ppyolo/yoloV3模型r34量化导出模型无法正常预测

song121982 commented 2 years ago

使用的功能：

使用ppyolo/yoloV3中r34模型搭配slim量化训练

环境：

paddlepaddle2.2 paddledetection2.1 paddleslim2.1

问题：

训练过程没有问题，边训练边评估mAP可以达到80+ 模型导出前，使用训练得到的动态模型，指令tools/infer.py xxxxxx可以得到正确的预测结果。

但模型导出后使用导出模型预测：python deploy/python/infer.py --model_dir=inference_model/ppyolov2_r34vd_dcn_vehicletrain --image_file=demo/03237.jpg --use_gpu=True

预测没有任何效果，没有任何预测框生成

后又尝试在导出模型过程中不配置slim文件，即不指定 --slim_config，导出的部署模型才有预测框，虽然是乱的。

这个问题困扰好多天了，一直没人解答，麻烦解答一下

song121982 commented 2 years ago

附配置文件：

# runtime
use_gpu: true
log_iter: 10
save_dir: output
snapshot_epoch: 2

# set collate_batch to false because ground-truth info is needed
# on voc dataset and should not collate data in batch when batch size
# is larger than 1.

# datasets
metric: VOC
map_type: 11point
num_classes: 1

TrainDataset:
  !VOCDataSet
    dataset_dir: dataset/VSvehicle_vehicle6000
    anno_path: train.txt
    label_list: label_list.txt
    data_fields: ['image', 'gt_bbox', 'gt_class', 'difficult']

EvalDataset:
  !VOCDataSet
    dataset_dir: dataset/VSvehicle_vehicle6000
    anno_path: valid.txt
    label_list: label_list.txt
    data_fields: ['image', 'gt_bbox', 'gt_class', 'difficult']

TestDataset:
  !ImageFolder
    anno_path: dataset/VSvehicle_vehicle6000/label_list.txt

# optimizer
epoch: 50

LearningRate:
  base_lr: 0.0001
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones:
    - 30
    - 40
  - !LinearWarmup
    start_factor: 0.
    steps: 2000

OptimizerBuilder:
  clip_grad_by_norm: 35.
  optimizer:
    momentum: 0.9
    type: Momentum
  regularizer:
    factor: 0.0005
    type: L2

# ppyolov2_r50vd_dcn
architecture: YOLOv3
#pretrain_weights: output/r34_pre_vehicle/best_model.pdparams
norm_type: sync_bn
use_ema: true
ema_decay: 0.9998

YOLOv3:
  backbone: ResNet
  neck: PPYOLOFPN
  yolo_head: YOLOv3Head
  post_process: BBoxPostProcess

ResNet:
  depth: 34
  variant: d
  return_idx: [1, 2, 3]
#  dcn_v2_stages: [3]
  freeze_at: -1
  freeze_norm: false
  norm_decay: 0.

PPYOLOFPN:
  coord_conv: true
  drop_block: false
  block_size: 3
  keep_prob: 0.9
  spp: true

YOLOv3Head:
  anchors: [[29, 17], [16, 34], [39, 24],
            [20, 47], [34, 40], [47, 50],
            [87, 36], [31, 104], [86, 92]]
  anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
  loss: YOLOv3Loss
  iou_aware: true
  iou_aware_factor: 0.4

YOLOv3Loss:
  ignore_thresh: 0.6
  downsample: [32, 16, 8]
  label_smooth: false
  scale_x_y: 1.05
  iou_loss: IouLoss
  iou_aware_loss: IouAwareLoss

IouLoss:
  loss_weight: 2.5
  loss_square: true

IouAwareLoss:
  loss_weight: 1.0

BBoxPostProcess:
  decode:
    name: YOLOBox
    conf_thresh: 0.01
    downsample_ratio: 32
    clip_bbox: true
    scale_x_y: 1.05
  nms:
    name: MatrixNMS
    keep_top_k: 100
    score_threshold: 0.01
    post_threshold: 0.01
    nms_top_k: -1
    background_label: -1

# reader
worker_num: 2
TrainReader:
  inputs_def:
    num_max_boxes: 50
  sample_transforms:
    - Decode: {}
    - Mixup: {alpha: 1.5, beta: 1.5}
    - RandomDistort: {}
    - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
    - RandomCrop: {}
    - RandomFlip: {}
  batch_transforms:
    - BatchRandomResize: {target_size: [320, 384, 416, 448, 480, 512, 544, 608], random_size: True, random_interp: True, keep_ratio: False}
    - NormalizeBox: {}
    - PadBox: {num_max_boxes: 50}
    - BboxXYXY2XYWH: {}
    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
    - Permute: {}
    - Gt2YoloTarget: {anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]],
                      anchors: [[29, 17], [16, 34], [39, 24],[20, 47], [34, 40], [47, 50],[87, 36], [31, 104], [86, 92]],
                      downsample_ratios: [32, 16, 8]}
  batch_size: 6
  shuffle: true
  drop_last: true
  mixup_epoch: 25000
  use_shared_memory: true

