alex✳space
机器学习 / A LEAF IN THE GARDEN

CIFAR-10图像分类

机器学习 · 动手深度学习。来自 Daily Learning 的学习笔记。

15 分钟阅读更新于 2026/08/29
#机器学习#动手深度学习#Notebook
代码 2 · In [1]
import torch
import torchvision
from torch import nn
import util
import shutil
import pandas as pd
import math
import os
import kagglehub
from pathlib import Path
import collections
已保存输出
/root/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/requests/__init__.py:92: RequestsDependencyWarning: Unable to find acceptable character detection dependency (chardet or charset_normalizer).
  warnings.warn(

加载数据

代码 4 · In [2]
data_dir = Path('./data/cifar-10')
label_file = data_dir/'trainLabels.csv'
def read_csv_labels(fname: str):
    with open(fname, 'r') as f:
        lines = f.readlines()[1:]  # 去掉表头
    tokens = [l.rstrip().split(',') for l in lines]
    return dict((name, label) for name, label in tokens)
read_csv_labels(label_file)
已保存输出
{'1': 'frog',
 '2': 'truck',
 '3': 'truck',
 '4': 'deer',
 '5': 'automobile',
 '6': 'automobile',
 '7': 'bird',
 '8': 'horse',
 '9': 'ship',
 '10': 'cat',
 '11': 'deer',
 '12': 'horse',
 '13': 'horse',
 '14': 'bird',
 '15': 'truck',
 '16': 'truck',
 '17': 'truck',
 '18': 'cat',
 '19': 'bird',
 '20': 'frog',
 '21': 'deer',
 '22': 'cat',
 '23': 'frog',
 '24': 'frog',
 '25': 'bird',
 '26': 'frog',
 '27': 'cat',
 '28': 'dog',
 '29': 'deer',
 '30': 'airplane',
 '31': 'airplane',
 '32': 'truck',
 '33': 'automobile',
 '34': 'cat',
 '35': 'deer',
 '36': 'airplane',
 '37': 'cat',
 '38': 'horse',
 '39': 'cat',
 '40': 'cat',
 '41': 'dog',
 '42': 'bird',
 '43': 'bird',
 '44': 'horse',
 '45': 'automobile',
 '46': 'automobile',
 '47': 'automobile',
 '48': 'bird',
 '49': 'bird',
 '50': 'airplane',
 '51': 'truck',
 '52': 'dog',
 '53': 'horse',
 '54': 'truck',
 '55': 'bird',
 '56': 'bird',
 '57': 'dog',
 '58': 'bird',
 '59': 'deer',
 '60': 'cat',
 '61': 'automobile',
 '62': 'automobile',
 '63': 'ship',
 '64': 'bird',
 '65': 'automobile',
 '66': 'automobile',
 '67': 'deer',
 '68': 'truck',
 '69': 'horse',
 '70': 'ship',
 '71': 'dog',
 '72': 'truck',
 '73': 'frog',
 '74': 'horse',
 '75': 'cat',
 '76': 'automobile',
 '77': 'truck',
 '78': 'airplane',
 '79': 'cat',
 '80': 'automobile',
 '81': 'cat',
 '82': 'dog',
 '83': 'deer',
 '84': 'dog',
 '85': 'horse',
 '86': 'horse',
 '87': 'deer',
 '88': 'horse',
 '89': 'truck',
 '90': 'deer',
 '91': 'bird',
 '92': 'cat',
 '93': 'ship',
 '94': 'airplane',
 '95': 'automobile',
 '96': 'frog',
 '97': 'automobile',
 '98': 'automobile',
 '99': 'deer',
 '100': 'automobile',
 '101': 'ship',
 '102': 'cat',
 '103': 'truck',
 '104': 'frog',
 '105': 'frog',
 '106': 'automobile',
 '107': 'ship',
 '108': 'dog',
 '109': 'bird',
 '110': 'truck',
 '111': 'truck',
 '112': 'ship',
 '113': 'automobile',
 '114': 'horse',
 '115': 'horse',
 '116': 'airplane',
 '117': 'airplane',
 '118': 'frog',
 '119': 'truck',
 '120': 'automobile',
 '121': 'bird',
 '122': 'bird',
 '123': 'truck',
 '124': 'bird',
 '125': 'frog',
 '126': 'frog',
 '127': 'automobile',
