使用迁移学习进行图像识别

工程目录：

代码实现：

# -*- coding: utf-8 -*-

import glob
import os.path
import random
import numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfile

# Inception-v3模型瓶颈层的节点个数
BOTTLENECK_TENSOR_SIZE = 2048

# Inception-v3模型中代表瓶颈层结果的张量名称。
# 在谷歌提出的Inception-v3模型中，这个张量名称就是'pool_3/_reshape:0'。
# 在训练模型时，可以通过tensor.name来获取张量的名称。
BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'

# 图像输入张量所对应的名称。
JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'

# 下载的谷歌训练好的Inception-v3模型文件目录
MODEL_DIR = 'model/'

# 下载的谷歌训练好的Inception-v3模型文件名
MODEL_FILE = 'tensorflow_inception_graph.pb'

# 因为一个训练数据会被使用多次，所以可以将原始图像通过Inception-v3模型计算得到的特征向量保存在文件中，免去重复的计算。
# 下面的变量定义了这些文件的存放地址。
CACHE_DIR = 'tmp/bottleneck/'

# 图片数据文件夹。
# 在这个文件夹中每一个子文件夹代表一个需要区分的类别，每个子文件夹中存放了对应类别的图片。
INPUT_DATA = 'touxiang/'

# 验证的数据百分比
VALIDATION_PERCENTAGE = 10
# 测试的数据百分比
TEST_PERCENTAGE = 10

# 定义神经网络的设置
LEARNING_RATE = 0.01
STEPS = 4000
BATCH = 100

# 这个函数从数据文件夹中读取所有的图片列表并按训练、验证、测试数据分开。
# testing_percentage和validation_percentage参数指定了测试数据集和验证数据集的大小。
def create_image_lists(testing_percentage, validation_percentage):
    # 得到的所有图片都存在result这个字典(dictionary)里。
    # 这个字典的key为类别的名称，value也是一个字典，字典里存储了所有的图片名称。
    result = {}
    # 获取当前目录下所有的子目录
    sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]
    # 得到的第一个目录是当前目录，不需要考虑
    is_root_dir = True
    for sub_dir in sub_dirs:
        if is_root_dir:
            is_root_dir = False
            continue

        # 获取当前目录下所有的有效图片文件。
        extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']
        file_list = []
        dir_name = os.path.basename(sub_dir)
        for extension in extensions:
            file_glob = os.path.join(INPUT_DATA, dir_name, '*.'+extension)
            file_list.extend(glob.glob(file_glob))
        if not file_list:
            continue

        # 通过目录名获取类别的名称。
        label_name = dir_name.lower()
        # 初始化当前类别的训练数据集、测试数据集和验证数据集
        training_images = []
        testing_images = []
        validation_images = []
        for file_name in file_list:
            base_name = os.path.basename(file_name)
            # 随机将数据分到训练数据集、测试数据集和验证数据集。
            chance = np.random.randint(100)
            if chance < validation_percentage:
                validation_images.append(base_name)
            elif chance < (testing_percentage + validation_percentage):
                testing_images.append(base_name)
            else:
                training_images.append(base_name)

        # 将当前类别的数据放入结果字典。
        result[label_name] = {
            'dir': dir_name,
            'training': training_images,
            'testing': testing_images,
            'validation': validation_images
            }
    # 返回整理好的所有数据
    return result


# 这个函数通过类别名称、所属数据集和图片编号获取一张图片的地址。
# image_lists参数给出了所有图片信息。
# image_dir参数给出了根目录。存放图片数据的根目录和存放图片特征向量的根目录地址不同。
# label_name参数给定了类别的名称。
# index参数给定了需要获取的图片的编号。
# category参数指定了需要获取的图片是在训练数据集、测试数据集还是验证数据集。
def get_image_path(image_lists, image_dir, label_name, index, category):
    # 获取给定类别中所有图片的信息。
    label_lists = image_lists[label_name]
    # 根据所属数据集的名称获取集合中的全部图片信息。
    category_list = label_lists[category]
    mod_index = index % len(category_list)
    # 获取图片的文件名。
    base_name = category_list[mod_index]
    sub_dir = label_lists['dir']
    # 最终的地址为数据根目录的地址 + 类别的文件夹 + 图片的名称
    full_path = os.path.join(image_dir, sub_dir, base_name)
    return full_path


# 这个函数通过类别名称、所属数据集和图片编号获取经过Inception-v3模型处理之后的特征向量文件地址。
def get_bottlenect_path(image_lists, label_name, index, category):
    return get_image_path(image_lists, CACHE_DIR, label_name, index, category) + '.txt';


# 这个函数使用加载的训练好的Inception-v3模型处理一张图片，得到这个图片的特征向量。
def run_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):
    # 这个过程实际上就是将当前图片作为输入计算瓶颈张量的值。这个瓶颈张量的值就是这张图片的新的特征向量。
    bottleneck_values = sess.run(bottleneck_tensor, {image_data_tensor: image_data})
    # 经过卷积神经网络处理的结果是一个四维数组，需要将这个结果压缩成一个特征向量（一维数组）
    bottleneck_values = np.squeeze(bottleneck_values)
    return bottleneck_values