EvalReader:
  sample_transforms:
    - Decode: {}
    - Resize: {target_size: [512, 512], keep_ratio: False, interp: 2}
    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
    - Permute: {}
  batch_size: 1
  collate_batch: false

TestReader:
  inputs_def:
    image_shape: [3, 512, 512]
  sample_transforms:
    - Decode: {}
    - Resize: {target_size: [512, 512], keep_ratio: False, interp: 2}
    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
    - Permute: {}
  batch_size: 1

slim配置文件：

pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet34_pretrained.pdparams

slim: QAT

QAT:
  quant_config: {
    'activation_preprocess_type': 'PACT',
    'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max',
    'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.99,
    'quantizable_layer_type': ['Conv2D', 'Linear']}
  print_model: False

heavengate commented 2 years ago

训练过程没有问题，边训练边评估mAP可以达到80+ 模型导出前，使用训练得到的动态模型，指令tools/infer.py xxxxxx可以得到正确的预测结果。

但模型导出后使用导出模型预测：python deploy/python/infer.py --model_dir=inference_model/ppyolov2_r34vd_dcn_vehicletrain --image_file=demo/03237.jpg --use_gpu=True

你tools/infer.py预测有结果的命令和导出的全量命令麻烦发一下还有你说导出了有框但是是乱的的预测结果麻烦也发一下

song121982 commented 2 years ago

训练过程没有问题，边训练边评估mAP可以达到80+ 模型导出前，使用训练得到的动态模型，指令tools/infer.py xxxxxx可以得到正确的预测结果。但模型导出后使用导出模型预测：python deploy/python/infer.py --model_dir=inference_model/ppyolov2_r34vd_dcn_vehicletrain --image_file=demo/03237.jpg --use_gpu=True

你tools/infer.py预测有结果的命令和导出的全量命令麻烦发一下还有你说导出了有框但是是乱的的预测结果麻烦也发一下

导出模型前预测：

python tools/infer.py -c testtest/r34_qat/ppyolov2_r34vd_vehicletrain.yml --slim_config testtest/r34_qat/ppyolov2_r34vd_dcn_vehicletrain.yml --infer_img=demo/Visdrone_Vehicle00416.jpg -o weights=testtest/r34_qat/best_model.pdparams 输出结果正常（这里标签值忘改了）:

导出指令： python tools/export_model.py -c testtest/r34_qat/ppyolov2_r34vd_vehicletrain.yml --slim_config testtest/r34_qat/ppyolov2_r34vd_dcn_vehicletrain.yml -o weights=output/best_model.pdparams

使用导出模型预测：

python deploy/python/infer.py --model_dir=testtest/ppyolov2_r34vd_dcn_vehicletrain --image_file=./demo/Visdrone_vehicle_val01348.jpg print输出：

(paddle) E:\paddle\PaddleDetection-release-2.1>python deploy/python/infer.py --model_dir=testtest/ppyolov2_r
34vd_dcn_vehicletrain --image_file=./demo/Visdrone_vehicle_val01348.jpg
-----------  Running Arguments -----------
batch_size: 1
camera_id: -1
cpu_threads: 1
enable_mkldnn: False
image_dir: None
image_file: ./demo/Visdrone_vehicle_val01348.jpg
model_dir: testtest/ppyolov2_r34vd_dcn_vehicletrain
output_dir: output
run_benchmark: False
run_mode: fluid
threshold: 0.5
trt_calib_mode: False
trt_max_shape: 1280
trt_min_shape: 1
trt_opt_shape: 640
use_dynamic_shape: False
use_gpu: False
video_file: None
------------------------------------------
-----------  Model Configuration -----------
Model Arch: YOLO
Transform Order:
--transform op: Resize
--transform op: NormalizeImage
--transform op: Permute
--------------------------------------------
e[37m---    Fused 0 subgraphs into layer_norm op.e[0m

------------------ Inference Time Info ----------------------
total_time(ms): 1349.1, img_num: 1
average latency time(ms): 1349.10, QPS: 0.741235
preprocess_time(ms): 22.40, inference_time(ms): 1326.70, postprocess_time(ms): 0.00

结果没有任何预测框生成：

导出模型过程中不指定slim_config,然后使用导出模型预测： python tools/export_model.py -c testtest/r34_qat/ppyolov2_r34vd_vehicletrain.yml -o weights=testtest/r34_qat/best_model.pdparams

python deploy/python/infer.py --model_dir=output_inference/ppyolov2_r34vd_vehicletrain --image_file=./demo/Visdrone_vehicle_val01348.jpg 输出结果：