 '128': 'truck',
 '129': 'dog',
 '130': 'airplane',
 '131': 'deer',
 '132': 'horse',
 '133': 'frog',
 '134': 'horse',
 '135': 'automobile',
 '136': 'ship',
 '137': 'automobile',
 '138': 'automobile',
 '139': 'bird',
 '140': 'ship',
 '141': 'automobile',
 '142': 'cat',
 '143': 'cat',
 '144': 'frog',
 '145': 'bird',
 '146': 'deer',
 '147': 'truck',
 '148': 'truck',
 '149': 'dog',
 '150': 'deer',
 '151': 'cat',
 '152': 'frog',
 '153': 'horse',
 '154': 'deer',
 '155': 'frog',
 '156': 'ship',
 '157': 'dog',
 '158': 'dog',
 '159': 'deer',
 '160': 'cat',
 '161': 'automobile',
 '162': 'ship',
 '163': 'deer',
 '164': 'horse',
 '165': 'frog',
 '166': 'airplane',
 '167': 'truck',
 '168': 'dog',
 '169': 'automobile',
 '170': 'cat',
 '171': 'ship',
 '172': 'bird',
 '173': 'horse',
 '174': 'dog',
 '175': 'cat',
 '176': 'deer',
 '177': 'automobile',
 '178': 'dog',
 '179': 'horse',
 '180': 'airplane',
 '181': 'deer',
 '182': 'horse',
 '183': 'dog',
 '184': 'dog',
 '185': 'automobile',
 '186': 'airplane',
 '187': 'truck',
 '188': 'frog',
 '189': 'truck',
 '190': 'airplane',
 '191': 'ship',
 '192': 'horse',
 '193': 'ship',
 '194': 'ship',
 '195': 'bird',
 '196': 'dog',
 '197': 'bird',
 '198': 'cat',
 '199': 'dog',
 '200': 'airplane',
 '201': 'frog',
 '202': 'automobile',
 '203': 'truck',
 '204': 'cat',
 '205': 'frog',
 '206': 'truck',
 '207': 'automobile',
 '208': 'cat',
 '209': 'truck',
 '210': 'frog',
 '211': 'frog',
 '212': 'horse',
 '213': 'automobile',
 '214': 'airplane',
 '215': 'truck',
 '216': 'dog',
 '217': 'ship',
 '218': 'dog',
 '219': 'bird',
 '220': 'truck',
 '221': 'airplane',
 '222': 'ship',
 '223': 'ship',
 '224': 'airplane',
 '225': 'frog',
 '226': 'truck',
 '227': 'automobile',
 '228': 'automobile',
 '229': 'frog',
 '230': 'cat',
 '231': 'horse',
 '232': 'frog',
 '233': 'frog',
 '234': 'airplane',
 '235': 'frog',
 '236': 'frog',
 '237': 'automobile',
 '238': 'horse',
 '239': 'automobile',
 '240': 'dog',
 '241': 'ship',
 '242': 'cat',
 '243': 'frog',
 '244': 'frog',
 '245': 'ship',
 '246': 'frog',
 '247': 'ship',
 '248': 'deer',
 '249': 'frog',
 '250': 'frog',
 '251': 'automobile',
 '252': 'cat',
 '253': 'ship',
 '254': 'cat',
 '255': 'deer',
 '256': 'automobile',
 '257': 'horse',
 '258': 'automobile',
 '259': 'cat',
 '260': 'ship',
 '261': 'dog',
 '262': 'automobile',
 '263': 'automobile',
 '264': 'deer',
 '265': 'airplane',
 '266': 'truck',
 '267': 'cat',
 '268': 'horse',
 '269': 'deer',
 '270': 'truck',
 '271': 'truck',
 '272': 'bird',
 '273': 'deer',
 '274': 'truck',
 '275': 'truck',
 '276': 'automobile',
 '277': 'airplane',
 '278': 'dog',
 '279': 'truck',
 '280': 'airplane',
 '281': 'ship',
 '282': 'bird',
 '283': 'automobile',
 '284': 'bird',
 '285': 'airplane',
 '286': 'dog',
 '287': 'frog',
 '288': 'cat',
 '289': 'bird',
 '290': 'horse',
 '291': 'ship',
 '292': 'ship',
 '293': 'frog',
 '294': 'airplane',
 '295': 'horse',
 '296': 'truck',
 '297': 'deer',
 '298': 'dog',
 '299': 'frog',
 '300': 'deer',
 '301': 'bird',
