EfficientNetB0使ったベースライン紹介

skywalker

EfficientNetB0使ったベースライン紹介

Efficient Baseline

Google Colabで回しています。
pytorch EfficientNetB0でのベースラインモデル作成までを紹介します。

Settings

!nvidia-smi

Tue Jun 29 09:16:33 2021       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 465.27       Driver Version: 460.32.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla P100-PCIE...  Off  | 00000000:00:04.0 Off |                    0 |
| N/A   41C    P0    27W / 250W |      0MiB / 16280MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

!pip uninstall albumentations -y

Uninstalling albumentations-1.0.0:
  Successfully uninstalled albumentations-1.0.0

!pip install timm -q
!pip install albumentations -q

from google.colab import drive
drive.mount('/content/drive')

Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).

import warnings
warnings.simplefilter('ignore')

import os
import gc
gc.enable()
import sys
import glob
import time
import random

import numpy as np
import pandas as pd
from PIL import Image
import matplotlib.pyplot as plt
%matplotlib inline

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, models, transforms
from sklearn.model_selection import GroupKFold, StratifiedKFold

from datetime import datetime
from tqdm.autonotebook import tqdm as tqdm

import albumentations as A
from albumentations.pytorch import ToTensorV2
from sklearn.metrics import roc_auc_score

import timm

class CFG:
    def __init__(self):
        
        self.debug=False
        self.num_workers=4
        self.model_name='efficientnet_b0'
        self.size=224
        self.scheduler='CosineAnnealingLR' 
        self.epochs= 20
        self.T_max=4
        self.lr=1e-3
        self.min_lr=1e-4
        self.batch_size= 16
        self.weight_decay=5e-5
        self.dropout=0.5
        self.seed=42
        self.n_fold=5
        self.alpha = 1.0
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

CONFIG = CFG()

def seed_everything(seed:int==42):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True

seed_everything(CONFIG.seed)

Data Load

## Google Driveに保存されているディレクトリ指定
INPUT_DIR = "/content/drive/MyDrive/probspace/religion/data/raw/"
BASE_DIR = "/content/drive/MyDrive/probspace/religion/data/interim/"

MODEL_DIR = f'{BASE_DIR}for_topic/'
OUTPUT_DIR = f'{BASE_DIR}for_topic/'

os.makedirs(MODEL_DIR, exist_ok=True)
os.makedirs(OUTPUT_DIR, exist_ok=True)

file_list = sorted(glob.glob(os.path.join(INPUT_DIR, '*')))

for i, file in enumerate(file_list):
    
    print(f'{i} | {file}')

0 | /content/drive/MyDrive/probspace/religion/data/raw/christ-test-imgs.npz
1 | /content/drive/MyDrive/probspace/religion/data/raw/christ-train-imgs.npz
2 | /content/drive/MyDrive/probspace/religion/data/raw/christ-train-labels.npz

train_label = np.load("/content/drive/MyDrive/probspace/religion/data/raw/christ-train-labels.npz")['arr_0']
train_image = np.load("/content/drive/MyDrive/probspace/religion/data/raw/christ-train-imgs.npz")['arr_0']
test_image = np.load("/content/drive/MyDrive/probspace/religion/data/raw/christ-test-imgs.npz")['arr_0']

train_label.shape, train_image.shape, test_image.shape

((654,), (654, 224, 224, 3), (497, 224, 224, 3))

plt.figure(figsize=(12, 2))
for i in range(5):
    plt.subplot(1, 5, i+1)
    plt.imshow(train_image[i], aspect='auto')
    plt.title(train_label[i])
plt.show()

print(f'クラス数: {np.unique(train_label).size}')

クラス数: 13

Training

DataSet

class TrainDataset:
    
    def __init__(self, image, target, transform=None):
        
        self.image = image
        self.target = target
        self.transform = transform
        
    def __len__(self):
        return len(self.image)
    
    def __getitem__(self, item):
        
        image = self.image[item]
        target = self.target[item]
        
        if self.transform is not None:
            transformed = self.transform(image = image)
            image = transformed['image']
            
        else:
            image = image[np.newaxis, :, :]

        return {
            'image' : torch.tensor(image, dtype=torch.float),
            'target' : torch.tensor(target, dtype=torch.float).long(),
        }

class TestDataset:
    
    def __init__(self, image, transform=None):
        