# 这个函数获取一张图片经过Inception-v3模型处理之后的特征向量。
# 这个函数会先试图寻找已经计算且保存下来的特征向量，如果找不到则先计算这个特征向量，然后保存到文件。
def get_or_create_bottleneck(sess, image_lists, label_name, index, category, jpeg_data_tensor, bottleneck_tensor):
    # 获取一张图片对应的特征向量文件的路径。
    label_lists = image_lists[label_name]
    sub_dir = label_lists['dir']
    sub_dir_path = os.path.join(CACHE_DIR, sub_dir)
    if not os.path.exists(sub_dir_path):
        os.makedirs(sub_dir_path)
    bottleneck_path = get_bottlenect_path(image_lists, label_name, index, category)
    # 如果这个特征向量文件不存在，则通过Inception-v3模型来计算特征向量，并将计算的结果存入文件。
    if not os.path.exists(bottleneck_path):
        # 获取原始的图片路径
        image_path = get_image_path(image_lists, INPUT_DATA, label_name, index, category)
        # 获取图片内容。
        image_data = gfile.FastGFile(image_path, 'rb').read()
        # print(len(image_data))
        # 由于输入的图片大小不一致，此处得到的image_data大小也不一致（已验证），但却都能通过加载的inception-v3模型生成一个2048的特征向量。具体原理不详。
        # 通过Inception-v3模型计算特征向量
        bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)
        # 将计算得到的特征向量存入文件
        bottleneck_string = ','.join(str(x) for x in bottleneck_values)
        with open(bottleneck_path, 'w') as bottleneck_file:
            bottleneck_file.write(bottleneck_string)
    else:
        # 直接从文件中获取图片相应的特征向量。
        with open(bottleneck_path, 'r') as bottleneck_file:
            bottleneck_string = bottleneck_file.read()
        bottleneck_values = [float(x) for x in bottleneck_string.split(',')]
    # 返回得到的特征向量
    return bottleneck_values


# 这个函数随机获取一个batch的图片作为训练数据。
def get_random_cached_bottlenecks(sess, n_classes, image_lists, how_many, category,
                                  jpeg_data_tensor, bottleneck_tensor):
    bottlenecks = []
    ground_truths = []
    for _ in range(how_many):
        # 随机一个类别和图片的编号加入当前的训练数据。
        label_index = random.randrange(n_classes)
        label_name = list(image_lists.keys())[label_index]
        image_index = random.randrange(65536)
        bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, image_index, category,
                                              jpeg_data_tensor, bottleneck_tensor)
        ground_truth = np.zeros(n_classes, dtype=np.float32)
        ground_truth[label_index] = 1.0
        bottlenecks.append(bottleneck)
        ground_truths.append(ground_truth)
    return bottlenecks, ground_truths


# 这个函数获取全部的测试数据。在最终测试的时候需要在所有的测试数据上计算正确率。
def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):
    bottlenecks = []
    ground_truths = []
    label_name_list = list(image_lists.keys())
    # 枚举所有的类别和每个类别中的测试图片。
    for label_index, label_name in enumerate(label_name_list):
        category = 'testing'
        for index, unused_base_name in enumerate(image_lists[label_name][category]):
            # 通过Inception-v3模型计算图片对应的特征向量，并将其加入最终数据的列表。
            bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, index, category,
                                                  jpeg_data_tensor, bottleneck_tensor)
            ground_truth = np.zeros(n_classes, dtype = np.float32)
            ground_truth[label_index] = 1.0
            bottlenecks.append(bottleneck)
            ground_truths.append(ground_truth)
    return bottlenecks, ground_truths


def main(_):
    # 读取所有图片。
    image_lists = create_image_lists(TEST_PERCENTAGE, VALIDATION_PERCENTAGE)
    n_classes = len(image_lists.keys())
    # 读取已经训练好的Inception-v3模型。
    # 谷歌训练好的模型保存在了GraphDef Protocol Buffer中，里面保存了每一个节点取值的计算方法以及变量的取值。
    # TensorFlow模型持久化的问题在第5章中有详细的介绍。
    with gfile.FastGFile(os.path.join(MODEL_DIR, MODEL_FILE), 'rb') as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())
    # 加载读取的Inception-v3模型，并返回数据输入所对应的张量以及计算瓶颈层结果所对应的张量。
    bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(graph_def, return_elements=[BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME])
    # 定义新的神经网络输入，这个输入就是新的图片经过Inception-v3模型前向传播到达瓶颈层时的结点取值。
    # 可以将这个过程类似的理解为一种特征提取。
    bottleneck_input = tf.placeholder(tf.float32, [None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder')
    # 定义新的标准答案输入
    ground_truth_input = tf.placeholder(tf.float32, [None, n_classes], name='GroundTruthInput')
    # 定义一层全连接层来解决新的图片分类问题。
    # 因为训练好的Inception-v3模型已经将原始的图片抽象为了更加容易分类的特征向量了，所以不需要再训练那么复杂的神经网络来完成这个新的分类任务。
    with tf.name_scope('final_training_ops'):
        weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, n_classes], stddev=0.001))
        biases = tf.Variable(tf.zeros([n_classes]))
        logits = tf.matmul(bottleneck_input, weights) + biases
        final_tensor = tf.nn.softmax(logits)
    # 定义交叉熵损失函数
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=ground_truth_input)
    cross_entropy_mean = tf.reduce_mean(cross_entropy)
    train_step = tf.train.GradientDescentOptimizer(LEARNING_RATE).minimize(cross_entropy_mean)
    # 计算正确率
    with tf.name_scope('evaluation'):
        correct_prediction = tf.equal(tf.argmax(final_tensor, 1), tf.argmax(ground_truth_input, 1))
        evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