 '302': 'automobile',
 '303': 'automobile',
 '304': 'bird',
 '305': 'automobile',
 '306': 'dog',
 '307': 'truck',
 '308': 'truck',
 '309': 'airplane',
 '310': 'ship',
 '311': 'deer',
 '312': 'automobile',
 '313': 'automobile',
 '314': 'frog',
 '315': 'cat',
 '316': 'cat',
 '317': 'truck',
 '318': 'airplane',
 '319': 'horse',
 '320': 'truck',
 '321': 'horse',
 '322': 'horse',
 '323': 'truck',
 '324': 'automobile',
 '325': 'dog',
 '326': 'automobile',
 '327': 'frog',
 '328': 'frog',
 '329': 'ship',
 '330': 'horse',
 '331': 'automobile',
 '332': 'cat',
 '333': 'airplane',
 '334': 'cat',
 '335': 'cat',
 '336': 'bird',
 '337': 'deer',
 '338': 'dog',
 '339': 'horse',
 '340': 'dog',
 '341': 'truck',
 '342': 'airplane',
 '343': 'cat',
 '344': 'deer',
 '345': 'airplane',
 '346': 'deer',
 '347': 'deer',
 '348': 'frog',
 '349': 'airplane',
 '350': 'airplane',
 '351': 'frog',
 '352': 'frog',
 '353': 'airplane',
 '354': 'ship',
 '355': 'automobile',
 '356': 'frog',
 '357': 'bird',
 '358': 'truck',
 '359': 'bird',
 '360': 'dog',
 '361': 'truck',
 '362': 'frog',
 '363': 'horse',
 '364': 'deer',
 '365': 'automobile',
 '366': 'ship',
 '367': 'horse',
 '368': 'cat',
 '369': 'frog',
 '370': 'truck',
 '371': 'cat',
 '372': 'airplane',
 '373': 'deer',
 '374': 'airplane',
 '375': 'dog',
 '376': 'automobile',
 '377': 'airplane',
 '378': 'cat',
 '379': 'deer',
 '380': 'ship',
 '381': 'dog',
 '382': 'deer',
 '383': 'horse',
 '384': 'bird',
 '385': 'cat',
 '386': 'truck',
 '387': 'horse',
 '388': 'frog',
 '389': 'horse',
 '390': 'automobile',
 '391': 'deer',
 '392': 'horse',
 '393': 'airplane',
 '394': 'automobile',
 '395': 'horse',
 '396': 'cat',
 '397': 'automobile',
 '398': 'ship',
 '399': 'deer',
 '400': 'deer',
 '401': 'bird',
 '402': 'airplane',
 '403': 'bird',
 '404': 'bird',
 '405': 'airplane',
 '406': 'airplane',
 '407': 'truck',
 '408': 'airplane',
 '409': 'truck',
 '410': 'frog',
 '411': 'ship',
 '412': 'bird',
 '413': 'horse',
 '414': 'horse',
 '415': 'deer',
 '416': 'airplane',
 '417': 'cat',
 '418': 'airplane',
 '419': 'ship',
 '420': 'truck',
 '421': 'deer',
 '422': 'bird',
 '423': 'horse',
 '424': 'bird',
 '425': 'dog',
 '426': 'bird',
 '427': 'dog',
 '428': 'automobile',
 '429': 'truck',
 '430': 'deer',
 '431': 'ship',
 '432': 'dog',
 '433': 'automobile',
 '434': 'horse',
 '435': 'deer',
 '436': 'deer',
 '437': 'airplane',
 '438': 'frog',
 '439': 'truck',
 '440': 'airplane',
 '441': 'horse',
 '442': 'ship',
 '443': 'ship',
 '444': 'truck',
 '445': 'truck',
 '446': 'cat',
 '447': 'cat',
 '448': 'deer',
 '449': 'airplane',
 '450': 'deer',
 '451': 'dog',
 '452': 'frog',
 '453': 'frog',
 '454': 'airplane',
 '455': 'automobile',
 '456': 'airplane',
 '457': 'ship',
 '458': 'airplane',
 '459': 'deer',
 '460': 'ship',
 '461': 'ship',
 '462': 'automobile',
 '463': 'dog',
 '464': 'bird',
 '465': 'frog',
 '466': 'ship',
 '467': 'automobile',
 '468': 'airplane',
 '469': 'airplane',
 '470': 'horse',
 '471': 'horse',
 '472': 'dog',
 '473': 'truck',
 '474': 'frog',