        self.image = image
        self.transform = transform
        
    def __len__(self):
        return len(self.image)
    
    def __getitem__(self, item):
        
        image = self.image[item]
        
        if self.transform is not None:
            transformed = self.transform(image = image)
            image = transformed['image']
            
        else:
            image = image[np.newaxis, :, :]

        return {
            'image' : torch.tensor(image, dtype=torch.float),
        }

def get_transforms(*, data):
    
    if data == 'train':
        return A.Compose([
            A.Resize(CONFIG.size, CONFIG.size),
            A.VerticalFlip(p=0.5),
            A.HorizontalFlip(p=0.5),
            A.Rotate(p=0.3),
            A.ToSepia(p=0.3),
            A.Normalize(),
            ToTensorV2(),
        ])
    
    elif data == 'valid':
        return A.Compose([
            A.Resize(CONFIG.size, CONFIG.size),
            A.Normalize(),
            ToTensorV2()
        ])

Model

class Model(nn.Module):
    
    def __init__(self, CONFIG, pretrained=False):
        super().__init__()
        self.CONFIG = CONFIG
        self.backbone = timm.create_model(self.CONFIG.model_name, 
                                          pretrained=pretrained,
                                          in_chans=3)
        
        self.n_features = self.backbone.classifier.in_features
        self.backbone.classifier = nn.Linear(self.n_features, 13)
        
        
    def forward(self, x):
        output = self.backbone(x)
        output = F.log_softmax(output, dim = 1)
        return output

Helper Function

def train_func(model, optimizer, scheduler, loss_fn, dataloader, device):
    
    model.train()
    final_loss = 0
    
    for step, data in enumerate(dataloader):
        
        s_t = time.time()
        optimizer.zero_grad()
        
        image = data['image'].to(device)
        label = data['target'].to(device)
        batch_size = label.size(0)
        
        preds = model(image)

        loss = loss_fn(preds, label)

        e_t = time.time()
        e_t = e_t - s_t

        loss.backward()
        optimizer.step()
        scheduler.step()
        
        final_loss += loss.item()
        
    final_loss /= len(dataloader)
    
    return final_loss

def valid_func(model, loss_fn, dataloader, device):
    
    model.eval()

    final_loss = 0
    valid_preds = []

    for step, data in enumerate(dataloader):
        
        s_t = time.time()
        image = data['image'].to(device)
        label = data['target'].to(device)

        preds = model(image)
            
        loss = loss_fn(preds, label)
        e_t = time.time()
        e_t = e_t - s_t

        valid_preds.append(preds.sigmoid().detach().cpu().numpy())
        final_loss += loss.item()
        
    final_loss /= len(dataloader)
    valid_preds = np.concatenate(valid_preds)
    
    return final_loss, valid_preds

def test_func(model, dataloader, device):
    
    model.eval()
        
    final_loss = 0
    test_preds = []

    for step, data in enumerate(dataloader):
        
        s_t = time.time()
        image = data['image'].to(device)

        preds = model(image)
            
        e_t = time.time()
        e_t = e_t - s_t
        
        test_preds.append(preds.sigmoid().detach().cpu().numpy())
                
    test_preds = np.concatenate(test_preds)
    
    return test_preds

from sklearn.metrics import accuracy_score

CV = StratifiedKFold(n_splits=CONFIG.n_fold, random_state=CONFIG.seed)

oof = np.zeros((train_image.shape[0], 13))
pred = np.zeros((test_image.shape[0], 13))

device = CONFIG.device

for fold, (tr, te) in enumerate(CV.split(train_image, train_label)):

    print(f'==================== Fold {fold+1} ======================')

    tr_image = train_image[tr]
    va_image = train_image[te]

    tr_target = train_label[tr]
    va_target = train_label[te]


    train_dataset = TrainDataset(tr_image, tr_target, 
                                    get_transforms(data='train'))
    valid_dataset = TrainDataset(va_image, va_target,
                                    get_transforms(data='valid'))

    train_loader = torch.utils.data.DataLoader(train_dataset,
                                                batch_size=CONFIG.batch_size,
                                                num_workers=CONFIG.num_workers,
                                                pin_memory=True,
                                                shuffle=True)

    valid_loader = torch.utils.data.DataLoader(valid_dataset,
                                                batch_size=CONFIG.batch_size,
                                                num_workers=CONFIG.num_workers,
                                                pin_memory=True,
                                                shuffle=False)
    

    model = Model(CONFIG, pretrained=True)
    model.to(device)

    optimizer = torch.optim.AdamW(model.parameters(),
                                  lr=CONFIG.lr,
                                  weight_decay=CONFIG.weight_decay)

