https://touch-sp.hatenablog.com/entry/2021/12/30/231939

データセットの作成

データはこちらからダウンロードさせて頂きました。

import glob
import random
import pandas as pd

cat_files = glob.glob('train/cat*')
dog_files = glob.glob('train/dog*')

cat_train = random.sample(cat_files, 10000)
dog_train = random.sample(dog_files, 10000)


cat_test = list(set(cat_files) - set(cat_train))
dog_test = list(set(dog_files) - set(dog_train))

train_dataset_list = []
for image_path in cat_train:
    train_dataset_list.append({
        'image': image_path,
        'label': 0
        })
for image_path in dog_train:
    train_dataset_list.append({
        'image': image_path,
        'label': 1
        })
train_df = pd.DataFrame(train_dataset_list)
train_df.to_pickle('train_df.pkl')


test_dataset_list = []
for image_path in cat_test:
    test_dataset_list.append({
        'image': image_path,
        'label': 0
        })
for image_path in dog_test:
    test_dataset_list.append({
        'image': image_path,
        'label': 1
        })
test_df = pd.DataFrame(test_dataset_list)
test_df.to_pickle('test_df.pkl')

学習

Pythonスクリプト

import pandas as pd
from autogluon.vision import ImageDataset
from autogluon.vision import ImagePredictor
from autogluon .core.space import Categorical

train_df = pd.read_pickle('train_df.pkl')

train_dataset = ImageDataset(train_df)

predictor = ImagePredictor()
model = Categorical('resnet101_v1d')
hyperparameters = {'model':model, 'batch_size':32, 'epochs': 5}
hyperparameter_tune_kwargs={'num_trials': 2}
predictor.fit(
    train_dataset, 
    hyperparameters = hyperparameters,
    hyperparameter_tune_kwargs = hyperparameter_tune_kwargs
    )
predictor.save('predictor.ag')

学習結果

Finished, total runtime is 2082.66 s
{ 'best_config': { 'estimator': <class 'gluoncv.auto.estimators.image_classification.image_classification.ImageClassificationEstimator'>,
                   'gpus': [0],
                   'img_cls': { 'batch_norm': False,
                                'last_gamma': False,
                                'model': 'resnet101_v1d',
                                'use_gn': False,
                                'use_pretrained': True,
                                'use_se': False},
                   'train': { 'batch_size': 32,
                              'crop_ratio': 0.875,
                              'data_dir': 'auto',
                              'dtype': 'float32',
                              'early_stop_baseline': -inf,
                              'early_stop_max_value': inf,
                              'early_stop_min_delta': 0.001,
                              'early_stop_patience': 10,
                              'epochs': 5,
                              'hard_weight': 0.5,
                              'input_size': 224,
                              'label_smoothing': False,
                              'log_interval': 50,
                              'lr': 0.01,
                              'lr_decay': 0.1,
                              'lr_decay_epoch': '40, 60',
                              'lr_decay_period': 0,
                              'lr_mode': 'step',
                              'mixup': False,
                              'mixup_alpha': 0.2,
                              'mixup_off_epoch': 0,
                              'mode': '',
                              'momentum': 0.9,
                              'no_wd': False,
                              'num_training_samples': -1,
                              'num_workers': 8,
                              'output_lr_mult': 0.1,
                              'pretrained_base': True,
                              'rec_train': 'auto',
                              'rec_train_idx': 'auto',
                              'rec_val': 'auto',
                              'rec_val_idx': 'auto',
                              'resume_epoch': 0,
                              'start_epoch': 0,
                              'teacher': None,
                              'temperature': 20,
                              'transfer_lr_mult': 0.01,
                              'use_rec': False,
                              'warmup_epochs': 0,
                              'warmup_lr': 0.0,
                              'wd': 0.0001},
                   'valid': {'batch_size': 32, 'num_workers': 8}},
  'total_time': 2082.6618723869324,
  'train_acc': 0.9678047153024911,
  'valid_acc': 0.99}

テストデータを用いた検証

Pythonスクリプト

import pandas as pd
from autogluon.vision import ImageDataset
from autogluon.vision import ImagePredictor

test_df = pd.read_pickle('test_df.pkl')

test_dataset = ImageDataset(test_df)

predictor = ImagePredictor.load('predictor.ag')

result = predictor.evaluate(test_dataset)

print('Top-1 test acc: %.3f' % result['top1'])

検証結果

Top-1 test acc: 0.994

推論

Pythonスクリプト

from autogluon.vision import ImagePredictor

predictor = ImagePredictor.load('predictor.ag')

proba = predictor.predict_proba('test1.jpg')

print(proba)

推論結果

          0         1
0  0.000141  0.999859

動作環境

Intel Core i7-7700K
RAM 32G
NVIDIA GTX 1080 (VRAM 8G)

Ubuntu 20.04 on WSL2
Python 3.8.10

mxnet-cu112==1.9.0
autogluon==0.3.2b20211230

追記

AutoGluonのAutoMMPredcitorを使う方法も別記事に書きました。
touch-sp.hatenablog.com