    with tf.Session() as sess:
        tf.global_variables_initializer().run()
        # 训练过程
        for i in range(STEPS):
            # 每次获取一个batch的训练数据
            train_bottlenecks, train_ground_truth = get_random_cached_bottlenecks(
                sess, n_classes, image_lists, BATCH, 'training', jpeg_data_tensor, bottleneck_tensor)
            sess.run(train_step, feed_dict={bottleneck_input: train_bottlenecks, ground_truth_input: train_ground_truth})
            # 在验证集上测试正确率。
            if i%10 == 0 or i+1 == STEPS:
                validation_bottlenecks, validation_ground_truth = get_random_cached_bottlenecks(
                    sess, n_classes, image_lists, BATCH, 'validation', jpeg_data_tensor, bottleneck_tensor)
                validation_accuracy = sess.run(evaluation_step, feed_dict={
                    bottleneck_input:validation_bottlenecks, ground_truth_input: validation_ground_truth})
                print('Step %d: Validation accuracy on random sampled %d examples = %.1f%%'
                      % (i, BATCH, validation_accuracy*100))
        # 在最后的测试数据上测试正确率
        test_bottlenecks, test_ground_truth = get_test_bottlenecks(sess, image_lists, n_classes,
                                                                       jpeg_data_tensor, bottleneck_tensor)
        test_accuracy = sess.run(evaluation_step, feed_dict={bottleneck_input: test_bottlenecks,
                                                                 ground_truth_input: test_ground_truth})
        print('Final test accuracy = %.1f%%' % (test_accuracy * 100))


if __name__ == '__main__':
    tf.app.run()

