DATA: dataset: classification dataset_json_file: /data02/xy/dataEngine/json_data/LuojiaHOG(test)_.json # dataset_json_file: /data02/xy/dataEngine/json_data/merged_output_combined_9w_resplit.json # dataset_json_file: /data02/xy/dataEngine/json_data/merged_output_combined_9w_resplit.json exp_name: classifi ratio: 0 dataset_train_split: 0.6 dataset_query_split: 0.2 imgs_folder: /data02/xy/Clip-hash/datasets/image/ label_path: /data02/xy/Clip-hash/labels.txt num_classes: 10 # num_classes: 131 TRAIN: # Base Arch # clip_pretrain: /data02/xy/Clip-hash/pretrain/RS5M_ViT-B-32.pt clip_pretrain: ./cisen/pretrain/RS5M_ViT-B-32.pt model_name: ViT-B-32 ckpt_path: /data02/xy/GeoRSCLIP/codebase/inference/pretrain/RS5M_ViT-B-32.pt input_size: 224 word_len: 328 word_dim: 1024 vis_dim: 512 fpn_in: [ 512, 768, 768 ] fpn_out: [ 768, 768, 768, 512 ] sync_bn: True # Decoder num_layers: 3 num_head: 8 dim_ffn: 2048 dropout: 0.1 intermediate: False # Training Setting workers: 32 # data loader workers workers_val: 16 epochs: 50 milestones: [50] start_epoch: 0 batch_size: 256 # batch size for training batch_size_val: 256 # batch size for validation during training, memory and speed tradeoff 11111 base_lr: 0.0001 min_lr: 0.00000001 lr_decay: 0.5 lr_multi: 0.1 weight_decay: 0. max_norm: 0. manual_seed: 0 print_freq: 1 lamda1: 0.5 lamda2: 0.5 beta1: 0.5 beta2: 0.5 eta: 0.2 warmup_epochs: 0 contrastive: [0.4, 0.3, 0.3] # Resume & Save output_folder: /data02/xy/Clip-hash/exp/ save_freq: 1 weight: # path to initial weight (default: none) resume: False # path to latest checkpoint (default: none) evaluate: True # evaluate on validation set, extra gpu memory needed and small batch_size_val is recommend Distributed: dist_url: tcp://localhost:3693 dist_backend: 'nccl' multiprocessing_distributed: True world_size: 1 rank: 0 TEST: test_split: val-test gpu : [0] test_lmdb: /data02/xy/Clip-hash/datasets/lmdb/refcoco/val.lmdb visualize: False topk: 5 test_batch_size: 256 #1111111 val_batch_size: 1