keras

Tweet

画像分類のプログラム

cudaとかcudnnとかtensorflowのバージョンによって書き方が異なるのがしんどい。
以下はwindows版。

• tensorflow-gpu==1.13.1
• keras-nightly==2.5.0.dev2021032900
• cudaは10.0 cuDNNは7.4

# coding: UTF-8
import os
import csv
import random
import numpy as np
from PIL import Image
from keras.models import Sequential, Model
from keras.layers import Dense, Activation, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D, BatchNormalization
from keras.callbacks import EarlyStopping, ModelCheckpoint, CSVLogger
from keras.utils import to_categorical
from keras.optimizers import SGD, Adam


def load_train_data():
    train_data_list = []
    with open(r"./data/train_master.tsv", "r") as f:
        reader = csv.reader(f, delimiter='\t')
        # ヘッダスキップ
        next(reader)
        for record in reader:
            full_path = os.path.join(r"./data/train", record[0])
            label = record[1]
            train_data_list.append([full_path, label])

    return train_data_list


def train_data_generator(train_data_list, batch_size=32, class_num=20):
    max_len = len(train_data_list)
    while True:
        rtn_images = []
        rtn_labels = []
        for i in range(batch_size):
            index = random.randint(0, max_len-1)
            file_path = train_data_list[index][0]
            label = train_data_list[index][1]
            img = Image.open(file_path)
            img = np.asarray(img)/255.
            label = to_categorical(label, num_classes=class_num)
            rtn_images.append(img)
            rtn_labels.append(label)

        yield np.array(rtn_images), np.array(rtn_labels)


def my_model():
    model = Sequential()
    model.add(Conv2D(filters=32, kernel_size=(3, 3), padding='same', input_shape=(32, 32, 3)))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(Conv2D(filters=32, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(3, 3), padding='same')) # 16×16

    model.add(Conv2D(filters=64, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(Conv2D(filters=64, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(3, 3), padding='same')) # 8×8

    model.add(Conv2D(filters=128, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(Conv2D(filters=128, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(3, 3), padding='same')) # 4×4

    model.add(Flatten())

    model.add(Dense(1024))
    model.add(Activation('relu'))
    model.add(Dense(20))
    model.add(Activation('softmax'))

    return model


def main():
    batch_size = 32
    save_weight_file = 'practice.{epoch:02d}-{loss:.2f}-{acc:.2f}.hdf5'
    weights_path = r"./weights"
    csv_dir = r"./csv"
    weight_path = os.path.join(weights_path, save_weight_file)
    csv_path = os.path.join(csv_dir, "log.csv")
    if not os.path.exists(weights_path):
        os.makedirs(weights_path)

    if not os.path.exists(csv_dir):
        os.makedirs(csv_dir)

    train_data_list = load_train_data()
    model = my_model()
    model.compile(
        optimizer=Adam(lr=1e-3),
        loss="categorical_crossentropy",
        metrics=["accuracy"]
    )
    model.summary()

    cb_weight = ModelCheckpoint(filepath=weight_path, monitor='acc', verbose=1, save_best_only=False, mode='auto')
    cb_csv = CSVLogger(csv_path)
    model.fit_generator(
        train_data_generator(train_data_list, batch_size=batch_size),
        steps_per_epoch=int(50000/batch_size),
        epochs=100,
        callbacks=[cb_weight, cb_csv]
    )


if __name__ == '__main__':
    main()

以下はLinux版。

• tensorflow==2.3.0
• kerasはtensorflowをインストールしたら自動的に入るものを使用
• export LD_LIBRARY_PATH=/usr/local/cuda-10.1/lib64

# coding: UTF-8
import os
import csv
import random
import numpy as np
from PIL import Image
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Dense, Activation, Dropout, Flatten
from tensorflow.keras.layers import BatchNormalization
from tensorflow.keras.layers import Conv2D, MaxPooling2D
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, CSVLogger
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.optimizers import SGD, Adam


def load_train_data():
    train_data_list = []
    with open(r"./data/train_master.tsv", "r") as f:
        reader = csv.reader(f, delimiter='\t')
        # ヘッダスキップ
        next(reader)
        for record in reader:
            full_path = os.path.join(r"./data/train", record[0])
            label = record[1]
            train_data_list.append([full_path, label])

    return train_data_list


def train_data_generator(train_data_list, batch_size=32, class_num=20):
    max_len = len(train_data_list)
    while True:
        rtn_images = []
        rtn_labels = []
        for i in range(batch_size):
            index = random.randint(0, max_len-1)
            file_path = train_data_list[index][0]
            label = train_data_list[index][1]
            img = Image.open(file_path)
            img = np.asarray(img)/255.
            label = to_categorical(label, num_classes=class_num)
            rtn_images.append(img)
            rtn_labels.append(label)

        yield np.array(rtn_images), np.array(rtn_labels)


def my_model():
    model = Sequential()
    model.add(Conv2D(filters=32, kernel_size=(3, 3), padding='same', input_shape=(32, 32, 3)))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(Conv2D(filters=32, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(3, 3), padding='same')) # 16×16

    model.add(Conv2D(filters=64, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(Conv2D(filters=64, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(3, 3), padding='same')) # 8×8

    model.add(Conv2D(filters=128, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(Conv2D(filters=128, kernel_size=(3, 3), padding='same'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(3, 3), padding='same')) # 4×4

    model.add(Flatten())

    model.add(Dense(1024))
    model.add(Activation('relu'))
    model.add(Dense(20))
    model.add(Activation('softmax'))

    return model


def main():
    batch_size = 256
    save_weight_file = "practice.{epoch:02d}-{loss:.2f}-{accuracy:.2f}.hdf5"
    weights_path = r"./weights"
    csv_dir = r"./csv"
    weight_path = os.path.join(weights_path, save_weight_file)
    csv_path = os.path.join(csv_dir, "log.csv")
    if not os.path.exists(weights_path):
        os.makedirs(weights_path)

    if not os.path.exists(csv_dir):
        os.makedirs(csv_dir)

    train_data_list = load_train_data()
    model = my_model()
    model.compile(
        optimizer=Adam(lr=1e-3),
        loss="categorical_crossentropy",
        metrics=["accuracy"]
    )
    model.summary()

    cb_weight = ModelCheckpoint(filepath=weight_path, monitor='acc', verbose=1, save_best_only=False, mode='auto')
    cb_csv = CSVLogger(csv_path)
    model.fit_generator(
        train_data_generator(train_data_list, batch_size=batch_size),
        steps_per_epoch=int(50000/batch_size),
        epochs=100,
        callbacks=[cb_weight, cb_csv]
    )


if __name__ == '__main__':
    main()