 '475': 'bird',
 '476': 'ship',
 '477': 'cat',
 '478': 'deer',
 '479': 'horse',
 '480': 'cat',
 '481': 'truck',
 '482': 'airplane',
 '483': 'automobile',
 '484': 'bird',
 '485': 'deer',
 '486': 'ship',
 '487': 'automobile',
 '488': 'ship',
 '489': 'frog',
 '490': 'deer',
 '491': 'deer',
 '492': 'dog',
 '493': 'horse',
 '494': 'automobile',
 '495': 'cat',
 '496': 'truck',
 '497': 'ship',
 '498': 'airplane',
 '499': 'automobile',
 '500': 'horse',
 '501': 'dog',
 '502': 'ship',
 '503': 'bird',
 '504': 'ship',
 '505': 'airplane',
 '506': 'deer',
 '507': 'automobile',
 '508': 'ship',
 '509': 'truck',
 '510': 'ship',
 '511': 'bird',
 '512': 'truck',
 '513': 'truck',
 '514': 'bird',
 '515': 'horse',
 '516': 'dog',
 '517': 'horse',
 '518': 'cat',
 '519': 'ship',
 '520': 'ship',
 '521': 'deer',
 '522': 'deer',
 '523': 'bird',
 '524': 'horse',
 '525': 'automobile',
 '526': 'frog',
 '527': 'deer',
 '528': 'airplane',
 '529': 'deer',
 '530': 'frog',
 '531': 'truck',
 '532': 'horse',
 '533': 'frog',
 '534': 'bird',
 '535': 'dog',
 '536': 'dog',
 '537': 'automobile',
 '538': 'horse',
 '539': 'bird',
 '540': 'bird',
 '541': 'bird',
 '542': 'truck',
 '543': 'dog',
 '544': 'deer',
 '545': 'bird',
 '546': 'horse',
 '547': 'ship',
 '548': 'automobile',
 '549': 'cat',
 '550': 'deer',
 '551': 'cat',
 '552': 'horse',
 '553': 'frog',
 '554': 'truck',
 '555': 'ship',
 '556': 'airplane',
 '557': 'frog',
 '558': 'airplane',
 '559': 'bird',
 '560': 'bird',
 '561': 'bird',
 '562': 'automobile',
 '563': 'ship',
 '564': 'deer',
 '565': 'airplane',
 '566': 'automobile',
 '567': 'ship',
 '568': 'ship',
 '569': 'automobile',
 '570': 'dog',
 '571': 'horse',
 '572': 'frog',
 '573': 'deer',
 '574': 'dog',
 '575': 'ship',
 '576': 'horse',
 '577': 'automobile',
 '578': 'truck',
 '579': 'automobile',
 '580': 'truck',
 '581': 'ship',
 '582': 'deer',
 '583': 'horse',
 '584': 'cat',
 '585': 'ship',
 '586': 'ship',
 '587': 'bird',
 '588': 'frog',
 '589': 'frog',
 '590': 'horse',
 '591': 'automobile',
 '592': 'frog',
 '593': 'ship',
 '594': 'automobile',
 '595': 'truck',
 '596': 'horse',
 '597': 'ship',
 '598': 'cat',
 '599': 'airplane',
 '600': 'automobile',
 '601': 'airplane',
 '602': 'ship',
 '603': 'ship',
 '604': 'cat',
 '605': 'airplane',
 '606': 'airplane',
 '607': 'automobile',
 '608': 'dog',
 '609': 'airplane',
 '610': 'ship',
 '611': 'ship',
 '612': 'horse',
 '613': 'truck',
 '614': 'truck',
 '615': 'airplane',
 '616': 'truck',
 '617': 'deer',
 '618': 'automobile',
 '619': 'cat',
 '620': 'frog',
 '621': 'frog',
 '622': 'deer',
 '623': 'deer',
 '624': 'horse',
 '625': 'dog',
 '626': 'frog',
 '627': 'airplane',
 '628': 'ship',
 '629': 'airplane',
 '630': 'cat',
 '631': 'bird',
 '632': 'ship',
 '633': 'deer',
 '634': 'frog',
 '635': 'truck',
 '636': 'truck',
 '637': 'horse',
 '638': 'airplane',
 '639': 'cat',
 '640': 'cat',
 '641': 'frog',
 '642': 'horse',
 '643': 'deer',
 '644': 'truck',
 '645': 'automobile',
 '646': 'frog',
 '647': 'bird',
 '648': 'horse',
