一、模型配置

一、backbone
resnet50
二、neck
GlobalAveragePooling
三、head
fc
四、loss
type=‘LabelSmoothLoss’,
label_smooth_val=0.1,
num_classes=30,
reduction=‘mean’,
loss_weight=1.0
五、optimizer
lr=0.1, momentum=0.9, type=‘SGD’, weight_decay=0.0001
六、scheduler
T_max=260, begin=20, by_epoch=True, end=300, type='CosineAnnealingLR
七、evaluator
topk=(1, 5 ), type=‘Accuracy’
八、max_epochs
300

九、Config

auto_scale_lr = dict(base_batch_size=256)
data_preprocessor = dict(
    mean=[
        123.675,
        116.28,
        103.53,
    ],
    num_classes=30,
    std=[
        58.395,
        57.12,
        57.375,
    ],
    to_rgb=True)
dataset_type = 'ImageNet'
data_root = 'data/PlantsClassification'
default_hooks = dict(
    checkpoint=dict(interval=1, type='CheckpointHook', max_keep_ckpts=2, save_best="auto"),
    logger=dict(interval=100, type='LoggerHook'),
    param_scheduler=dict(type='ParamSchedulerHook'),
    sampler_seed=dict(type='DistSamplerSeedHook'),
    timer=dict(type='IterTimerHook'),
    visualization=dict(enable=False, type='VisualizationHook'))
default_scope = 'mmpretrain'
env_cfg = dict(
    cudnn_benchmark=False,
    dist_cfg=dict(backend='nccl'),
    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0))
launcher = 'none'
load_from = './work_dirs/resnet50_8xb32-coslr_in1k/resnet50_8xb32_in1k_20210831-ea4938fc.pth'
log_level = 'INFO'
model = dict(
    backbone=dict(
        depth=50,
        num_stages=4,
        out_indices=(3,),
        style='pytorch',
        type='ResNet'),
    head=dict(
        in_channels=2048,
        # loss=dict(loss_weight=1.0, type='CrossEntropyLoss'),
        loss=dict(
                    type='LabelSmoothLoss',
                    label_smooth_val=0.1,
                    num_classes=30,
                    reduction='mean',
                    loss_weight=1.0),
        num_classes=30,
        topk=(
            1,
            5,
        ),
        type='LinearClsHead'),
    data_preprocessor=data_preprocessor,
    neck=dict(type='GlobalAveragePooling'),
    type='ImageClassifier')
train_cfg = dict(by_epoch=True, max_epochs=300, val_interval=1)
optim_wrapper = dict(
    optimizer=dict(lr=0.1, momentum=0.9, type='SGD', weight_decay=0.0001))
param_scheduler = dict(
    T_max=260, begin=20, by_epoch=True, end=300, type='CosineAnnealingLR')
randomness = dict(deterministic=False, seed=None)
resume = False
test_cfg = dict()
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(edge='short', scale=256, type='ResizeEdge'),
    dict(crop_size=224, type='CenterCrop'),
    dict(type='PackInputs'),
]
test_dataloader = dict(
    batch_size=32,
    collate_fn=dict(type='default_collate'),
    dataset=dict(
        data_root=data_root,
        pipeline=test_pipeline,
        split='test',
        ann_file='test.txt',
        type=dataset_type),
    num_workers=1,
    persistent_workers=True,
    pin_memory=True,
    sampler=dict(shuffle=False, type='DefaultSampler'))
test_evaluator = dict(
    topk=(
        1,
        5,
    ), type='Accuracy')

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(scale=224, type='RandomResizedCrop'),
    dict(direction='horizontal', prob=0.5, type='RandomFlip'),
    dict(type='PackInputs'),
]
train_dataloader = dict(
    batch_size=45,
    collate_fn=dict(type='default_collate'),
    dataset=dict(
        data_root=data_root,
        pipeline=train_pipeline,
        split='train',
        ann_file='train.txt',
        type=dataset_type),
    num_workers=1,
    persistent_workers=True,
    pin_memory=True,
    sampler=dict(shuffle=True, type='DefaultSampler'))

val_cfg = dict()
val_dataloader = dict(
    batch_size=45,
    collate_fn=dict(type='default_collate'),
    dataset=dict(
        data_root=data_root,
        pipeline=test_pipeline,
        split='val',
        ann_file='val.txt',
        type=dataset_type),
    num_workers=1,
    persistent_workers=True,
    pin_memory=True,
    sampler=dict(shuffle=False, type='DefaultSampler'))
val_evaluator = test_evaluator
vis_backends = [
    dict(type='LocalVisBackend'),
]
visualizer = dict(
    type='UniversalVisualizer', vis_backends=[
        dict(type='LocalVisBackend'),
    ])
work_dir = './work_dirs\\resnet50_8xb32-coslr_in1k'

二、训练结果

采用kaggle植物分类数据集，30分类，标签：
IMAGENET_CATEGORIES = [‘aloevera’, ‘banana’, ‘bilimbi’, ‘cantaloupe’, ‘cassava’, ‘coconut’, ‘corn’, ‘cucumber’,
‘curcuma’, ‘eggplant’, ‘galangal’, ‘ginger’, ‘guava’, ‘kale’, ‘longbeans’, ‘mango’, ‘melon’,
‘orange’, ‘paddy’, ‘papaya’, ‘peperchili’, ‘pineapple’, ‘pomelo’, ‘shallot’, ‘soybeans’,
‘spinach’, ‘sweetpotatoes’, ‘tobacco’, ‘waterapple’, ‘watermelon’]
“accuracy/top1”: 90.00000762939453, “accuracy/top5”: 98.0666732788086

三、结果分析

分析分类结果发现，更改不同的训练策略，结果不会增加，且总有个别类别分类错误，仔细分析数据发现，引起该问题的主要原因是数据集本身引起的，在个别类别中混入了其他类别的图片，甚至出现两个类别使用的数据完全一致的情况，比如melon和cantaloupe使用的是相同的数据集，且内部混入了西瓜的数据集

四、预测测试

五、总结

在训练前，必须检查数据集