    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
                                                         T_max=CONFIG.T_max,
                                                         eta_min=CONFIG.min_lr,
                                                         last_epoch=-1)

    loss_fn = nn.CrossEntropyLoss()

    best_score = 0

    for epoch in range(CONFIG.epochs):

        start_time = time.time()
        train_loss = train_func(model, optimizer, scheduler,
                                loss_fn, train_loader, device)
        valid_loss, valid_preds = valid_func(model, loss_fn,
                                                valid_loader, device)

        score = accuracy_score(va_target, np.argmax(valid_preds, axis=1))

        end_time = time.time()
        print(f"FOLD: {fold+1} | EPOCH:{epoch+1:02d} | train_loss:{train_loss:.6f} | valid_loss:{valid_loss:.6f} | valid_score:{score:.4f} | time:{end_time-start_time:.1f}s ")


        if score > best_score:
            best_score = score
            oof[te] = valid_preds
                
            MODEL_PATH = f"{MODEL_DIR}topic_001_baseline_{CONFIG.model_name}_fold{fold+1}.pth"
            torch.save(model.state_dict(), MODEL_PATH)

        else:
            continue


    del train_dataset, valid_dataset, train_loader, valid_loader, valid_preds
    _ = gc.collect()


    ## Predict
    model.load_state_dict(torch.load(MODEL_PATH))
    test_dataset = TestDataset(test_image,
                                 get_transforms(data='valid'))
    test_loader = torch.utils.data.DataLoader(test_dataset,
                                              batch_size=CONFIG.batch_size,
                                              num_workers=CONFIG.num_workers,
                                              pin_memory=True,
                                              shuffle=False)
    test_preds = test_func(model, test_loader, device)
    pred += test_preds/CONFIG.n_fold