 '649': 'bird',
 '650': 'bird',
 '651': 'airplane',
 '652': 'frog',
 '653': 'horse',
 '654': 'dog',
 '655': 'horse',
 '656': 'frog',
 '657': 'ship',
 '658': 'truck',
 '659': 'airplane',
 '660': 'truck',
 '661': 'deer',
 '662': 'deer',
 '663': 'horse',
 '664': 'airplane',
 '665': 'truck',
 '666': 'deer',
 '667': 'truck',
 '668': 'frog',
 '669': 'truck',
 '670': 'deer',
 '671': 'dog',
 '672': 'horse',
 '673': 'truck',
 '674': 'bird',
 '675': 'deer',
 '676': 'dog',
 '677': 'automobile',
 '678': 'deer',
 '679': 'cat',
 '680': 'truck',
 '681': 'frog',
 '682': 'dog',
 '683': 'frog',
 '684': 'truck',
 '685': 'cat',
 '686': 'cat',
 '687': 'dog',
 '688': 'airplane',
 '689': 'horse',
 '690': 'bird',
 '691': 'automobile',
 '692': 'cat',
 '693': 'frog',
 '694': 'deer',
 '695': 'airplane',
 '696': 'airplane',
 '697': 'bird',
 '698': 'dog',
 '699': 'airplane',
 '700': 'automobile',
 '701': 'airplane',
 '702': 'bird',
 '703': 'cat',
 '704': 'truck',
 '705': 'ship',
 '706': 'deer',
 '707': 'truck',
 '708': 'ship',
 '709': 'airplane',
 '710': 'bird',
 '711': 'frog',
 '712': 'deer',
 '713': 'deer',
 '714': 'airplane',
 '715': 'automobile',
 '716': 'ship',
 '717': 'ship',
 '718': 'cat',
 '719': 'frog',
 '720': 'truck',
 '721': 'frog',
 '722': 'frog',
 '723': 'horse',
 '724': 'ship',
 '725': 'bird',
 '726': 'deer',
 '727': 'dog',
 '728': 'horse',
 '729': 'frog',
 '730': 'dog',
 '731': 'cat',
 '732': 'airplane',
 '733': 'dog',
 '734': 'airplane',
 '735': 'dog',
 '736': 'airplane',
 '737': 'ship',
 '738': 'bird',
 '739': 'frog',
 '740': 'horse',
 '741': 'cat',
 '742': 'ship',
 '743': 'bird',
 '744': 'automobile',
 '745': 'horse',
 '746': 'frog',
 '747': 'horse',
 '748': 'automobile',
 '749': 'airplane',
 '750': 'truck',
 '751': 'dog',
 '752': 'dog',
 '753': 'airplane',
 '754': 'automobile',
 '755': 'horse',
 '756': 'frog',
 '757': 'truck',
 '758': 'airplane',
 '759': 'deer',
 '760': 'horse',
 '761': 'horse',
 '762': 'automobile',
 '763': 'dog',
 '764': 'truck',
 '765': 'deer',
 '766': 'airplane',
 '767': 'ship',
 '768': 'dog',
 '769': 'truck',
 '770': 'truck',
 '771': 'frog',
 '772': 'horse',
 '773': 'automobile',
 '774': 'ship',
 '775': 'cat',
 '776': 'bird',
 '777': 'cat',
 '778': 'ship',
 '779': 'bird',
 '780': 'bird',
 '781': 'deer',
 '782': 'frog',
 '783': 'airplane',
 '784': 'airplane',
 '785': 'dog',
 '786': 'cat',
 '787': 'ship',
 '788': 'bird',
 '789': 'cat',
 '790': 'horse',
 '791': 'bird',
 '792': 'truck',
 '793': 'cat',
 '794': 'ship',
 '795': 'horse',
 '796': 'ship',
 '797': 'bird',
 '798': 'horse',
 '799': 'truck',
 '800': 'airplane',
 '801': 'bird',
 '802': 'cat',
 '803': 'bird',
 '804': 'bird',
 '805': 'bird',
 '806': 'cat',
 '807': 'cat',
 '808': 'frog',
 '809': 'bird',
 '810': 'cat',
 '811': 'bird',
 '812': 'ship',
 '813': 'airplane',
 '814': 'dog',
 '815': 'dog',
 '816': 'automobile',
 '817': 'deer',
 '818': 'dog',
 '819': 'frog',
 '820': 'frog',
 '821': 'bird',
 '822': 'horse',
 '823': 'airplane',
 '824': 'automobile',
 '825': 'horse',
