# Waymo Open Dataset 開發紀錄 contributed by < `gyes00205` > ###### tags: `waymo` ## Download Dataset [Waymo Open Dataset](https://console.cloud.google.com/storage/browser/waymo_open_dataset_v_1_2_0) domain_adaptation 的資料沒有 label ，所以請下載 training/ 的資料去訓練 ## tfrecord 內容 * 五種 camera 所拍攝的照片以及 Lidar 資訊 * num_classes: 0: Unknown, 1: Vehicle, 2: Pedestrian, 3: Sign, 4: Cyclist 這次 project 不需要 sign 和 Unknown 這兩個 classes ，因此 label_map.pbtxt 修改如下: ```pbtxt item { id: 1 name: 'vehicle' } item { id: 2 name: 'pedestrian' } item { id: 4 name: 'cyclist' } ``` * camera 種類: FRONT, FRONT_LEFT, FRONT_RIGHT, SIDE_LEFT, SIDE_RIGHT <img src="https://i.imgur.com/Q68Lepf.jpg"> * bbox (x, y, w, h) 座標: , xy 代表 bbox 中心座標 , wh 代表寬和高 ## 環境配置 ### 安裝 Waymo open dataset ```shell pip3 install waymo-open-dataset-tf-2-1-0==1.2.0 ``` ### 安裝 COCO API ```shell pip install cython pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI ``` ### 安裝 Tensorflow 2 Object Detection API 參考 [TensorFlow 2 Object Detection API tutorial](https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/index.html) 安裝套件 * git clone Tensorflow 2 Object Detection API ```shell git clone https://github.com/tensorflow/models.git ``` * 到 models/research/ 執行 ```shell protoc object_detection/protos/*.proto --python_out=. ``` * 將 API 加到環境變數 ``` export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim ``` * 複製 setup.py 到 models/research/ ```shell cp object_detection/packages/tf2/setup.py ./ ``` * 安裝 setup.py ```shell python -m pip install . ``` * 測試是否安裝成功 ``` python object_detection/builders/model_builder_tf2_test.py ``` ### 建立資料夾結構 ``` Waymo ├───models #Tensorflow Object Detection API ├───training_configs #訓練用的 config ├───pre-trained-models #預訓練模型 ├───exported-models #輸出模型 └───data #訓練資料 └───segment-???.tfrecord ``` ## 轉換 tfrecord 格式 因為 waymo 的 tfrecord 除了有 Lidar 的資訊之外，他的 bbox 格式如下: (x0, y0): 為中心點座標。 (w, h): 為長寬。 ![](https://i.imgur.com/WSDKAQZ.png) 而我們的目標是過濾掉 Lidar 並將 bbox 轉為以下格式: (x1, y1): 為左上角座標。 (x2, y2): 為右下角座標。 ![](https://i.imgur.com/HyR6xS0.png) 轉換 tfrecord 的程式碼參考 [LevinJ/tf_obj_detection_api](https://github.com/LevinJ/tf_obj_detection_api)，並且做一些小修改。 **create_record.py:** filepath: tfrecord 的路徑 data_dir: 轉換過後的 tfrecord 會儲存在 data_dir/processed 目錄下 執行方式如下: ```shell python create_record.py \ --filepath=data/segment-???.tfrecord \ --data_dir=data/ ``` 執行完後 data/processed 便會出現處理完的 tfrecord。 ``` Waymo ├───models ├───training_configs ├───pre-trained-models ├───exported-models └───data ├───processed │ └───segment-???.tfrecord #處理後的 tfrecord └───segment-???.tfrecord ``` ## 下載預訓練模型 到 [Tensorflow Model Zoo](https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md) 下載 pretrained model。 ![](https://i.imgur.com/x34zpZL.png) 我下載的是 `SSD ResNet50 V1 FPN 640x640 (RetinaNet50)`。 * 先到 pre-trained-models 目錄下 `cd pre-trained-models` * 下載 SSD ResNet50 的 pretrained model ``` wget http://download.tensorflow.org/models/object_detection/tf2/20200711/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8.tar.gz ``` * 解壓縮 `tar zxvf ssd_resnet50_v1_fpn_640x640_coco17_tpu-8.tar.gz` ``` Waymo ├───models ├───training_configs ├───pre-trained-models │ └───ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 │ ├─ checkpoint/ │ ├─ saved_model/ │ └─ pipeline.config ├───exported-models └───data ├───processed │ └───segment-???.tfrecord #處理後的 tfrecord └───segment-???.tfrecord ``` ## 修改訓練用 config 到 [configs/tf2](https://github.com/tensorflow/models/tree/master/research/object_detection/configs/tf2) 找到與 pretrained model 相對應的 config，也就是ssd_resnet50_v1_fpn_640x640_coco17_tpu-8.config * 在 training_configs 新增資料夾 ```shell cd training_configs mkdir ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 ``` * 在 ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 目錄下新增 pipeline.config，並將剛剛找到的 config 內容複製貼上，並且做一些修改。 * num_classes: 種類個數 * batch_size: bach size 大小，根據電腦的記憶體而有不同設置 * fine_tune_checkpoint: 更改成 pretrained model 的 ckpt-0 路徑 * num_steps: 訓練步數 * use_bfloat16: 是否使用 tpu，沒有使用設定為 false * label_map_path: label_map.pbtxt 路徑 * train_input_reader: 將 input_path 設定成訓練用的 tfrecord 路徑 * metrics_set: "coco_detection_metrics" * use_moving_averages: false * eval_input_reader: 將 input_path 設定成評估用用的 tfrecord 路徑 ```config # SSD with Resnet 50 v1 FPN feature extractor, shared box predictor and focal # loss (a.k.a Retinanet). # See Lin et al, https://arxiv.org/abs/1708.02002 # Trained on COCO, initialized from Imagenet classification checkpoint # Train on TPU-8 # # Achieves 34.3 mAP on COCO17 Val model { ssd { inplace_batchnorm_update: true freeze_batchnorm: false num_classes: 3 #因為種類有 3 個 box_coder { faster_rcnn_box_coder { y_scale: 10.0 x_scale: 10.0 height_scale: 5.0 width_scale: 5.0 } } matcher { argmax_matcher { matched_threshold: 0.5 unmatched_threshold: 0.5 ignore_thresholds: false negatives_lower_than_unmatched: true force_match_for_each_row: true use_matmul_gather: true } } similarity_calculator { iou_similarity { } } encode_background_as_zeros: true anchor_generator { multiscale_anchor_generator { min_level: 3 max_level: 7 anchor_scale: 4.0 aspect_ratios: [1.0, 2.0, 0.5] scales_per_octave: 2 } } image_resizer { fixed_shape_resizer { height: 640 width: 640 } } box_predictor { weight_shared_convolutional_box_predictor { depth: 256 class_prediction_bias_init: -4.6 conv_hyperparams { activation: RELU_6, regularizer { l2_regularizer { weight: 0.0004 } } initializer { random_normal_initializer { stddev: 0.01 mean: 0.0 } } batch_norm { scale: true, decay: 0.997, epsilon: 0.001, } } num_layers_before_predictor: 4 kernel_size: 3 } } feature_extractor { type: 'ssd_resnet50_v1_fpn_keras' fpn { min_level: 3 max_level: 7 } min_depth: 16 depth_multiplier: 1.0 conv_hyperparams { activation: RELU_6, regularizer { l2_regularizer { weight: 0.0004 } } initializer { truncated_normal_initializer { stddev: 0.03 mean: 0.0 } } batch_norm { scale: true, decay: 0.997, epsilon: 0.001, } } override_base_feature_extractor_hyperparams: true } loss { classification_loss { weighted_sigmoid_focal { alpha: 0.25 gamma: 2.0 } } localization_loss { weighted_smooth_l1 { } } classification_weight: 1.0 localization_weight: 1.0 } normalize_loss_by_num_matches: true normalize_loc_loss_by_codesize: true post_processing { batch_non_max_suppression { score_threshold: 1e-8 iou_threshold: 0.6 max_detections_per_class: 100 max_total_detections: 100 } score_converter: SIGMOID } } } train_config: { fine_tune_checkpoint_version: V2 #pretrained model 的 ckpt-0 位置 fine_tune_checkpoint: "pre-trained-models/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8/checkpoint/ckpt-0" fine_tune_checkpoint_type: "detection" #改為 detection batch_size: 2 sync_replicas: true startup_delay_steps: 0 replicas_to_aggregate: 8 use_bfloat16: false #因為沒有使用 tpu 所以改為 false num_steps: 6000 #訓練步數 data_augmentation_options { random_horizontal_flip { } } data_augmentation_options { random_crop_image { min_object_covered: 0.0 min_aspect_ratio: 0.75 max_aspect_ratio: 3.0 min_area: 0.75 max_area: 1.0 overlap_thresh: 0.0 } } optimizer { momentum_optimizer: { learning_rate: { cosine_decay_learning_rate { learning_rate_base: .04 total_steps: 25000 warmup_learning_rate: .013333 warmup_steps: 2000 } } momentum_optimizer_value: 0.9 } use_moving_average: false } max_number_of_boxes: 100 unpad_groundtruth_tensors: false } train_input_reader: { label_map_path: "./label_map.pbtxt" tf_record_input_reader { input_path: "data/processed/*.tfrecord" } } eval_config: { metrics_set: "coco_detection_metrics" use_moving_averages: false } eval_input_reader: { label_map_path: "./label_map.pbtxt" shuffle: false num_epochs: 1 tf_record_input_reader { input_path: "data/processed/*.tfrecord" } } ``` ``` Waymo ├───models ├───training_configs │ └───ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 │ └───pipeline.config #新增 pipeline.config ├───pre-trained-models │ └───ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 │ ├─ checkpoint/ │ ├─ saved_model/ │ └─ pipeline.config ├───exported-models └───data ├───processed │ └───segment-???.tfrecord └───segment-???.tfrecord ``` ## 訓練模型 **model_main_tf2.py** model_dir: 會將訓練的 checkpoint 儲存在 model_dir 目錄下 pipeline_config_path: pipeline.config 路徑 執行方式如下: ```shell python model_main_tf2.py \ --model_dir=training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 \ --pipeline_config_path=training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8/pipeline.config ``` 執行結果如下: 每 100 steps 會印一次。 ``` Step 2100 per-step time 0.320s INFO:tensorflow:{'Loss/classification_loss': 0.121629156, 'Loss/localization_loss': 0.16370133, 'Loss/regularization_loss': 0.2080817, 'Loss/total_loss': 0.4934122, 'learning_rate': 0.039998136} I0605 08:29:04.605577 139701982308224 model_lib_v2.py:700] {'Loss/classification_loss': 0.121629156, 'Loss/localization_loss': 0.16370133, 'Loss/regularization_loss': 0.2080817, 'Loss/total_loss': 0.4934122, 'learning_rate': 0.039998136} ``` ## 評估模型 (Optional) **model_main_tf2.py** checkpoint_dir: 讀取 checkpoint 的目錄。 執行方式如下: ```shell python model_main_tf2.py \ --model_dir=training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 \ --pipeline_config_path=training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8/pipeline.config \ --checkpoint_dir=training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8/ ``` 執行結果: 會計算 AP 和 AR $AP^{small}:$ AP for small object : area < $32^2$ $AP^{medium}:$ AP for medium object : $32^2$ < area < $96^2$ $AP^{large}:$ AP for large object : $96^2$ < area ![](https://i.imgur.com/RjN2dRf.png) ## 輸出模型 **exporter_main_v2.py** input_type: image_tensor pipeline_config_path: pipeline.config 的路徑 trained_checkpoint_dir: 儲存 checkpoint 的位置 output_directory: 輸出模型位置 執行方式如下: ```shell !python exporter_main_v2.py \ --input_type image_tensor \ --pipeline_config_path training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8/pipeline.config \ --trained_checkpoint_dir training_configs/ssd_resnet50_v1_fpn_640x640_coco17_tpu-8/ \ --output_directory exported-models/my_model_6000steps ``` 執行結果如下: ```shell INFO:tensorflow:Assets written to: exported-models/my_model_6000steps/saved_model/assets I0605 09:07:21.034602 139745385867136 builder_impl.py:775] Assets written to: exported-models/my_model_6000steps/saved_model/assets INFO:tensorflow:Writing pipeline config file to exported-models/my_model_6000steps/pipeline.config I0605 09:07:22.310333 139745385867136 config_util.py:254] Writing pipeline config file to exported-models/my_model_6000steps/pipeline.config ``` ``` Waymo ├───models ├───training_configs │ └───ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 │ └─pipeline.config #新增 pipeline.config ├───pre-trained-models │ └───ssd_resnet50_v1_fpn_640x640_coco17_tpu-8 │ ├─ checkpoint/ │ ├─ saved_model/ │ └─ pipeline.config ├───exported-models │ └───my_model_6000steps └───data ├───processed │ └─segment-???.tfrecord └───segment-???.tfrecord ``` ## 使用模型預測圖片 **detect.py** saved_model_path: 模型位置 test_path: 測試圖片位置 output_path: 輸出預測圖片位置 min_score_thresh: 信心水準 執行方式如下: ```shell !python detect.py \ --saved_model_path=exported-models/my_model_6000steps \ --test_path=test_image \ --output_path=output_image \ --min_score_thresh=.1 ``` 預測結果: <img src="https://i.imgur.com/NNE6OuI.png" width=250px height=200px> <img src="https://i.imgur.com/dyRuUpA.png" width=300px height=200px> <img src="https://i.imgur.com/vICSrnI.png" width=250px height=200px> <img src="https://i.imgur.com/it53kPf.png" width=300px height=200px> ## Reference 1. [LevinJ/tf_obj_detection_api](https://github.com/LevinJ/tf_obj_detection_api) 2. [Waymo Open Dataset](https://waymo.com/open/) 3. [Waymo quick start tutorial](https://colab.research.google.com/github/waymo-research/waymo-open-dataset/blob/r1.0/tutorial/tutorial.ipynb) 4. [Tensorflow Object Detection API Tutorial](https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html)