    del test_dataset, test_loader, model

==================== Fold 1 ======================
FOLD: 1 | EPOCH:01 | train_loss:2.406092 | valid_loss:2.540015 | valid_score:0.2977 | time:3.3s 
FOLD: 1 | EPOCH:02 | train_loss:1.935311 | valid_loss:2.252315 | valid_score:0.3740 | time:3.2s 
FOLD: 1 | EPOCH:03 | train_loss:1.648426 | valid_loss:2.121224 | valid_score:0.3740 | time:3.2s 
FOLD: 1 | EPOCH:04 | train_loss:1.271097 | valid_loss:2.065577 | valid_score:0.3817 | time:3.2s 
FOLD: 1 | EPOCH:05 | train_loss:0.998648 | valid_loss:2.270356 | valid_score:0.3893 | time:3.2s 
FOLD: 1 | EPOCH:06 | train_loss:0.871209 | valid_loss:2.469064 | valid_score:0.3740 | time:3.2s 
FOLD: 1 | EPOCH:07 | train_loss:0.701129 | valid_loss:2.459864 | valid_score:0.3893 | time:3.2s 
FOLD: 1 | EPOCH:08 | train_loss:0.572170 | valid_loss:2.552534 | valid_score:0.4198 | time:3.2s 
FOLD: 1 | EPOCH:09 | train_loss:0.490011 | valid_loss:2.832892 | valid_score:0.3282 | time:3.2s 
FOLD: 1 | EPOCH:10 | train_loss:0.365210 | valid_loss:2.921633 | valid_score:0.4122 | time:3.2s 
FOLD: 1 | EPOCH:11 | train_loss:0.426756 | valid_loss:2.526095 | valid_score:0.4122 | time:3.2s 
FOLD: 1 | EPOCH:12 | train_loss:0.438959 | valid_loss:2.853215 | valid_score:0.3740 | time:3.2s 
FOLD: 1 | EPOCH:13 | train_loss:0.396685 | valid_loss:2.757360 | valid_score:0.3664 | time:3.2s 
FOLD: 1 | EPOCH:14 | train_loss:0.363250 | valid_loss:2.712663 | valid_score:0.3588 | time:3.2s 
FOLD: 1 | EPOCH:15 | train_loss:0.319701 | valid_loss:2.969607 | valid_score:0.3817 | time:3.2s 
FOLD: 1 | EPOCH:16 | train_loss:0.232301 | valid_loss:3.066341 | valid_score:0.3664 | time:3.3s 
FOLD: 1 | EPOCH:17 | train_loss:0.269452 | valid_loss:2.896516 | valid_score:0.3664 | time:3.2s 
FOLD: 1 | EPOCH:18 | train_loss:0.185510 | valid_loss:2.989266 | valid_score:0.3664 | time:3.2s 
FOLD: 1 | EPOCH:19 | train_loss:0.196089 | valid_loss:2.742342 | valid_score:0.4046 | time:3.2s 
FOLD: 1 | EPOCH:20 | train_loss:0.198678 | valid_loss:3.293876 | valid_score:0.3893 | time:3.3s 
==================== Fold 2 ======================
FOLD: 2 | EPOCH:01 | train_loss:2.416383 | valid_loss:2.253022 | valid_score:0.3053 | time:3.2s 
FOLD: 2 | EPOCH:02 | train_loss:1.898816 | valid_loss:2.123590 | valid_score:0.3511 | time:3.2s 
FOLD: 2 | EPOCH:03 | train_loss:1.507994 | valid_loss:2.143511 | valid_score:0.3435 | time:3.2s 
FOLD: 2 | EPOCH:04 | train_loss:1.223553 | valid_loss:2.027303 | valid_score:0.3969 | time:3.2s 
FOLD: 2 | EPOCH:05 | train_loss:0.951261 | valid_loss:1.979834 | valid_score:0.4351 | time:3.3s 
FOLD: 2 | EPOCH:06 | train_loss:0.773565 | valid_loss:2.253961 | valid_score:0.4046 | time:3.2s 
FOLD: 2 | EPOCH:07 | train_loss:0.693143 | valid_loss:2.060942 | valid_score:0.4046 | time:3.2s 
FOLD: 2 | EPOCH:08 | train_loss:0.630672 | valid_loss:2.476328 | valid_score:0.4122 | time:3.2s 
FOLD: 2 | EPOCH:09 | train_loss:0.502894 | valid_loss:2.269202 | valid_score:0.4504 | time:3.3s 
FOLD: 2 | EPOCH:10 | train_loss:0.441415 | valid_loss:2.433585 | valid_score:0.4427 | time:3.3s 
FOLD: 2 | EPOCH:11 | train_loss:0.454266 | valid_loss:2.629570 | valid_score:0.4504 | time:3.3s 
FOLD: 2 | EPOCH:12 | train_loss:0.371607 | valid_loss:2.577382 | valid_score:0.3893 | time:3.3s 
FOLD: 2 | EPOCH:13 | train_loss:0.329626 | valid_loss:2.509102 | valid_score:0.4122 | time:3.3s 
