config_file_detail.md

NanoDet Config File Analysis

NanoDet using yacs to read yaml config file.

Saving path

save_dir: PATH_TO_SAVE

Change save_dir to where you want to save logs and models. If path not exist, NanoDet will create it.

Model

model:
    arch:
        name: OneStageDetector
        backbone: xxx
        fpn: xxx
        head: xxx

Most detection model architecture can be devided into 3 parts: backbone, task head and connector between them(e.g. FPN, BiFPN, PAN...).

Backbone

backbone:
    name: ShuffleNetV2
    model_size: 1.0x
    out_stages: [2,3,4]
    activation: LeakyReLU
    with_last_conv: False

NanoDet using ShuffleNetV2 as backbone. You can modify model size, output feature levels and activation function. Moreover, NanoDet provides other lightweight backbones like GhostNet and MobileNetV2. You can also add your backbone network by importing it in nanodet/model/backbone/__init__.py.

FPN

fpn:
    name: PAN
    in_channels: [116, 232, 464]
    out_channels: 96
    start_level: 0
    num_outs: 3

NanoDet using modified PAN (replace downsample convs with interpolation to reduce amount of computations).

in_channels : a list of feature map channels extracted from backbone.

out_channels : out put feature map channel.

Head

head:
    name: NanoDetHead
    num_classes: 80
    input_channel: 96
    feat_channels: 96
    stacked_convs: 2
    share_cls_reg: True
    octave_base_scale: 8
    scales_per_octave: 1
    strides: [8, 16, 32]
    reg_max: 7
    ignore_iof_thr: -1
    norm_cfg:
      type: BN
    loss:

name: Task head class name

num_classes: number of classes

input_channel: input feature map channel

feat_channels: channel of task head convs

stacked_convs: how many conv blocks use in one task head

share_cls_reg: use same conv blocks for classification and box regression

octave_base_scale: base box scale

scales_per_octave: anchor free model only have one base box, default value 1

strides: down sample stride of each feature map level

reg_max: max value of per-level l-r-t-b distance

ignore_iof_thr: thresh of iof for ignore box, default value -1

norm_cfg: normalization layer setting

loss: adjust loss functions and weights

Weight averaging

Nanodet supports weight averaging method like EMA:

model:
  weight_averager:
    name: ExpMovingAverager
    decay: 0.9998
  arch:
    ...

Data

data:
    train:
        name: CocoDataset
        img_path: coco/train2017
        ann_path: coco/annotations/instances_train2017.json
        input_size: [320,320]
        keep_ratio: True
        multi_scale: [0.6, 1.4]
        pipeline:
    val:
    .....

In data you need to set your train and validate dataset.

name: Dataset format name. You can create your own dataset format in nanodet/data/dataset. input_size: [width, height] keep_ratio: whether to maintain the original image ratio when resizing to input size multi_scale: Scaling range for multi-scale training. Set to None to turn off. pipeline: data preprocessing and augmentation pipeline

Device

device:
    gpu_ids: [0]
    workers_per_gpu: 12
    batchsize_per_gpu: 160
    precision: 32

gpu_ids: CUDA device id. For multi-gpu training, set [0, 1, 2...].

workers_per_gpu: how many dataloader processes for each gpu

batchsize_per_gpu: amount of images in one batch for each gpu

precision: Training precision. The default value 32 means FP32 training. Set to 16 to enable AMP training.

schedule

schedule:
#  resume:
#  load_model: YOUR_MODEL_PATH
  optimizer:
    name: SGD
    lr: 0.14
    momentum: 0.9
    weight_decay: 0.0001
  warmup:
    name: linear
    steps: 300
    ratio: 0.1
  total_epochs: 70
  lr_schedule:
    name: MultiStepLR
    milestones: [40,55,60,65]
    gamma: 0.1
  val_intervals: 10

Set training schedule.

resume: whether to restore last training process

load_model: path to trained weight

optimizer: Support all optimizer provided by pytorch.

You should adjust the lr with batch_size. Following linear scaling rule in paper Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

warmup: Warm up your network before training. Support constant, exp and linear three types of warm up.

total_epochs: total epochs to train

lr_schedule: please refer to pytorch lr_scheduler documentation

val_intervals: epoch interval of evaluating during training

Evaluate

evaluator:
  name: CocoDetectionEvaluator
  save_key: mAP

Currently only support coco eval.

save_key: Metric of best model. Support mAP, AP50, AP75....

---

class_names: used in visualization