 '826': 'horse',
 '827': 'ship',
 '828': 'bird',
 '829': 'truck',
 '830': 'bird',
 '831': 'bird',
 '832': 'deer',
 '833': 'bird',
 '834': 'automobile',
 '835': 'automobile',
 '836': 'automobile',
 '837': 'frog',
 '838': 'frog',
 '839': 'frog',
 '840': 'dog',
 '841': 'automobile',
 '842': 'automobile',
 '843': 'horse',
 '844': 'airplane',
 '845': 'deer',
 '846': 'cat',
 '847': 'cat',
 '848': 'horse',
 '849': 'automobile',
 '850': 'bird',
 '851': 'cat',
 '852': 'dog',
 '853': 'dog',
 '854': 'dog',
 '855': 'frog',
 '856': 'automobile',
 '857': 'deer',
 '858': 'cat',
 '859': 'horse',
 '860': 'ship',
 '861': 'ship',
 '862': 'cat',
 '863': 'frog',
 '864': 'frog',
 '865': 'bird',
 '866': 'cat',
 '867': 'airplane',
 '868': 'truck',
 '869': 'deer',
 '870': 'cat',
 '871': 'ship',
 '872': 'airplane',
 '873': 'airplane',
 '874': 'automobile',
 '875': 'automobile',
 '876': 'dog',
 '877': 'deer',
 '878': 'truck',
 '879': 'cat',
 '880': 'automobile',
 '881': 'ship',
 '882': 'truck',
 '883': 'cat',
 '884': 'truck',
 '885': 'truck',
 '886': 'bird',
 '887': 'truck',
 '888': 'deer',
 '889': 'ship',
 '890': 'bird',
 '891': 'truck',
 '892': 'ship',
 '893': 'ship',
 '894': 'automobile',
 '895': 'dog',
 '896': 'cat',
 '897': 'frog',
 '898': 'ship',
 '899': 'horse',
 '900': 'frog',
 '901': 'truck',
 '902': 'ship',
 '903': 'airplane',
 '904': 'frog',
 '905': 'deer',
 '906': 'airplane',
 '907': 'airplane',
 '908': 'bird',
 '909': 'dog',
 '910': 'ship',
 '911': 'bird',
 '912': 'airplane',
 '913': 'bird',
 '914': 'horse',
 '915': 'frog',
 '916': 'truck',
 '917': 'horse',
 '918': 'automobile',
 '919': 'dog',
 '920': 'dog',
 '921': 'frog',
 '922': 'frog',
 '923': 'cat',
 '924': 'frog',
 '925': 'bird',
 '926': 'deer',
 '927': 'horse',
 '928': 'airplane',
 '929': 'dog',
 '930': 'frog',
 '931': 'deer',
 '932': 'frog',
 '933': 'dog',
 '934': 'bird',
 '935': 'deer',
 '936': 'frog',
 '937': 'automobile',
 '938': 'frog',
 '939': 'airplane',
 '940': 'deer',
 '941': 'airplane',
 '942': 'cat',
 '943': 'automobile',
 '944': 'ship',
 '945': 'dog',
 '946': 'deer',
 '947': 'deer',
 '948': 'automobile',
 '949': 'horse',
 '950': 'cat',
 '951': 'truck',
 '952': 'deer',
 '953': 'horse',
 '954': 'truck',
 '955': 'horse',
 '956': 'cat',
 '957': 'horse',
 '958': 'bird',
 '959': 'ship',
 '960': 'deer',
 '961': 'frog',
 '962': 'frog',
 '963': 'automobile',
 '964': 'bird',
 '965': 'truck',
 '966': 'airplane',
 '967': 'deer',
 '968': 'ship',
 '969': 'horse',
 '970': 'cat',
 '971': 'truck',
 '972': 'ship',
 '973': 'horse',
 '974': 'horse',
 '975': 'airplane',
 '976': 'bird',
 '977': 'deer',
 '978': 'automobile',
 '979': 'automobile',
 '980': 'deer',
 '981': 'automobile',
 '982': 'dog',
 '983': 'deer',
 '984': 'airplane',
 '985': 'dog',
 '986': 'frog',
 '987': 'bird',
 '988': 'ship',
 '989': 'dog',
 '990': 'airplane',
 '991': 'bird',
 '992': 'automobile',
 '993': 'cat',
 '994': 'dog',
 '995': 'horse',
 '996': 'cat',
 '997': 'dog',
 '998': 'automobile',
 '999': 'cat',
 '1000': 'dog',
 ...}