FOLD: 2 | EPOCH:14 | train_loss:0.315804 | valid_loss:2.751275 | valid_score:0.4427 | time:3.2s 
FOLD: 2 | EPOCH:15 | train_loss:0.216425 | valid_loss:2.791086 | valid_score:0.4656 | time:3.2s 
FOLD: 2 | EPOCH:16 | train_loss:0.225850 | valid_loss:2.615117 | valid_score:0.5038 | time:3.2s 
FOLD: 2 | EPOCH:17 | train_loss:0.236258 | valid_loss:2.922496 | valid_score:0.4351 | time:3.3s 
FOLD: 2 | EPOCH:18 | train_loss:0.240287 | valid_loss:3.070802 | valid_score:0.3969 | time:3.2s 
FOLD: 2 | EPOCH:19 | train_loss:0.224027 | valid_loss:2.906561 | valid_score:0.4198 | time:3.3s 
FOLD: 2 | EPOCH:20 | train_loss:0.245791 | valid_loss:3.438557 | valid_score:0.4122 | time:3.2s 
==================== Fold 3 ======================
FOLD: 3 | EPOCH:01 | train_loss:2.413378 | valid_loss:2.199042 | valid_score:0.3664 | time:3.2s 
FOLD: 3 | EPOCH:02 | train_loss:1.944740 | valid_loss:1.924229 | valid_score:0.3969 | time:3.3s 
FOLD: 3 | EPOCH:03 | train_loss:1.635726 | valid_loss:2.430089 | valid_score:0.4351 | time:3.3s 
FOLD: 3 | EPOCH:04 | train_loss:1.328843 | valid_loss:2.144016 | valid_score:0.4275 | time:3.3s 
FOLD: 3 | EPOCH:05 | train_loss:1.084501 | valid_loss:1.990621 | valid_score:0.4198 | time:3.2s 
FOLD: 3 | EPOCH:06 | train_loss:0.836173 | valid_loss:2.582394 | valid_score:0.4504 | time:3.3s 
FOLD: 3 | EPOCH:07 | train_loss:0.732399 | valid_loss:2.284314 | valid_score:0.4733 | time:3.3s 
FOLD: 3 | EPOCH:08 | train_loss:0.541313 | valid_loss:2.446321 | valid_score:0.4504 | time:3.3s 
FOLD: 3 | EPOCH:09 | train_loss:0.513164 | valid_loss:2.397102 | valid_score:0.4351 | time:3.2s 
FOLD: 3 | EPOCH:10 | train_loss:0.425108 | valid_loss:2.654123 | valid_score:0.4733 | time:3.3s 
FOLD: 3 | EPOCH:11 | train_loss:0.358090 | valid_loss:2.897399 | valid_score:0.4275 | time:3.2s 
FOLD: 3 | EPOCH:12 | train_loss:0.414298 | valid_loss:2.652705 | valid_score:0.4809 | time:3.3s 
FOLD: 3 | EPOCH:13 | train_loss:0.318489 | valid_loss:2.737102 | valid_score:0.4351 | time:3.3s 
FOLD: 3 | EPOCH:14 | train_loss:0.262174 | valid_loss:2.565899 | valid_score:0.5115 | time:3.3s 
FOLD: 3 | EPOCH:15 | train_loss:0.310047 | valid_loss:2.982922 | valid_score:0.4809 | time:3.2s 
FOLD: 3 | EPOCH:16 | train_loss:0.329359 | valid_loss:3.046944 | valid_score:0.4351 | time:3.2s 
FOLD: 3 | EPOCH:17 | train_loss:0.322574 | valid_loss:2.811296 | valid_score:0.4656 | time:3.3s 
FOLD: 3 | EPOCH:18 | train_loss:0.253226 | valid_loss:2.694476 | valid_score:0.4885 | time:3.3s 
FOLD: 3 | EPOCH:19 | train_loss:0.196119 | valid_loss:3.254169 | valid_score:0.4427 | time:3.3s 
FOLD: 3 | EPOCH:20 | train_loss:0.200497 | valid_loss:2.784538 | valid_score:0.4122 | time:3.3s 
==================== Fold 4 ======================
FOLD: 4 | EPOCH:01 | train_loss:2.422234 | valid_loss:2.397270 | valid_score:0.2366 | time:3.3s 
FOLD: 4 | EPOCH:02 | train_loss:1.934286 | valid_loss:2.099750 | valid_score:0.3588 | time:3.2s 
FOLD: 4 | EPOCH:03 | train_loss:1.521994 | valid_loss:2.600684 | valid_score:0.3359 | time:3.3s 
FOLD: 4 | EPOCH:04 | train_loss:1.179104 | valid_loss:2.075456 | valid_score:0.4198 | time:3.2s 
FOLD: 4 | EPOCH:05 | train_loss:0.887440 | valid_loss:2.269339 | valid_score:0.3435 | time:3.3s 
FOLD: 4 | EPOCH:06 | train_loss:0.775537 | valid_loss:2.328922 | valid_score:0.3969 | time:3.3s 
