TensorFlow models/object_detection使用说明

1、下载安装

  1. cd xxx/models/research/
  2. python setup.py intsall
  3. cd xxx/models/research/slim/
  4. python setup.py install
  6. pip install --user package # 代表仅该用户的安装，安装后仅该用户可用
  7. pip install package        # 代表进行全局安装，安装后全局可用

修改xxx/models/research/object_detection/protos/路径下源文件时需要执行protoc才可产生效。

2、 配置object_detection训练pipeline

配置文件路径xxx/models/research/object_detection/samples/configs/

Step1: Preparing Inputs

Generating the dataset label map file.

如：xxx/models/research/object_detection/data/person_coco.pbtxt、xxx/models/research/object_detection/data/person_voc.pbtxt

Generating the AI TFRecord files.
  1. ## voc style
  2. python models/research/object_detection/dataset_tools/create_ai_tf_record_voc.py
  3.     --output_path /workspace/codes/models/research/object_detection/data/val_dataset_voc.record
  4.     --annotations_file /workspace/datasets/ai_challenger/ai_challenger_keypoint_validation_20170911/keypoint_validation_annotations_20170911.json
  5.     --image_dir /workspace/datasets/ai_challenger/ai_challenger_keypoint_validation_20170911/keypoint_validation_images_20170911/
  7. python models/research/object_detection/dataset_tools/create_ai_tf_record_voc.py    
  8.     --output_path /workspace/codes/models/research/object_detection/data/train_dataset_voc.record
  9.     --annotations_file /workspace/datasets/ai_challenger/ai_challenger_keypoint_train_20170909/keypoint_train_annotations_20170909.json
  10.     --image_dir  /workspace/datasets/ai_challenger/ai_challenger_keypoint_train_20170909/keypoint_train_images_20170902/    
  13. ## coco style
  14. python models/research/object_detection/dataset_tools/create_ai_tf_record_coco.py 
  15.        --output_path /workspace/codes/models/research/object_detection/data/val_dataset_coco.record
  16.        --annotations_file /workspace/datasets/ai_challenger/ai_challenger_keypoint_validation_20170911/keypoint_validation_annotations_20170911.json
  17.        --image_dir /workspace/datasets/ai_challenger/ai_challenger_keypoint_validation_20170911/keypoint_validation_images_20170911/
  18. python models/research/object_detection/dataset_tools/create_ai_tf_record_coco.py
  19.        --output_path /workspace/codes/models/research/object_detection/data/train_dataset_coco.record
  20.        --annotations_file /workspace/datasets/ai_challenger/ai_challenger_keypoint_train_20170909/keypoint_train_annotations_20170909.json
  21.        --image_dir  /workspace/datasets/ai_challenger/ai_challenger_keypoint_train_20170909/keypoint_train_images_20170902/

Step2: Runing Locally

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_locally.md

Step3: 训练与量化训练

可能仅需要更改配置文件。

Step4: Exporting a trained model for inference

Step5: Running on mobile with TensorFlow Lite（量化的

Step6: Pb to tflite

  1. ## env: 1.12.3
  2. # 命令行
  3.     Model: Frozen Graph(.pb)、SavedModel
  4.     Model Format: FLOAT、QUANTIZED_UINT8
  5.     Input Format: FLOAT、QUANTIZED_UINT8
  6.     post_training_quantize: 
  7. # api
  8. ## env: 1.15.2
  9. # 命令行
  10.     Model: Frozen Graph(.pb)、SavedModel
  11.     Model Format: FLOAT、QUANTIZED_UINT8
  12.     Input Format: FLOAT、QUANTIZED_UINT8
  13.     post_training_quantize: quantize_to_float16
  14. # api
  15. ## env: 2.1.0
  16. # 命令行
  17.     Model: SavedModel
  18. # api
  19.     Model: SavedModel
  20.     Model Format: int8、flot32、flot16

3、Inference

  1. ## infer代码
  2. python inference/infer_detections.py  
  3.     --input_tfrecord_paths=../object_detection/data/val_dataset_coco.record
  4.     --output_tfrecord_path=val_dataset_coco_detectionss.tfrecord
  5.     --inference_graph=../object_detection/test_ckpt/ssd_mv1/frozen_inference_graph.pb
  6. or
  8. python inference/infer_pb.py       # include pipeline

4、Eval

  1. ## ckpt模型eval流程(训练时候用的)
  2. python /legacy/eval.py 
  3.     --logtostderr 
  4.     --checkpoint_dir=../test_ckpt/ssd_mobilenet_v1_coco_2018_01_28/model.ckpt 
  5.     --eval_dir=./ 
  6.     --pipeline_config_path=../test_ckpt/ssd_mobilenet_v1_coco_2018_01_28/pipeline.config
  9. ## pb模型eval流程
  10. python inference/infer_detections.py 
  11.     --input_tfrecord_paths=../object_detection/data/val_dataset_coco.record
  12.     --output_tfrecord_path=val_dataset_coco_detectionss.tfrecord
  13.     --inference_graph=../object_detection/test_ckpt/ssd_mv1/frozen_inference_graph.pb
  14. ## 修改validation_eval_metrics/validation_eval_config.pbtxt文件  (eval格式选择：coco or voc）
  15. metrics_set: 'coco_detection_metrics'
  16. ## 修改validation_eval_metrics/validation_input_config.pbtxt文件  （标签文件：xx.pbtxt; 输出的tfrecord文件）
  17. label_map_path: '/workspace/codes/models/research/object_detection/data/person_coco.pbtxt'
  18. tf_record_input_reader: { input_path: '/workspace/codes/models/research/ai/val_dataset_coco_detectionss_3464.tfrecord' }
  19. SPLIT=validation
  20. PYTHONPATH=$PYTHONPATH:$(readlink -f ..) \
  21. python ../object_detection/metrics/offline_eval_map_corloc.py
  22.     --eval_dir=${SPLIT}_eval_metrics \
  23.     --eval_config_path=${SPLIT}_eval_metrics/${SPLIT}_eval_config.pbtxt \
  24.     --input_config_path=${SPLIT}_eval_metrics/${SPLIT}_input_config.pbtxt