Can't reproduce MSMT17 performance

[RRoundTable]

Based on README, the performace of ViT-B/16+ICS is 75.1%. But I got 71.2% with MSMT17_V2 dataset. ViT-B/16+ICS is evaluated on MSMT17_V2?

``` 023-07-05 13:54:03 transreid INFO: Namespace(config_file='configs/msmt17/vit_base_ics_384.yml', opts=['MODEL.DEVICE_ID', "('0')"]) 2023-07-05 13:54:03 transreid INFO: Loaded configuration file configs/msmt17/vit_base_ics_384.yml 2023-07-05 13:54:03 transreid INFO: MODEL: PRETRAIN_PATH: '/home/hpds/Repositories/ml-models/proto/TransReID-SSL/checkpoint/vit_base_ics_cfs_lup.pth' PRETRAIN_HW_RATIO: 2 METRIC_LOSS_TYPE: 'triplet' IF_LABELSMOOTH: 'off' IF_WITH_CENTER: 'no' NAME: 'transformer' NO_MARGIN: True DEVICE_ID: ('2') TRANSFORMER_TYPE: 'vit_base_patch16_224_TransReID' STRIDE_SIZE: [16, 16] STEM_CONV: True # False for vanilla ViT-S # DIST_TRAIN: True

INPUT:
SIZE_TRAIN: [384, 128]
SIZE_TEST: [384, 128]
PROB: 0.5 # random horizontal flip
RE_PROB: 0.5 # random erasing
PADDING: 10
PIXEL_MEAN: [0.5, 0.5, 0.5]
PIXEL_STD: [0.5, 0.5, 0.5]

DATASETS:
NAMES: ('MSMT17_V2')
ROOT_DIR: ('/home/hpds/Repositories/ml-models/dataset')

DATALOADER:
SAMPLER: 'softmax_triplet'
NUM_INSTANCE: 4
NUM_WORKERS: 8

SOLVER:
OPTIMIZER_NAME: 'SGD'
MAX_EPOCHS: 120
BASE_LR: 0.0004
WARMUP_EPOCHS: 20
IMS_PER_BATCH: 64
WARMUP_METHOD: 'cosine'
LARGE_FC_LR: False
CHECKPOINT_PERIOD: 120
LOG_PERIOD: 20
EVAL_PERIOD: 120
WEIGHT_DECAY: 1e-4
WEIGHT_DECAY_BIAS: 1e-4
BIAS_LR_FACTOR: 2

TEST:
EVAL: True
IMS_PER_BATCH: 256
RE_RANKING: False
WEIGHT: '/home/hpds/Repositories/ml-models/proto/TransReID-SSL/checkpoint/transformer_120.pth'
NECK_FEAT: 'before'
FEAT_NORM: 'yes'

OUTPUT_DIR: '../../log/transreid/msmt17/vit_base_ics_cfs_lup_384'

2023-07-05 13:54:03 transreid INFO: Running with config:
DATALOADER:
NUM_INSTANCE: 4
NUM_WORKERS: 8
REMOVE_TAIL: 0
SAMPLER: softmax_triplet
DATASETS:
NAMES: MSMT17_V2
ROOT_DIR: /home/hpds/Repositories/ml-models/dataset
ROOT_TRAIN_DIR: ../data
ROOT_VAL_DIR: ../data
INPUT:
PADDING: 10
PIXEL_MEAN: [0.5, 0.5, 0.5]
PIXEL_STD: [0.5, 0.5, 0.5]
PROB: 0.5
RE_PROB: 0.5
SIZE_TEST: [384, 128]
SIZE_TRAIN: [384, 128]
MODEL:
ATT_DROP_RATE: 0.0
COS_LAYER: False
DEVICE: cuda
DEVICE_ID: 0
DEVIDE_LENGTH: 4
DIST_TRAIN: False
DROPOUT_RATE: 0.0
DROP_OUT: 0.0
DROP_PATH: 0.1
FEAT_DIM: 512
GEM_POOLING: False
ID_LOSS_TYPE: softmax
ID_LOSS_WEIGHT: 1.0
IF_LABELSMOOTH: off
IF_WITH_CENTER: no
JPM: False
LAST_STRIDE: 1
METRIC_LOSS_TYPE: triplet
NAME: transformer
NECK: bnneck
NO_MARGIN: True
PRETRAIN_CHOICE: imagenet
PRETRAIN_HW_RATIO: 2
PRETRAIN_PATH: /home/hpds/Repositories/ml-models/proto/TransReID-SSL/checkpoint/vit_base_ics_cfs_lup.pth
REDUCE_FEAT_DIM: False
RE_ARRANGE: True
SHIFT_NUM: 5
SHUFFLE_GROUP: 2
SIE_CAMERA: False
SIE_COE: 3.0
SIE_VIEW: False
STEM_CONV: True
STRIDE_SIZE: [16, 16]
TRANSFORMER_TYPE: vit_base_patch16_224_TransReID
TRIPLET_LOSS_WEIGHT: 1.0
OUTPUT_DIR: ../../log/transreid/msmt17/vit_base_ics_cfs_lup_384
SOLVER:
BASE_LR: 0.0004
BIAS_LR_FACTOR: 2
CENTER_LOSS_WEIGHT: 0.0005
CENTER_LR: 0.5
CHECKPOINT_PERIOD: 120