FOLD: 4 | EPOCH:07 | train_loss:0.609571 | valid_loss:2.209479 | valid_score:0.4046 | time:3.3s 
FOLD: 4 | EPOCH:08 | train_loss:0.484529 | valid_loss:2.402753 | valid_score:0.4504 | time:3.2s 
FOLD: 4 | EPOCH:09 | train_loss:0.446730 | valid_loss:2.492520 | valid_score:0.4198 | time:3.2s 
FOLD: 4 | EPOCH:10 | train_loss:0.493044 | valid_loss:2.506052 | valid_score:0.4351 | time:3.3s 
FOLD: 4 | EPOCH:11 | train_loss:0.525597 | valid_loss:2.753054 | valid_score:0.3740 | time:3.3s 
FOLD: 4 | EPOCH:12 | train_loss:0.414995 | valid_loss:2.700784 | valid_score:0.4427 | time:3.3s 
FOLD: 4 | EPOCH:13 | train_loss:0.358577 | valid_loss:2.480349 | valid_score:0.4809 | time:3.3s 
FOLD: 4 | EPOCH:14 | train_loss:0.315230 | valid_loss:2.585382 | valid_score:0.4809 | time:3.3s 
FOLD: 4 | EPOCH:15 | train_loss:0.259825 | valid_loss:2.581717 | valid_score:0.4580 | time:3.3s 
FOLD: 4 | EPOCH:16 | train_loss:0.272259 | valid_loss:2.791074 | valid_score:0.3893 | time:3.2s 
FOLD: 4 | EPOCH:17 | train_loss:0.343921 | valid_loss:3.005155 | valid_score:0.4122 | time:3.3s 
FOLD: 4 | EPOCH:18 | train_loss:0.258963 | valid_loss:2.672385 | valid_score:0.4122 | time:3.3s 
FOLD: 4 | EPOCH:19 | train_loss:0.248019 | valid_loss:2.963132 | valid_score:0.4351 | time:3.2s 
FOLD: 4 | EPOCH:20 | train_loss:0.260328 | valid_loss:2.786731 | valid_score:0.4122 | time:3.2s 
==================== Fold 5 ======================
FOLD: 5 | EPOCH:01 | train_loss:2.401482 | valid_loss:2.235001 | valid_score:0.2923 | time:3.3s 
FOLD: 5 | EPOCH:02 | train_loss:2.010715 | valid_loss:2.199522 | valid_score:0.3462 | time:3.3s 
FOLD: 5 | EPOCH:03 | train_loss:1.592346 | valid_loss:2.140982 | valid_score:0.3615 | time:3.3s 
FOLD: 5 | EPOCH:04 | train_loss:1.227384 | valid_loss:1.992489 | valid_score:0.4154 | time:3.3s 
FOLD: 5 | EPOCH:05 | train_loss:0.984638 | valid_loss:2.271417 | valid_score:0.3615 | time:3.2s 
FOLD: 5 | EPOCH:06 | train_loss:0.755655 | valid_loss:2.478073 | valid_score:0.4308 | time:3.2s 
FOLD: 5 | EPOCH:07 | train_loss:0.726328 | valid_loss:2.203939 | valid_score:0.4538 | time:3.3s 
FOLD: 5 | EPOCH:08 | train_loss:0.689626 | valid_loss:2.559072 | valid_score:0.4385 | time:3.3s 
FOLD: 5 | EPOCH:09 | train_loss:0.457736 | valid_loss:2.454785 | valid_score:0.4692 | time:3.3s 
FOLD: 5 | EPOCH:10 | train_loss:0.403398 | valid_loss:2.358973 | valid_score:0.4385 | time:3.3s 
FOLD: 5 | EPOCH:11 | train_loss:0.362042 | valid_loss:2.395575 | valid_score:0.4077 | time:3.3s 
FOLD: 5 | EPOCH:12 | train_loss:0.387101 | valid_loss:2.071558 | valid_score:0.4846 | time:3.3s 
FOLD: 5 | EPOCH:13 | train_loss:0.405359 | valid_loss:2.509265 | valid_score:0.4923 | time:3.3s 
FOLD: 5 | EPOCH:14 | train_loss:0.339272 | valid_loss:2.563571 | valid_score:0.4077 | time:3.3s 
FOLD: 5 | EPOCH:15 | train_loss:0.336142 | valid_loss:2.695094 | valid_score:0.4308 | time:3.3s 
FOLD: 5 | EPOCH:16 | train_loss:0.306162 | valid_loss:2.540406 | valid_score:0.4077 | time:3.3s 
FOLD: 5 | EPOCH:17 | train_loss:0.274088 | valid_loss:2.501376 | valid_score:0.4846 | time:3.2s 
FOLD: 5 | EPOCH:18 | train_loss:0.224032 | valid_loss:2.473349 | valid_score:0.4385 | time:3.3s 
FOLD: 5 | EPOCH:19 | train_loss:0.228130 | valid_loss:2.214555 | valid_score:0.4538 | time:3.3s 
FOLD: 5 | EPOCH:20 | train_loss:0.201584 | valid_loss:2.768242 | valid_score:0.4308 | time:3.3s