效果：

2018-01-18 18:23:12.780704: I C: f_jenkinshomeworkspace el-winMwindowsPY36 ensorflowcoreplatformcpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2018-01-18 18:23:13.824129: W C: f_jenkinshomeworkspace el-winMwindowsPY36 ensorflowcoreframeworkop_def_util.cc:334] Op BatchNormWithGlobalNormalization is deprecated. It will cease to work in GraphDef version 9. Use tf.nn.batch_normalization().
Step 0: Validation accuracy on random sampled 100 examples = 48.0%
Step 10: Validation accuracy on random sampled 100 examples = 70.0%
Step 20: Validation accuracy on random sampled 100 examples = 57.0%
Step 30: Validation accuracy on random sampled 100 examples = 58.0%
Step 40: Validation accuracy on random sampled 100 examples = 70.0%
Step 50: Validation accuracy on random sampled 100 examples = 83.0%
Step 60: Validation accuracy on random sampled 100 examples = 72.0%
Step 70: Validation accuracy on random sampled 100 examples = 74.0%
Step 80: Validation accuracy on random sampled 100 examples = 70.0%
Step 90: Validation accuracy on random sampled 100 examples = 74.0%
Step 100: Validation accuracy on random sampled 100 examples = 79.0%
Step 110: Validation accuracy on random sampled 100 examples = 82.0%
Step 120: Validation accuracy on random sampled 100 examples = 80.0%
Step 130: Validation accuracy on random sampled 100 examples = 76.0%
Step 140: Validation accuracy on random sampled 100 examples = 81.0%
Step 150: Validation accuracy on random sampled 100 examples = 74.0%
Step 160: Validation accuracy on random sampled 100 examples = 82.0%
Step 170: Validation accuracy on random sampled 100 examples = 79.0%
Step 180: Validation accuracy on random sampled 100 examples = 75.0%
Step 190: Validation accuracy on random sampled 100 examples = 77.0%
Step 200: Validation accuracy on random sampled 100 examples = 74.0%
Step 210: Validation accuracy on random sampled 100 examples = 80.0%
Step 220: Validation accuracy on random sampled 100 examples = 85.0%
Step 230: Validation accuracy on random sampled 100 examples = 73.0%
Step 240: Validation accuracy on random sampled 100 examples = 79.0%
Step 250: Validation accuracy on random sampled 100 examples = 77.0%
Step 260: Validation accuracy on random sampled 100 examples = 78.0%
Step 270: Validation accuracy on random sampled 100 examples = 80.0%
Step 280: Validation accuracy on random sampled 100 examples = 76.0%
Step 290: Validation accuracy on random sampled 100 examples = 83.0%
Step 300: Validation accuracy on random sampled 100 examples = 83.0%
Step 310: Validation accuracy on random sampled 100 examples = 77.0%
Step 320: Validation accuracy on random sampled 100 examples = 73.0%
Step 330: Validation accuracy on random sampled 100 examples = 83.0%
Step 340: Validation accuracy on random sampled 100 examples = 77.0%
Step 350: Validation accuracy on random sampled 100 examples = 81.0%
Step 360: Validation accuracy on random sampled 100 examples = 79.0%
Step 370: Validation accuracy on random sampled 100 examples = 76.0%
Step 380: Validation accuracy on random sampled 100 examples = 83.0%
Step 390: Validation accuracy on random sampled 100 examples = 80.0%
Step 400: Validation accuracy on random sampled 100 examples = 80.0%
Step 410: Validation accuracy on random sampled 100 examples = 79.0%
Step 420: Validation accuracy on random sampled 100 examples = 83.0%
Step 430: Validation accuracy on random sampled 100 examples = 83.0%
Step 440: Validation accuracy on random sampled 100 examples = 71.0%
Step 450: Validation accuracy on random sampled 100 examples = 75.0%
Step 460: Validation accuracy on random sampled 100 examples = 86.0%
Step 470: Validation accuracy on random sampled 100 examples = 83.0%
Step 480: Validation accuracy on random sampled 100 examples = 82.0%
Step 490: Validation accuracy on random sampled 100 examples = 81.0%
Step 500: Validation accuracy on random sampled 100 examples = 82.0%
Step 510: Validation accuracy on random sampled 100 examples = 85.0%
Step 520: Validation accuracy on random sampled 100 examples = 77.0%
Step 530: Validation accuracy on random sampled 100 examples = 79.0%
Step 540: Validation accuracy on random sampled 100 examples = 87.0%
Step 550: Validation accuracy on random sampled 100 examples = 79.0%
Step 560: Validation accuracy on random sampled 100 examples = 75.0%
Step 570: Validation accuracy on random sampled 100 examples = 83.0%
Step 580: Validation accuracy on random sampled 100 examples = 81.0%
Step 590: Validation accuracy on random sampled 100 examples = 82.0%
Step 600: Validation accuracy on random sampled 100 examples = 82.0%
Step 610: Validation accuracy on random sampled 100 examples = 80.0%
Step 620: Validation accuracy on random sampled 100 examples = 87.0%
Step 630: Validation accuracy on random sampled 100 examples = 82.0%
Step 640: Validation accuracy on random sampled 100 examples = 81.0%
Step 650: Validation accuracy on random sampled 100 examples = 80.0%
Step 660: Validation accuracy on random sampled 100 examples = 75.0%
Step 670: Validation accuracy on random sampled 100 examples = 79.0%
Step 680: Validation accuracy on random sampled 100 examples = 82.0%
Step 690: Validation accuracy on random sampled 100 examples = 89.0%
Step 700: Validation accuracy on random sampled 100 examples = 76.0%
Step 710: Validation accuracy on random sampled 100 examples = 85.0%
Step 720: Validation accuracy on random sampled 100 examples = 77.0%
Step 730: Validation accuracy on random sampled 100 examples = 81.0%
Step 740: Validation accuracy on random sampled 100 examples = 80.0%
Step 750: Validation accuracy on random sampled 100 examples = 82.0%
Step 760: Validation accuracy on random sampled 100 examples = 79.0%
