spcl+在UDA上不能复现结果

[keenJMS]

葛博，你好。我用的是4卡2080Ti，pytorch1.7+cuda10.1+python3.8.5，spcl+ 在duke->msmt map只有22 ，在market->msmt只有23.3，下面是我的log文件

==========

Args:Namespace(config='SpCL/config_duke_msmt.yaml', launcher='pytorch', resume_from=None, set_cfgs=None, tcp_port='10010', work_dir='SpCL/duke_msmt/4gpu_16per/800iter')

==========
cfg.LOCAL_RANK: 0
cfg.DATA_ROOT: ../datasets
cfg.LOGS_ROOT: /data/OpenUnlogs/logs

cfg.MODEL = edict()
cfg.MODEL.backbone: resnet50
cfg.MODEL.pooling: gem
cfg.MODEL.embed_feat: 0
cfg.MODEL.dropout: 0.0
cfg.MODEL.dsbn: True
cfg.MODEL.sync_bn: True
cfg.MODEL.samples_per_bn: 16
cfg.MODEL.mean_net: False
cfg.MODEL.alpha: 0.999
cfg.MODEL.imagenet_pretrained: True
cfg.MODEL.source_pretrained: None

cfg.DATA = edict()
cfg.DATA.height: 256
cfg.DATA.width: 128
cfg.DATA.norm_mean: [0.485, 0.456, 0.406]
cfg.DATA.norm_std: [0.229, 0.224, 0.225]

cfg.DATA.TRAIN = edict()
cfg.DATA.TRAIN.is_autoaug: False
cfg.DATA.TRAIN.is_flip: True
cfg.DATA.TRAIN.flip_prob: 0.5
cfg.DATA.TRAIN.is_pad: True
cfg.DATA.TRAIN.pad_size: 10
cfg.DATA.TRAIN.is_blur: False
cfg.DATA.TRAIN.blur_prob: 0.5
cfg.DATA.TRAIN.is_erase: True
cfg.DATA.TRAIN.erase_prob: 0.5
cfg.DATA.TRAIN.is_mutual_transform: False
cfg.DATA.TRAIN.mutual_times: 2

cfg.TRAIN = edict()
cfg.TRAIN.seed: 1
cfg.TRAIN.deterministic: True
cfg.TRAIN.amp: False

cfg.TRAIN.datasets = edict()
cfg.TRAIN.datasets.msmt17: trainval
cfg.TRAIN.datasets.dukemtmcreid: trainval
cfg.TRAIN.unsup_dataset_indexes: [0]
cfg.TRAIN.epochs: 50
cfg.TRAIN.iters: 800

cfg.TRAIN.LOSS = edict()

cfg.TRAIN.LOSS.losses = edict()
cfg.TRAIN.LOSS.losses.hybrid_memory: 1.0
cfg.TRAIN.LOSS.temp: 0.05
cfg.TRAIN.LOSS.momentum: 0.2
cfg.TRAIN.val_dataset: msmt17
cfg.TRAIN.val_freq: 5

cfg.TRAIN.SAMPLER = edict()
cfg.TRAIN.SAMPLER.num_instances: 4
cfg.TRAIN.SAMPLER.is_shuffle: True

cfg.TRAIN.LOADER = edict()
cfg.TRAIN.LOADER.samples_per_gpu: 16
cfg.TRAIN.LOADER.workers_per_gpu: 2

cfg.TRAIN.PSEUDO_LABELS = edict()
cfg.TRAIN.PSEUDO_LABELS.freq: 1
cfg.TRAIN.PSEUDO_LABELS.use_outliers: True
cfg.TRAIN.PSEUDO_LABELS.norm_feat: True
cfg.TRAIN.PSEUDO_LABELS.norm_center: True
cfg.TRAIN.PSEUDO_LABELS.cluster: dbscan
cfg.TRAIN.PSEUDO_LABELS.eps: [0.58, 0.6, 0.62]
cfg.TRAIN.PSEUDO_LABELS.min_samples: 4
cfg.TRAIN.PSEUDO_LABELS.dist_metric: jaccard
cfg.TRAIN.PSEUDO_LABELS.k1: 30
cfg.TRAIN.PSEUDO_LABELS.k2: 6
cfg.TRAIN.PSEUDO_LABELS.search_type: 0
cfg.TRAIN.PSEUDO_LABELS.cluster_num: None

cfg.TRAIN.OPTIM = edict()
cfg.TRAIN.OPTIM.optim: adam
cfg.TRAIN.OPTIM.lr: 0.00035
cfg.TRAIN.OPTIM.weight_decay: 0.0005

cfg.TRAIN.SCHEDULER = edict()
cfg.TRAIN.SCHEDULER.lr_scheduler: single_step
cfg.TRAIN.SCHEDULER.stepsize: 20
cfg.TRAIN.SCHEDULER.gamma: 0.1

cfg.TEST = edict()
cfg.TEST.datasets: ['msmt17']

cfg.TEST.LOADER = edict()
cfg.TEST.LOADER.samples_per_gpu: 32
cfg.TEST.LOADER.workers_per_gpu: 2
cfg.TEST.dist_metric: euclidean
cfg.TEST.norm_feat: True
cfg.TEST.dist_cuda: True
cfg.TEST.rerank: False
cfg.TEST.search_type: 0
cfg.TEST.k1: 20
cfg.TEST.k2: 6
cfg.TEST.lambda_value: 0.3
cfg.launcher: pytorch
cfg.tcp_port: 10010
cfg.work_dir: /data/OpenUnlogs/logs/SpCL/duke_msmt/4gpu_16per/800iter
cfg.rank: 0
cfg.ngpus_per_node: 4
cfg.gpu: 0
cfg.total_gpus: 4
cfg.world_size: 4
The training is in a un/semi-supervised manner with 2 dataset(s) (['msmt17', 'dukemtmcreid']),
where ['msmt17'] have no labels.

Mean AP: 22.0%
CMC Scores:
top-1 46.6%
top-5 59.3%
top-10 64.6%
Testing time: 0:03:25.443005

******************************* Finished testing *******************************

Total running time: 5:10:33.865417