oof_df = pd.DataFrame(columns=['target', 'pred'])
oof_df['target'] = train_label
oof_df['pred'] = np.argmax(oof, axis=1).astype(int)

for i in range(13):
    oof_df[f'pred_{i}'] = oof[:, i]

oof_score = accuracy_score(oof_df['target'], oof_df['pred'])
oof_df.to_csv(os.path.join(OUTPUT_DIR, f"001_topic_baseline_{oof_score:.5f}.csv"), index=False)
print(f'OOF_Score: {oof_score:.6f}')

display(oof_df.head())

OOF_Score: 0.481651

	target	pred	pred_0	pred_1	pred_2	pred_3	pred_4	pred_5	pred_6	pred_7	pred_8	pred_9	pred_10	pred_11	pred_12
0	5	0	0.426871	0.034438	0.004412	0.000526	0.000096	0.171342	0.005283	0.000192	0.000027	0.000892	0.000080	0.001074	0.000123
1	11	5	0.006598	0.040137	0.088920	0.003676	0.003733	0.410994	0.000544	0.000175	0.000176	0.021127	0.075572	0.019143	0.024378
2	8	8	0.000255	0.000837	0.006759	0.004273	0.001920	0.001171	0.002650	0.002013	0.492987	0.000207	0.003402	0.000732	0.003342
3	2	12	0.001176	0.000667	0.005934	0.000579	0.007171	0.015054	0.002181	0.000143	0.000196	0.006507	0.000722	0.002037	0.489082
4	6	6	0.028367	0.011791	0.186052	0.018460	0.073440	0.029008	0.188610	0.009138	0.015732	0.117745	0.049877	0.078932	0.068070

pred_df = pd.DataFrame(columns=['id', 'y'])
pred_df['id'] = [i+1 for i in range(test_image.shape[0])]
pred_df['y'] = np.argmax(pred, axis=1).astype('int')

for i in range(13):
    pred_df[f'pred_{i}'] = pred[:, i]

## Submit
pred_df[['id', 'y']].to_csv(os.path.join(OUTPUT_DIR, f"001_topic_baseline_submit_{oof_score:.5f}.csv"), index=False)

## for Ensemble
pred_df.to_csv(os.path.join(OUTPUT_DIR, f"001_topic_baseline_pred_{oof_score:.5f}.csv"), index=False)

display(pred_df.head())

	id	y	pred_0	pred_1	pred_2	pred_3	pred_4	pred_5	pred_6	pred_7	pred_8	pred_9	pred_10	pred_11	pred_12
0	1	5	0.063115	0.085578	0.128370	0.002261	0.004221	0.306202	0.075674	0.001421	0.000390	0.000843	0.001934	0.001913	0.000272
1	2	2	0.026866	0.024289	0.292164	0.136144	0.114584	0.001354	0.003841	0.005285	0.000213	0.003804	0.014728	0.007462	0.002829
2	3	10	0.001633	0.001689	0.070191	0.007089	0.056124	0.071473	0.001689	0.001037	0.015338	0.009341	0.392848	0.004888	0.028331
3	4	0	0.257828	0.009941	0.025205	0.101705	0.061008	0.038380	0.042961	0.003941	0.142874	0.015365	0.009926	0.003194	0.025524
4	5	1	0.059157	0.197781	0.076790	0.066885	0.016137	0.140100	0.048319	0.006185	0.006618	0.062361	0.064546	0.040767	0.017302

from sklearn.metrics import confusion_matrix
import seaborn as sns

cm = confusion_matrix(oof_df['target'], np.argmax(oof, axis=1).astype(int))

plt.figure(figsize=(18, 12))
sns.heatmap(cm, annot=True, fmt='d')
plt.xlim(12, 0)
plt.ylim(0, 12)
plt.show()

添付データ

religion_art_baseline_for_Beginner.ipynb?X-Amz-Expires=10800&X-Amz-Date=20241121T084738Z&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAIP7GCBGMWPMZ42PQ

EfficientNetB0使ったベースライン紹介

Settings

Data Load

Training

DataSet

Model

Helper Function

添付データ

new user