How to reproduce paper's result

[cswwp]

Hi, @amaralibey thank you for share so great repo. Now i'm reproducing the paper's result, but i can't get same result as paper shows with just modify the data path in this repo. the best pitts250k test top1 recall is 92.85 as follows, not 94.6 in paper

and my parameters is here

agg_arch: MixVPR
agg_config:
in_channels: 1024
in_h: 20
in_w: 20
mix_depth: 4
mlp_ratio: 1
out_channels: 1024
out_rows: 4
backbone_arch: resnet50
batch_sampler: null
batch_size: 120
cities:

• Bangkok
• BuenosAires
• LosAngeles
• MexicoCity
• OSL
• Rome
• Barcelona
• Chicago
• Madrid
• Miami
• Phoenix
• TRT
• Boston
• Lisbon
• Medellin
• Minneapolis
• PRG
• WashingtonDC
• Brussels
• London
• Melbourne
• Osaka
• PRS
  faiss_gpu: false
  image_size: !!python/tuple
• 320
• 320
  img_per_place: 4
  layers_to_crop:
• 4
  layers_to_freeze: 2
  loss_name: MultiSimilarityLoss
  lr: 0.05
  lr_mult: 0.3
  mean_std:
  mean:
  • 0.485
  • 0.456
  • 0.406
    std:
  • 0.229
  • 0.224
  • 0.225
    milestones:
• 5
• 10
• 15
• 25
• 45
  min_img_per_place: 4
  miner_margin: 0.1
  miner_name: MultiSimilarityMiner
  momentum: 0.9
  num_workers: 28
  optimizer: sgd
  pretrained: true
  random_sample_from_each_place: true
  show_data_stats: true
  shuffle_all: false
  val_set_names:
• pitts250k_test
  warmpup_steps: 650
  weight_decay: 0.001

Do you mind to supply your parameters to reproduce your result? thank you

[lv162]

hello,@cswwp,I also want to use this as a comparative experiment. Can you share how you debugged it or how you wrote the evaluation code? I want to use the code to test other datasets.Thank you very much.