使用GeneralRecognitionV2_PPLCNetV2_base.yaml训练自己的数据集，有些参数如何调整

[yaphet266]

是否有文档解释下这个yaml文件中每个配置项的含义

global configs

Global: checkpoints: null pretrained_model: null output_dir: ./output device: gpu save_interval: 1 eval_during_train: True eval_interval: 1 epochs: 100 print_batch_step: 20 use_visualdl: False eval_mode: retrieval retrieval_feature_from: features # 'backbone' or 'features' re_ranking: False use_dali: False

used for static mode and model export

image_shape: [3, 224, 224] save_inference_dir: ./inference

AMP: scale_loss: 65536 use_dynamic_loss_scaling: True

O1: mixed fp16

level: O1

model architecture

Arch: name: RecModel infer_output_key: features infer_add_softmax: False

Backbone: name: PPLCNetV2_base_ShiTu pretrained: True use_ssld: True class_expand: &feat_dim 512 BackboneStopLayer: name: flatten Neck: name: BNNeck num_features: feat_dim weight_attr: initializer: name: Constant value: 1.0 bias_attr: initializer: name: Constant value: 0.0 learning_rate: 1.0e-20 # NOTE: Temporarily set lr small enough to freeze the bias to zero Head: name: FC embedding_size: feat_dim class_num: 192612 weight_attr: initializer: name: Normal std: 0.001 bias_attr: False

loss function config for traing/eval process

Loss: Train:

• CELoss: weight: 1.0 epsilon: 0.1
• TripletAngularMarginLoss: weight: 1.0 feature_from: features margin: 0.5 reduction: mean add_absolute: True absolute_loss_weight: 0.1 normalize_feature: True ap_value: 0.8 an_value: 0.4 Eval:
• CELoss: weight: 1.0

Optimizer: name: Momentum momentum: 0.9 lr: name: Cosine learning_rate: 0.06 # for 8gpu x 256bs warmup_epoch: 5 regularizer: name: L2 coeff: 0.00001

data loader for train and eval

DataLoader: Train: dataset: name: ImageNetDataset image_root: ./dataset/ cls_label_path: ./dataset/train_reg_all_data_v2.txt relabel: True transform_ops:

• DecodeImage: to_rgb: True channel_first: False
• ResizeImage: size: [224, 224] return_numpy: False interpolation: bilinear backend: cv2
• RandFlipImage: flip_code: 1
• Pad: padding: 10 backend: cv2
• RandCropImageV2: size: [224, 224]
• RandomRotation: prob: 0.5 degrees: 90 interpolation: bilinear
• ResizeImage: size: [224, 224] return_numpy: False interpolation: bilinear backend: cv2

• NormalizeImage: scale: 1.0/255.0 mean: [0.485, 0.456, 0.406] std: [0.229, 0.224, 0.225] order: hwc sampler: name: PKSampler batch_size: 256 sample_per_id: 4 drop_last: False shuffle: True sample_method: "id_avg_prob" id_list: [50030, 80700, 92019, 96015] # be careful when set relabel=True ratio: [4, 4] loader: num_workers: 4 use_shared_memory: True

  Eval: Query: dataset: name: VeriWild image_root: ./dataset/Aliproduct/ cls_label_path: ./dataset/Aliproduct/val_list.txt transform_ops:

  • DecodeImage: to_rgb: True channel_first: False
  • ResizeImage: size: [224, 224] return_numpy: False interpolation: bilinear backend: cv2
  • NormalizeImage: scale: 1.0/255.0 mean: [0.485, 0.456, 0.406] std: [0.229, 0.224, 0.225] order: hwc sampler: name: DistributedBatchSampler batch_size: 64 drop_last: False shuffle: False loader: num_workers: 4 use_shared_memory: True

  Gallery: dataset: name: VeriWild image_root: ./dataset/Aliproduct/ cls_label_path: ./dataset/Aliproduct/val_list.txt transform_ops:

  • DecodeImage: to_rgb: True channel_first: False
  • ResizeImage: size: [224, 224] return_numpy: False interpolation: bilinear backend: cv2
  • NormalizeImage: scale: 1.0/255.0 mean: [0.485, 0.456, 0.406] std: [0.229, 0.224, 0.225] order: hwc sampler: name: DistributedBatchSampler batch_size: 64 drop_last: False shuffle: False loader: num_workers: 4 use_shared_memory: True

Metric: Eval:

• Recallk: topk: [1, 5]
• mAP: {}

[TingquanGao]

修改数据集路径(image_root: ./dataset/ cls_label_path: ./dataset/train_reg_all_data_v2.txt)后可以先训练试试，观察loss是否下降，已经最终的收敛情况、精度情况，再适当调整learning rate。

[yaphet266]

@TingquanGao 如果主干网络想改成ResNet50怎么修改，是否有可参考的yaml文件

[changdazhou]

更改一下Arch:即可