分离验证集

代码 6 · In [3]
new_root = 'train_valid_test'
def copyfile(filename, target_dir: Path):
    target_dir.mkdir(parents=True, exist_ok=True)
    shutil.copy(filename, target_dir)

def reorg_train_valid(data_dir: Path, labels: dict, valid_ratio: float):
    n = collections.Counter(labels.values()).most_common()[-1][1]
    n_valid_per_label = max(1, math.floor(n * valid_ratio))
    label_count = {}
    for train_file in os.listdir(data_dir/'train'):
        label = labels[train_file.split('.')[0]]
        fname = data_dir/'train'/train_file
        if label not in label_count or label_count[label] < n_valid_per_label:
            copyfile(fname, data_dir/new_root/'valid'/label)
            label_count[label] = label_count.get(label, 0) + 1
        else:
            copyfile(fname, data_dir/new_root/'train'/label)
    return n_valid_per_label

def reorg_test(data_dir: Path):
    for test_file in os.listdir(data_dir/'test'):
        copyfile(data_dir/'test'/test_file, data_dir/new_root/'test'/'unknown')

def reorg_cifar10_data(data_dir: Path, valid_ratio: float):
    labels = read_csv_labels(label_file)
    reorg_train_valid(data_dir, labels, valid_ratio)
    reorg_test(data_dir)
代码 7 · In [4]
reorg_cifar10_data(data_dir, valid_ratio=0.1)

图像增广

代码 9 · In [12]
normalize = torchvision.transforms.Normalize(mean=[0.4914, 0.4822, 0.4465],
                                             std=[0.2023, 0.1994, 0.2010])
train_aug = torchvision.transforms.Compose([
    torchvision.transforms.Resize(40), # 放大到 40x40
    torchvision.transforms.RandomResizedCrop(32, scale=(0.64, 1.0), ratio=(1.0, 1.0)), # 随机裁剪回 32x32
    torchvision.transforms.RandomHorizontalFlip(), # 随机水平翻转
    torchvision.transforms.ToTensor(), # 转为张量
    normalize
])
test_aug = torchvision.transforms.Compose([
    torchvision.transforms.Resize(32), # 调整图像大小为 32x32
    torchvision.transforms.ToTensor(), # 转为张量
    normalize
])

读取

代码 11 · In [13]
# 带增强
train_ds = torchvision.datasets.ImageFolder(data_dir/new_root/'train', transform=train_aug)
train_valid_ds = torchvision.datasets.ImageFolder(data_dir/new_root/'valid', transform=train_aug)
# 不带增强
valid_ds = torchvision.datasets.ImageFolder(data_dir/new_root/'valid', transform=test_aug)
test_ds = torchvision.datasets.ImageFolder(data_dir/new_root/'test', transform=test_aug)
代码 12 · In [14]
batch_size = 256
train_iter, train_valid_iter, valid_iter, test_iter = util.data.get_dataloader(
    train_ds, train_valid_ds, valid_ds, test_ds,
    batch_size=batch_size
)

