PaddlePaddle

PaddleDetection

dataset

coco

faster_rcnn_dcn_r50_fpn_1x.yml

类别+1(背景)

max_iters: 120000
use_gpu: true
snapshot_iter: 10000
log_iter: 1
save_dir: output
pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_dcn_r50_fpn_1x.tar
metric: COCO
weights: output/faster_rcnn_dcn_r50_fpn_1x/model_final
num_classes: 4
BBoxHead:
  head: TwoFCHead
  nms:
    keep_top_k: 15
    nms_threshold: 0.4
    score_threshold: 0.05
LearningRate:
  base_lr: 0.005
  schedulers:
     - !PiecewiseDecay
         gamma: 0.1
         milestones: [80000, 100000]
READER: ‘faster_fpn_reader.yml’
TrainReader:
    batch_size: 1

faster_rcnn_dcn_r50_fpn_1x.yml

TrainReader:
  inputs_def:
      fields: [‘image’, ‘im_info’, ‘im_id’, ‘gt_bbox’, ‘gt_class’, ‘is_crowd’]
  dataset:
      !COCODataSet
      image_dir: imgs
      anno_path: annotations/train.json
      dataset_dir: /data/yjl/dapeng
  batch_size: 1
  shuffle: true
  drop_empty: false
  worker_num: 2
  use_process: false
EvalReader:
  inputs_def:
      fields: [‘image’, ‘im_info’, ‘im_id’, ‘im_shape’]
      # for voc
      #fields: [‘image’, ‘im_info’, ‘im_id’, ‘im_shape’, ‘gt_bbox’, ‘gt_class’, ‘is_difficult’]
  dataset:
      !COCODataSet
      image_dir: imgs
      anno_path: annotations/valid.json
      dataset_dir: /data/yjl/dapeng
  batch_size: 1
  shuffle: false
  drop_empty: false
  worker_num: 2
TestReader:
  inputs_def:
      # set image_shape if needed
      fields: [‘image’, ‘im_info’, ‘im_id’, ‘im_shape’]
  dataset:
      !ImageFolder
      anno_path: /data/yjl/dapeng/annotations/valid.json

voc

ppyolo.yml

类别数不加背景

architecture: YOLOv3
use_gpu: true
max_iters: 130000
log_iter: 20
save_dir: output
snapshot_iter: 10000
metric: VOC
pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/ppyolo.pdparams
weights: output/ppyolo/model_final
num_classes: 1
use_fine_grained_loss: true
use_ema: true
ema_decay: 0.9998
LearningRate:
    base_lr: 0.002
    schedulers:
      - !PiecewiseDecay
        gamma: 0.1
        milestones:
        - 120000
        - 125000

ppyolo_reader.yml

TrainReader:
   inputs_def:
      fields: [‘image’, ‘gt_bbox’, ‘gt_class’, ‘gt_score’]
      num_max_boxes: 50
   dataset:
      !VOCDataSet
      anno_path: train.txt
       dataset_dir: /data/yjl/dapeng
       with_background: false
       use_default_label: false
   batch_size: 6
   shuffle: true
   mixup_epoch: 200
   drop_last: true
   worker_num: 8
   bufsize: 4
   use_process: true
EvalReader:
   inputs_def:
     fields: [‘image’, ‘im_size’, ‘im_id’, ‘gt_bbox’, ‘gt_class’, ‘is_difficult’]
     num_max_boxes: 50
   dataset:
     !VOCDataSet
     anno_path: valid.txt
     dataset_dir: /data/yjl/dapeng
     with_background: false
     use_default_label: false
   batch_size: 1
   drop_empty: false
   worker_num: 8
   bufsize: 4
TestReader:
   inputs_def:
     fields: [‘image’, ‘im_size’, ‘im_id’, ‘gt_bbox’, ‘gt_class’, ‘is_difficult’]
     num_max_boxes: 50
   dataset:
     !VOCDataSet
     anno_path: valid.txt
     dataset_dir: /data/yjl/dapeng
     with_background: false
     use_default_label: false
   batch_size: 1
   drop_empty: false
   worker_num: 8
   bufsize: 4

Train

nohup python -m paddle.distributed.launch --selected_gpus 2,3 tools/train.py  -c ppyolo.yml --eval -o use_gpu=true --use_vdl=True --vdl_log_dir=vdl_dir/scalar -r output/ppyolo/90000 &    

export CUDA_VISIBLE_DEVICES=2  
python tools/train.py -c faster_rcnn_dcn_r50_fpn_1x.yml --eval -o use_gpu=true

Eval

 CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c ppyolo.yml -o use_gpu=true weights=output/ppyolo/90000

Infer

python tools/infer.py -c mask_rcnn_r50_fpn_1x.yml --infer_dir val --output_dir val_result -o weights=output/mask_rcnn_r50_fpn_1x/90000

Export

python tools/export_model.py -c faster_rcnn_dcn_r50_fpn_1x.yml \
        --output_dir=./inference_model \
        -o weights=output/faster_rcnn_dcn_r50_fpn_1x/model_final

python deploy/python/infer.py --model_dir=/path/to/models --image_file=/path/to/image
--use_gpu=(False/True) --output_dir=./re

自定义

CUDA_VISIBLE_DEVICES=2 python infer.py --model_dir=./inference_model/ppyolo --image_file=./e_ps --use_gpu=True --GPU_id=2  
注：CUDA_VISIBLE_DEVICES 与GPU_id 指定的设备一致 查看调用模型的显存消耗

PaddleClas

train

python -m paddle.distributed.launch --selected_gpus="0" tools/train.py \  
        -c ../ neu.yaml  \   #配置文件路径
-o pretrained_model=./ResNet50_vd_pretrained  #预训练模型存放路径
-o use_gpu=True

# 恢复训练  
python tools/train.py \
    -c configs/quick_start/MobileNetV3_large_x1_0_finetune.yaml \
    -o checkpoints="./output/MobileNetV3_large_x1_0/5/ppcls" \
    -o last_epoch=5 \
    -o use_gpu=True

eval

python -m paddle.distributed.launch --selected_gpus="0" tools/eval.py \  
    -c .. /eval.yaml \   #配置文件路径
    -o pretrained_model=output/ResNet50_vd/best_model/ppcls  #评估模型路径

export infer

python tools/export_model.py \  
    --model='ResNet50_vd' \   #模型名称
    --pretrained_model=output/ResNet50_vd/best_model/ppcls \  #需要转换的模型路径
    --output_path=./inference  #输出的预测模型保存路径  
   
 
python tools/infer/predict.py --use_gpu=1  -m cls_infer_model/model -p cls_infer_model/params -i  1.jpg --use_tensorrt=False