Step 770: Validation accuracy on random sampled 100 examples = 78.0%
Step 780: Validation accuracy on random sampled 100 examples = 88.0%
Step 790: Validation accuracy on random sampled 100 examples = 86.0%
Step 800: Validation accuracy on random sampled 100 examples = 76.0%
Step 810: Validation accuracy on random sampled 100 examples = 83.0%
Step 820: Validation accuracy on random sampled 100 examples = 85.0%
Step 830: Validation accuracy on random sampled 100 examples = 82.0%
Step 840: Validation accuracy on random sampled 100 examples = 83.0%
Step 850: Validation accuracy on random sampled 100 examples = 75.0%
Step 860: Validation accuracy on random sampled 100 examples = 79.0%
Step 870: Validation accuracy on random sampled 100 examples = 79.0%
Step 880: Validation accuracy on random sampled 100 examples = 83.0%
Step 890: Validation accuracy on random sampled 100 examples = 86.0%
Step 900: Validation accuracy on random sampled 100 examples = 82.0%
Step 910: Validation accuracy on random sampled 100 examples = 72.0%
Step 920: Validation accuracy on random sampled 100 examples = 79.0%
Step 930: Validation accuracy on random sampled 100 examples = 83.0%
Step 940: Validation accuracy on random sampled 100 examples = 81.0%
Step 950: Validation accuracy on random sampled 100 examples = 76.0%
Step 960: Validation accuracy on random sampled 100 examples = 77.0%
Step 970: Validation accuracy on random sampled 100 examples = 84.0%
Step 980: Validation accuracy on random sampled 100 examples = 78.0%
Step 990: Validation accuracy on random sampled 100 examples = 84.0%
Step 1000: Validation accuracy on random sampled 100 examples = 82.0%
Step 1010: Validation accuracy on random sampled 100 examples = 76.0%
Step 1020: Validation accuracy on random sampled 100 examples = 81.0%
Step 1030: Validation accuracy on random sampled 100 examples = 85.0%
Step 1040: Validation accuracy on random sampled 100 examples = 84.0%
Step 1050: Validation accuracy on random sampled 100 examples = 76.0%
Step 1060: Validation accuracy on random sampled 100 examples = 83.0%
Step 1070: Validation accuracy on random sampled 100 examples = 78.0%
Step 1080: Validation accuracy on random sampled 100 examples = 83.0%
Step 1090: Validation accuracy on random sampled 100 examples = 80.0%
Step 1100: Validation accuracy on random sampled 100 examples = 93.0%
Step 1110: Validation accuracy on random sampled 100 examples = 85.0%
Step 1120: Validation accuracy on random sampled 100 examples = 74.0%
Step 1130: Validation accuracy on random sampled 100 examples = 77.0%
Step 1140: Validation accuracy on random sampled 100 examples = 75.0%
Step 1150: Validation accuracy on random sampled 100 examples = 78.0%
Step 1160: Validation accuracy on random sampled 100 examples = 88.0%
Step 1170: Validation accuracy on random sampled 100 examples = 81.0%
Step 1180: Validation accuracy on random sampled 100 examples = 84.0%
Step 1190: Validation accuracy on random sampled 100 examples = 77.0%
Step 1200: Validation accuracy on random sampled 100 examples = 87.0%
Step 1210: Validation accuracy on random sampled 100 examples = 81.0%
Step 1220: Validation accuracy on random sampled 100 examples = 87.0%
Step 1230: Validation accuracy on random sampled 100 examples = 83.0%
Step 1240: Validation accuracy on random sampled 100 examples = 87.0%
Step 1250: Validation accuracy on random sampled 100 examples = 81.0%
Step 1260: Validation accuracy on random sampled 100 examples = 82.0%
Step 1270: Validation accuracy on random sampled 100 examples = 79.0%
Step 1280: Validation accuracy on random sampled 100 examples = 80.0%
Step 1290: Validation accuracy on random sampled 100 examples = 81.0%
Step 1300: Validation accuracy on random sampled 100 examples = 84.0%
Step 1310: Validation accuracy on random sampled 100 examples = 75.0%
Step 1320: Validation accuracy on random sampled 100 examples = 82.0%
Step 1330: Validation accuracy on random sampled 100 examples = 82.0%
Step 1340: Validation accuracy on random sampled 100 examples = 82.0%
Step 1350: Validation accuracy on random sampled 100 examples = 85.0%
Step 1360: Validation accuracy on random sampled 100 examples = 80.0%
Step 1370: Validation accuracy on random sampled 100 examples = 86.0%
Step 1380: Validation accuracy on random sampled 100 examples = 86.0%
Step 1390: Validation accuracy on random sampled 100 examples = 80.0%
Step 1400: Validation accuracy on random sampled 100 examples = 75.0%
Step 1410: Validation accuracy on random sampled 100 examples = 81.0%
Step 1420: Validation accuracy on random sampled 100 examples = 77.0%
Step 1430: Validation accuracy on random sampled 100 examples = 78.0%
Step 1440: Validation accuracy on random sampled 100 examples = 77.0%
Step 1450: Validation accuracy on random sampled 100 examples = 83.0%
Step 1460: Validation accuracy on random sampled 100 examples = 77.0%
Step 1470: Validation accuracy on random sampled 100 examples = 75.0%
Step 1480: Validation accuracy on random sampled 100 examples = 88.0%
Step 1490: Validation accuracy on random sampled 100 examples = 87.0%
Step 1500: Validation accuracy on random sampled 100 examples = 79.0%
Step 1510: Validation accuracy on random sampled 100 examples = 78.0%
Step 1520: Validation accuracy on random sampled 100 examples = 82.0%
Step 1530: Validation accuracy on random sampled 100 examples = 80.0%
Step 1540: Validation accuracy on random sampled 100 examples = 77.0%
Step 1550: Validation accuracy on random sampled 100 examples = 74.0%
Step 1560: Validation accuracy on random sampled 100 examples = 83.0%
Step 1570: Validation accuracy on random sampled 100 examples = 78.0%