COSINE_MARGIN: 0.5
COSINE_SCALE: 30
EVAL_PERIOD: 120
GAMMA: 0.1
IMS_PER_BATCH: 64
LARGE_FC_LR: False
LOG_PERIOD: 20
MARGIN: 0.3
MAX_EPOCHS: 120
MOMENTUM: 0.9
OPTIMIZER_NAME: SGD
SEED: 1234
STEPS: (40, 70)
TRP_L2: False
WARMUP_EPOCHS: 20
WARMUP_FACTOR: 0.01
WARMUP_METHOD: cosine
WEIGHT_DECAY: 0.0001
WEIGHT_DECAY_BIAS: 0.0001
TEST:
DIST_MAT: dist_mat.npy
EVAL: True
FEAT_NORM: yes
IMS_PER_BATCH: 256
NECK_FEAT: before
RE_RANKING: False
WEIGHT: /home/hpds/Repositories/ml-models/proto/TransReID-SSL/checkpoint/transformer_120.pth
MSMT17_V2 /home/hpds/Repositories/ml-models/dataset
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15} cam_container
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15} cam_container
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15} cam_container
{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15} cam_container
=> MSMT17 loaded
2023-07-05 13:54:03 transreid.check INFO: Dataset statistics:
2023-07-05 13:54:03 transreid.check INFO: ----------------------------------------
2023-07-05 13:54:03 transreid.check INFO: subset | # ids | # images | # cameras
2023-07-05 13:54:03 transreid.check INFO: ----------------------------------------
2023-07-05 13:54:03 transreid.check INFO: train | 1041 | 32621 | 15
2023-07-05 13:54:03 transreid.check INFO: query | 3060 | 11659 | 15
2023-07-05 13:54:03 transreid.check INFO: gallery | 3060 | 82161 | 15
2023-07-05 13:54:03 transreid.check INFO: ----------------------------------------
using img_triplet sampler
using Transformer_type: vit_base_patch16_224_TransReID as a backbone
using stride: [16, 16], and patch number is num_y24 * num_x8
Resized position embedding from size:torch.Size([1, 129, 768]) to size: torch.Size([1, 193, 768]) with height:24 width: 8
Load 172 / 174 layers.
Loading pretrained ImageNet model......from /home/hpds/Repositories/ml-models/proto/TransReID-SSL/checkpoint/vit_base_ics_cfs_lup.pth
===========building transformer===========
Loading pretrained model from /home/hpds/Repositories/ml-models/proto/TransReID-SSL/checkpoint/transformer_120.pth
2023-07-05 13:54:05 transreid.test INFO: Enter inferencing
True
torch.cuda.device_count() 1

The test feature is normalized
=> Computing DistMat with euclidean_distance
/home/hpds/Repositories/ml-models/proto/TransReID-SSL/transreid_pytorch/utils/metrics.py:12: UserWarning: This overload of addmm_ is deprecated:
addmm_(Number beta, Number alpha, Tensor mat1, Tensor mat2)
Consider using one of the following signatures instead:
addmm_(Tensor mat1, Tensor mat2, *, Number beta, Number alpha) (Triggered internally at ../torch/csrc/utils/python_arg_parser.cpp:1485.)
dist_mat.addmm_(1, -2, qf, gf.t())
distmat (11659, 82161) <class 'numpy.ndarray'>
2023-07-05 14:04:03 transreid.test INFO: Validation Results
2023-07-05 14:04:03 transreid.test INFO: mAP: 71.2%
2023-07-05 14:04:03 transreid.test INFO: CMC curve, Rank-1 :87.9%
2023-07-05 14:04:03 transreid.test INFO: CMC curve, Rank-5 :93.6%
2023-07-05 14:04:03 transreid.test INFO: CMC curve, Rank-10 :95.1%