模型

代码 14 · In [18]
def get_pretrained_net():
    net = torchvision.models.resnet18(weights=torchvision.models.ResNet18_Weights.DEFAULT)
    net.fc = nn.Linear(net.fc.in_features, 10)
    nn.init.xavier_normal_(net.fc.weight)
    return net

def get_scratch_net():
    return util.model.resnet18(num_classes=10, in_channels=3)

loss = nn.CrossEntropyLoss(reduction="none")

训练

代码 16 · In [19]
def train(net,
          train_iter, valid_iter,
          num_epochs,
          lr, wd,
          devices,
          lr_period, lr_decay,
          finetune=True):
    if finetune:
        trainer = torch.optim.Adam([
            {'params': [param for name, param in net.named_parameters() if name not in ['fc.weight', 'fc.bias']]},
            {'params': net.fc.parameters(), 'lr': lr * 10}
        ], lr=lr, weight_decay=wd)
    else:
        trainer = torch.optim.Adam(net.parameters(), lr=lr, weight_decay=wd)
    scheduler = torch.optim.lr_scheduler.StepLR(trainer, step_size=lr_period, gamma=lr_decay)
    num_batches = len(train_iter)
    timer = util.Timer()
    legend = ['train loss', 'train acc']
    if valid_iter is not None:
        legend.append('valid acc')
    animator = util.Animator(xlabel='epoch', xlim=[1, num_epochs], legend=legend, ylim=[0.2, 1])
    net = nn.DataParallel(net, device_ids=devices).to(devices[0])
    for epoch in range(num_epochs):
        net.train()
        metric = util.Accumulator(3)
        for i, (X, y) in enumerate(train_iter):
            timer.start()
            l, acc = util.train.train_batch_ch13(net, X, y, loss, trainer, devices)
            metric.add(l, acc, y.shape[0])
            timer.stop()
            # 每个 epoch 画两次
            if (i + 1) % (num_batches // 2) == 0 or i == num_batches - 1:
                animator.add(epoch + (i + 1) / num_batches,
                             (metric[0] / metric[2], metric[1] / metric[2], None))
        if valid_iter is not None:
            valid_acc = util.train.evaluate_accuracy_gpu(net, valid_iter)
            animator.add(epoch + 1, (None, None, valid_acc))
        scheduler.step()
    measures = (f'train loss {metric[0] / metric[2]:.3f}, ',
                f'train acc {metric[1] / metric[2]:.3f}')
    if valid_iter is not None:
        measures += (f'valid acc {valid_acc:.3f}, ',)
    print(measures + (f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on {str(devices)}',))
代码 17
devices = util.try_all_gpus()
net = get_scratch_net()
num_epochs = 100
lr = 2e-4
wd = 5e-4
lr_period = 4 # 每隔 4 个 epoch 学习率衰减一次
lr_decay = 0.9 # 每次衰减为原来的 0.9 倍
train(net,
      train_iter, valid_iter,
      num_epochs=num_epochs,
      lr=lr, wd=wd,
      devices=devices,
      lr_period=lr_period, lr_decay=lr_decay,
      finetune=False)
已保存输出
('train loss 0.006, ', 'train acc 0.998', 'valid acc 0.969, ', "9931.7 examples/sec on [device(type='cuda', index=0)]")

代码 17 的已保存图表

预测

代码 19
def predict(net, test_iter):
    preds = []
    for X, _ in test_iter:
        y_hat = net(X.to(devices[0]))
        y_hat = y_hat.argmax(axis=1)
        preds.extend(y_hat.type(torch.int32).cpu().numpy())
        if len(preds) % 2000 == 0:
            print(f'Processed {len(preds)} examples, {len(test_iter.dataset)-len(preds)} left')
    sorted_ids = list(range(1, len(test_ds) + 1))
    sorted_ids.sort(key=lambda x: str(x))
    df = pd.DataFrame({'id': sorted_ids, 'label': preds})
    df['label'] = df['label'].map(lambda x: train_valid_ds.classes[x])
    return df

pred = predict(net, test_iter)
pred.to_csv(data_dir/'submission.csv', index=False)