Step 1580: Validation accuracy on random sampled 100 examples = 78.0%
Step 1590: Validation accuracy on random sampled 100 examples = 79.0%
Step 1600: Validation accuracy on random sampled 100 examples = 83.0%
Step 1610: Validation accuracy on random sampled 100 examples = 82.0%
Step 1620: Validation accuracy on random sampled 100 examples = 80.0%
Step 1630: Validation accuracy on random sampled 100 examples = 84.0%
Step 1640: Validation accuracy on random sampled 100 examples = 83.0%
Step 1650: Validation accuracy on random sampled 100 examples = 76.0%
Step 1660: Validation accuracy on random sampled 100 examples = 83.0%
Step 1670: Validation accuracy on random sampled 100 examples = 86.0%
Step 1680: Validation accuracy on random sampled 100 examples = 82.0%
Step 1690: Validation accuracy on random sampled 100 examples = 85.0%
Step 1700: Validation accuracy on random sampled 100 examples = 87.0%
Step 1710: Validation accuracy on random sampled 100 examples = 80.0%
Step 1720: Validation accuracy on random sampled 100 examples = 82.0%
Step 1730: Validation accuracy on random sampled 100 examples = 80.0%
Step 1740: Validation accuracy on random sampled 100 examples = 79.0%
Step 1750: Validation accuracy on random sampled 100 examples = 85.0%
Step 1760: Validation accuracy on random sampled 100 examples = 80.0%
Step 1770: Validation accuracy on random sampled 100 examples = 76.0%
Step 1780: Validation accuracy on random sampled 100 examples = 84.0%
Step 1790: Validation accuracy on random sampled 100 examples = 78.0%
Step 1800: Validation accuracy on random sampled 100 examples = 84.0%
Step 1810: Validation accuracy on random sampled 100 examples = 91.0%
Step 1820: Validation accuracy on random sampled 100 examples = 78.0%
Step 1830: Validation accuracy on random sampled 100 examples = 72.0%
Step 1840: Validation accuracy on random sampled 100 examples = 85.0%
Step 1850: Validation accuracy on random sampled 100 examples = 80.0%
Step 1860: Validation accuracy on random sampled 100 examples = 93.0%
Step 1870: Validation accuracy on random sampled 100 examples = 83.0%
Step 1880: Validation accuracy on random sampled 100 examples = 88.0%
Step 1890: Validation accuracy on random sampled 100 examples = 83.0%
Step 1900: Validation accuracy on random sampled 100 examples = 81.0%
Step 1910: Validation accuracy on random sampled 100 examples = 76.0%
Step 1920: Validation accuracy on random sampled 100 examples = 80.0%
Step 1930: Validation accuracy on random sampled 100 examples = 81.0%
Step 1940: Validation accuracy on random sampled 100 examples = 79.0%
Step 1950: Validation accuracy on random sampled 100 examples = 86.0%
Step 1960: Validation accuracy on random sampled 100 examples = 84.0%
Step 1970: Validation accuracy on random sampled 100 examples = 79.0%
Step 1980: Validation accuracy on random sampled 100 examples = 80.0%
Step 1990: Validation accuracy on random sampled 100 examples = 83.0%
Step 2000: Validation accuracy on random sampled 100 examples = 78.0%
Step 2010: Validation accuracy on random sampled 100 examples = 81.0%
Step 2020: Validation accuracy on random sampled 100 examples = 83.0%
Step 2030: Validation accuracy on random sampled 100 examples = 82.0%
Step 2040: Validation accuracy on random sampled 100 examples = 82.0%
Step 2050: Validation accuracy on random sampled 100 examples = 86.0%
Step 2060: Validation accuracy on random sampled 100 examples = 87.0%
Step 2070: Validation accuracy on random sampled 100 examples = 87.0%
Step 2080: Validation accuracy on random sampled 100 examples = 87.0%
Step 2090: Validation accuracy on random sampled 100 examples = 73.0%
Step 2100: Validation accuracy on random sampled 100 examples = 82.0%
Step 2110: Validation accuracy on random sampled 100 examples = 79.0%
Step 2120: Validation accuracy on random sampled 100 examples = 85.0%
Step 2130: Validation accuracy on random sampled 100 examples = 85.0%
Step 2140: Validation accuracy on random sampled 100 examples = 87.0%
Step 2150: Validation accuracy on random sampled 100 examples = 83.0%
Step 2160: Validation accuracy on random sampled 100 examples = 84.0%
Step 2170: Validation accuracy on random sampled 100 examples = 86.0%
Step 2180: Validation accuracy on random sampled 100 examples = 76.0%
Step 2190: Validation accuracy on random sampled 100 examples = 83.0%
Step 2200: Validation accuracy on random sampled 100 examples = 78.0%
Step 2210: Validation accuracy on random sampled 100 examples = 80.0%
Step 2220: Validation accuracy on random sampled 100 examples = 78.0%
Step 2230: Validation accuracy on random sampled 100 examples = 76.0%
Step 2240: Validation accuracy on random sampled 100 examples = 83.0%
Step 2250: Validation accuracy on random sampled 100 examples = 80.0%
Step 2260: Validation accuracy on random sampled 100 examples = 79.0%
Step 2270: Validation accuracy on random sampled 100 examples = 80.0%
Step 2280: Validation accuracy on random sampled 100 examples = 77.0%
Step 2290: Validation accuracy on random sampled 100 examples = 84.0%
Step 2300: Validation accuracy on random sampled 100 examples = 84.0%
Step 2310: Validation accuracy on random sampled 100 examples = 78.0%
Step 2320: Validation accuracy on random sampled 100 examples = 78.0%
Step 2330: Validation accuracy on random sampled 100 examples = 78.0%
Step 2340: Validation accuracy on random sampled 100 examples = 85.0%
Step 2350: Validation accuracy on random sampled 100 examples = 87.0%
Step 2360: Validation accuracy on random sampled 100 examples = 89.0%
Step 2370: Validation accuracy on random sampled 100 examples = 82.0%
Step 2380: Validation accuracy on random sampled 100 examples = 88.0%
