|
- # Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
- enable_modelarts: True
- # Url for modelarts
- data_url: ""
- train_url: ""
- checkpoint_url: ""
-
- # Path of local
- data_path: "/cache/data/"
- output_path: "/cache/train"
- load_path: "/cache/ckpt"
-
- # hyperparameters of training
- momentum: 0.9
- decay: 0.0001
- lr: 0.01
- epochs: 40
- batch_size: 4
-
- # other
- num_classes: 1
- k_max: 400
- img_size: [1088, 608]
- track_buffer: 30
- keep_checkpoint_max: 6
-
- # model initialization parameters
- backbone_input_shape: [32, 64, 128, 256, 512]
- backbone_shape: [64, 128, 256, 512, 1024]
- backbone_layers: [1, 2, 8, 8, 4]
- out_channel: 24 # 3 * (num_classes + 5)
- embedding_dim: 512
-
- # evaluation thresholds
- iou_thres: 0.50
- conf_thres: 0.50
- nms_thres: 0.40
- min_box_area: 200
-
- # h -> w
- anchor_scales: [
- [8, 24],
- [11, 34],
- [16, 48],
- [23, 68],
- [32, 96],
- [45, 135],
- [64, 192],
- [90, 271],
- [128, 384],
- [180, 540],
- [256, 640],
- [512, 640],
- ]
-
-
- # data configs
- col_names_train: [
- 'imgs',
- 'tconf_s',
- 'tbox_s',
- 'tid_s',
- 'tconf_m',
- 'tbox_m',
- 'tid_m',
- 'tconf_b',
- 'tbox_b',
- 'tid_b',
- 'emb_indices_s',
- 'emb_indices_m',
- 'emb_indices_b',
- ]
-
- col_names_val: [
- 'imgs',
- 'targets',
- 'lens',
- ]
-
- # profiling
- enable_profiling: false
- train_profiling_dir: "save_train_profiling"
-
- # other
- is_distributed: True
- dataset_root: "/cache/data/MOT"
- device_target: 'Ascend'
- device_id: 0
- device_start: 0
- ckpt_url: "./darknet53_pretrained_imagenet.ckpt"
- logs_dir: "/cache/train/"
- input_video: '/path/to/input/video'
- output_format: 'video'
- output_root: './results'
- save_images: False
- save_videos: False
- file_format: "MINDIR" # ["AIR", "ONNX", "MINDIR"]
- ---
- # Config description for each option
- momentum: 'Momentum for SGD optimizer.'
- decay: 'Weight_decay for SGD optimizer.'
- lr: 'Init learning rate.'
- epochs: 'Number of epochs to train.'
- batch_size: 'Batch size per one device'
- num_classes: 'Number of object classes.'
- k_max: 'Max predictions per one map (made for optimization of FC layer embedding computation).'
- img_size: 'Size of input images.'
- track_buffer: 'Tracking buffer.'
- keep_checkpoint_max: 'Keep saved last N checkpoints.'
- backbone_input_shape: 'Input filters of backbone layers.'
- backbone_shape: 'Input filters of backbone layers.'
- backbone_layers: 'Output filters of backbone layers.'
- out_channel: 'Number of channels for detection.'
- embedding_dim: 'Number of channels for embeddings.'
- iou_thres: 'IOU thresholds.'
- conf_thres: 'Confidence threshold.'
- nms_thres: 'Threshold for Non-max suppression.'
- min_box_area: 'Filter out tiny boxes.'
- anchor_scales: '12 predefined anchor boxes. Different 4 per each of 3 feature maps.'
- col_names_train: 'Names of columns for training GeneratorDataset.'
- col_names_val: 'Names of columns for validation GeneratorDataset.'
- is_distributed: 'Distribute training or not.'
- dataset_root: 'Path to datasets root folder.'
- device_target: 'Device GPU or any.'
- device_id: 'Device id of target device.'
- device_start: 'Start device id.'
- ckpt_url: 'Location of checkpoint.'
- logs_dir: 'Dir to save logs and ckpt.'
- input_video: 'Path to the input video.'
- output_format: 'Expected output format.'
- output_root: 'Expected output root path.'
- save_images: 'Save tracking results (image).'
- save_videos: 'Save tracking results (video).'
|