset & DataLoader¶

train / valid 나누기¶

get_loader¶

모델¶

conv 모델¶

efficientnet1¶

timm¶

훈련¶

초기 세팅¶

criterion¶

get_model¶

get_optimizer¶

get_scheduler¶

train¶

validate¶

pseudo_labeling¶

run¶

데이터 불러오기¶

훈련 진행¶

추론¶

앙상블¶

kfold 훈련¶

soft voting¶

'취업 이야기 > 데브매칭 문제 해설' 카테고리의 다른 글

관련글 관련글 더보기

티스토리툴바