Step 2390: Validation accuracy on random sampled 100 examples = 87.0%
Step 2400: Validation accuracy on random sampled 100 examples = 84.0%
Step 2410: Validation accuracy on random sampled 100 examples = 84.0%
Step 2420: Validation accuracy on random sampled 100 examples = 82.0%
Step 2430: Validation accuracy on random sampled 100 examples = 77.0%
Step 2440: Validation accuracy on random sampled 100 examples = 85.0%
Step 2450: Validation accuracy on random sampled 100 examples = 88.0%
Step 2460: Validation accuracy on random sampled 100 examples = 86.0%
Step 2470: Validation accuracy on random sampled 100 examples = 79.0%
Step 2480: Validation accuracy on random sampled 100 examples = 78.0%
Step 2490: Validation accuracy on random sampled 100 examples = 83.0%
Step 2500: Validation accuracy on random sampled 100 examples = 84.0%
Step 2510: Validation accuracy on random sampled 100 examples = 87.0%
Step 2520: Validation accuracy on random sampled 100 examples = 83.0%
Step 2530: Validation accuracy on random sampled 100 examples = 82.0%
Step 2540: Validation accuracy on random sampled 100 examples = 79.0%
Step 2550: Validation accuracy on random sampled 100 examples = 79.0%
Step 2560: Validation accuracy on random sampled 100 examples = 75.0%
Step 2570: Validation accuracy on random sampled 100 examples = 79.0%
Step 2580: Validation accuracy on random sampled 100 examples = 83.0%
Step 2590: Validation accuracy on random sampled 100 examples = 82.0%
Step 2600: Validation accuracy on random sampled 100 examples = 77.0%
Step 2610: Validation accuracy on random sampled 100 examples = 80.0%
Step 2620: Validation accuracy on random sampled 100 examples = 84.0%
Step 2630: Validation accuracy on random sampled 100 examples = 82.0%
Step 2640: Validation accuracy on random sampled 100 examples = 82.0%
Step 2650: Validation accuracy on random sampled 100 examples = 78.0%
Step 2660: Validation accuracy on random sampled 100 examples = 84.0%
Step 2670: Validation accuracy on random sampled 100 examples = 90.0%
Step 2680: Validation accuracy on random sampled 100 examples = 73.0%
Step 2690: Validation accuracy on random sampled 100 examples = 78.0%
Step 2700: Validation accuracy on random sampled 100 examples = 85.0%
Step 2710: Validation accuracy on random sampled 100 examples = 81.0%
Step 2720: Validation accuracy on random sampled 100 examples = 84.0%
Step 2730: Validation accuracy on random sampled 100 examples = 78.0%
Step 2740: Validation accuracy on random sampled 100 examples = 81.0%
Step 2750: Validation accuracy on random sampled 100 examples = 82.0%
Step 2760: Validation accuracy on random sampled 100 examples = 76.0%
Step 2770: Validation accuracy on random sampled 100 examples = 86.0%
Step 2780: Validation accuracy on random sampled 100 examples = 82.0%
Step 2790: Validation accuracy on random sampled 100 examples = 75.0%
Step 2800: Validation accuracy on random sampled 100 examples = 85.0%
Step 2810: Validation accuracy on random sampled 100 examples = 84.0%
Step 2820: Validation accuracy on random sampled 100 examples = 80.0%
Step 2830: Validation accuracy on random sampled 100 examples = 80.0%
Step 2840: Validation accuracy on random sampled 100 examples = 82.0%
Step 2850: Validation accuracy on random sampled 100 examples = 83.0%
Step 2860: Validation accuracy on random sampled 100 examples = 83.0%
Step 2870: Validation accuracy on random sampled 100 examples = 84.0%
Step 2880: Validation accuracy on random sampled 100 examples = 80.0%
Step 2890: Validation accuracy on random sampled 100 examples = 91.0%
Step 2900: Validation accuracy on random sampled 100 examples = 83.0%
Step 2910: Validation accuracy on random sampled 100 examples = 85.0%
Step 2920: Validation accuracy on random sampled 100 examples = 80.0%
Step 2930: Validation accuracy on random sampled 100 examples = 79.0%
Step 2940: Validation accuracy on random sampled 100 examples = 86.0%
Step 2950: Validation accuracy on random sampled 100 examples = 79.0%
Step 2960: Validation accuracy on random sampled 100 examples = 86.0%
Step 2970: Validation accuracy on random sampled 100 examples = 78.0%
Step 2980: Validation accuracy on random sampled 100 examples = 79.0%
Step 2990: Validation accuracy on random sampled 100 examples = 79.0%
Step 3000: Validation accuracy on random sampled 100 examples = 89.0%
Step 3010: Validation accuracy on random sampled 100 examples = 79.0%
Step 3020: Validation accuracy on random sampled 100 examples = 82.0%
Step 3030: Validation accuracy on random sampled 100 examples = 81.0%
Step 3040: Validation accuracy on random sampled 100 examples = 79.0%
Step 3050: Validation accuracy on random sampled 100 examples = 87.0%
Step 3060: Validation accuracy on random sampled 100 examples = 70.0%
Step 3070: Validation accuracy on random sampled 100 examples = 78.0%
Step 3080: Validation accuracy on random sampled 100 examples = 83.0%
Step 3090: Validation accuracy on random sampled 100 examples = 78.0%
Step 3100: Validation accuracy on random sampled 100 examples = 83.0%
Step 3110: Validation accuracy on random sampled 100 examples = 85.0%
Step 3120: Validation accuracy on random sampled 100 examples = 79.0%
Step 3130: Validation accuracy on random sampled 100 examples = 80.0%
Step 3140: Validation accuracy on random sampled 100 examples = 81.0%
Step 3150: Validation accuracy on random sampled 100 examples = 80.0%
Step 3160: Validation accuracy on random sampled 100 examples = 80.0%
Step 3170: Validation accuracy on random sampled 100 examples = 74.0%
Step 3180: Validation accuracy on random sampled 100 examples = 77.0%
Step 3190: Validation accuracy on random sampled 100 examples = 82.0%
Step 3200: Validation accuracy on random sampled 100 examples = 83.0%
Step 3210: Validation accuracy on random sampled 100 examples = 82.0%
Step 3220: Validation accuracy on random sampled 100 examples = 81.0%
Step 3230: Validation accuracy on random sampled 100 examples = 80.0%
Step 3240: Validation accuracy on random sampled 100 examples = 86.0%
Step 3250: Validation accuracy on random sampled 100 examples = 87.0%
Step 3260: Validation accuracy on random sampled 100 examples = 79.0%
Step 3270: Validation accuracy on random sampled 100 examples = 80.0%
Step 3280: Validation accuracy on random sampled 100 examples = 78.0%
Step 3290: Validation accuracy on random sampled 100 examples = 86.0%
Step 3300: Validation accuracy on random sampled 100 examples = 84.0%
Step 3310: Validation accuracy on random sampled 100 examples = 85.0%
Step 3320: Validation accuracy on random sampled 100 examples = 80.0%
Step 3330: Validation accuracy on random sampled 100 examples = 82.0%
Step 3340: Validation accuracy on random sampled 100 examples = 85.0%
Step 3350: Validation accuracy on random sampled 100 examples = 83.0%
Step 3360: Validation accuracy on random sampled 100 examples = 82.0%
Step 3370: Validation accuracy on random sampled 100 examples = 84.0%
Step 3380: Validation accuracy on random sampled 100 examples = 84.0%
Step 3390: Validation accuracy on random sampled 100 examples = 77.0%
Step 3400: Validation accuracy on random sampled 100 examples = 81.0%
Step 3410: Validation accuracy on random sampled 100 examples = 81.0%
Step 3420: Validation accuracy on random sampled 100 examples = 78.0%
Step 3430: Validation accuracy on random sampled 100 examples = 84.0%
Step 3440: Validation accuracy on random sampled 100 examples = 84.0%
Step 3450: Validation accuracy on random sampled 100 examples = 82.0%
Step 3460: Validation accuracy on random sampled 100 examples = 81.0%
Step 3470: Validation accuracy on random sampled 100 examples = 85.0%
Step 3480: Validation accuracy on random sampled 100 examples = 82.0%
Step 3490: Validation accuracy on random sampled 100 examples = 83.0%
Step 3500: Validation accuracy on random sampled 100 examples = 90.0%
Step 3510: Validation accuracy on random sampled 100 examples = 74.0%
Step 3520: Validation accuracy on random sampled 100 examples = 75.0%
Step 3530: Validation accuracy on random sampled 100 examples = 84.0%
Step 3540: Validation accuracy on random sampled 100 examples = 80.0%
Step 3550: Validation accuracy on random sampled 100 examples = 87.0%
Step 3560: Validation accuracy on random sampled 100 examples = 79.0%
Step 3570: Validation accuracy on random sampled 100 examples = 83.0%
Step 3580: Validation accuracy on random sampled 100 examples = 90.0%
Step 3590: Validation accuracy on random sampled 100 examples = 78.0%
Step 3600: Validation accuracy on random sampled 100 examples = 81.0%
Step 3610: Validation accuracy on random sampled 100 examples = 88.0%
Step 3620: Validation accuracy on random sampled 100 examples = 78.0%
Step 3630: Validation accuracy on random sampled 100 examples = 87.0%
Step 3640: Validation accuracy on random sampled 100 examples = 90.0%
Step 3650: Validation accuracy on random sampled 100 examples = 83.0%
Step 3660: Validation accuracy on random sampled 100 examples = 83.0%
Step 3670: Validation accuracy on random sampled 100 examples = 86.0%
Step 3680: Validation accuracy on random sampled 100 examples = 87.0%
Step 3690: Validation accuracy on random sampled 100 examples = 86.0%
Step 3700: Validation accuracy on random sampled 100 examples = 76.0%
Step 3710: Validation accuracy on random sampled 100 examples = 78.0%
Step 3720: Validation accuracy on random sampled 100 examples = 85.0%
Step 3730: Validation accuracy on random sampled 100 examples = 76.0%
Step 3740: Validation accuracy on random sampled 100 examples = 79.0%
Step 3750: Validation accuracy on random sampled 100 examples = 86.0%
Step 3760: Validation accuracy on random sampled 100 examples = 82.0%
Step 3770: Validation accuracy on random sampled 100 examples = 78.0%
Step 3780: Validation accuracy on random sampled 100 examples = 82.0%
Step 3790: Validation accuracy on random sampled 100 examples = 83.0%
Step 3800: Validation accuracy on random sampled 100 examples = 78.0%
Step 3810: Validation accuracy on random sampled 100 examples = 75.0%
Step 3820: Validation accuracy on random sampled 100 examples = 84.0%
Step 3830: Validation accuracy on random sampled 100 examples = 83.0%
Step 3840: Validation accuracy on random sampled 100 examples = 80.0%
Step 3850: Validation accuracy on random sampled 100 examples = 83.0%
Step 3860: Validation accuracy on random sampled 100 examples = 77.0%
Step 3870: Validation accuracy on random sampled 100 examples = 88.0%
Step 3880: Validation accuracy on random sampled 100 examples = 83.0%
Step 3890: Validation accuracy on random sampled 100 examples = 82.0%
Step 3900: Validation accuracy on random sampled 100 examples = 86.0%
Step 3910: Validation accuracy on random sampled 100 examples = 79.0%
Step 3920: Validation accuracy on random sampled 100 examples = 90.0%
Step 3930: Validation accuracy on random sampled 100 examples = 81.0%
Step 3940: Validation accuracy on random sampled 100 examples = 83.0%
Step 3950: Validation accuracy on random sampled 100 examples = 84.0%
Step 3960: Validation accuracy on random sampled 100 examples = 79.0%
Step 3970: Validation accuracy on random sampled 100 examples = 84.0%
Step 3980: Validation accuracy on random sampled 100 examples = 84.0%
Step 3990: Validation accuracy on random sampled 100 examples = 76.0%
Step 3999: Validation accuracy on random sampled 100 examples = 83.0%
Final test accuracy = 82.2%