# config yaml guide
KeyPoint config guide，Take an example of tinypose_256x192.yml
use_gpu: true #train with gpu or not
log_iter: 5 #print log every 5 iter
save_dir: output #the directory to save model
snapshot_epoch: 10 #save model every 10 epochs
weights: output/tinypose_256x192/model_final #the weight to load(without postfix “.pdparams”）
epoch: 420 #the total epoch number to train
num_joints: &num_joints 17 #number of joints
pixel_std: &pixel_std 200 #the standard pixel length（don't care）
metric: KeyPointTopDownCOCOEval #metric function
num_classes: 1 #number of classes（just for object detection, don't care）
train_height: &train_height 256 #the height of model input
train_width: &train_width 192 #the width of model input
trainsize: &trainsize [*train_width, *train_height] #the shape of model input
hmsize: &hmsize [48, 64] #the shape of model output
flip_perm: &flip_perm [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]] #the correspondence between left and right keypoint id, for example: left wrist become right wrist after image flip, and also the right wrist becomes left wrist
\#####model
architecture: TopDownHRNet #the model architecture
TopDownHRNet: #TopDownHRNet configs
backbone: LiteHRNet #which backbone to use
post_process: HRNetPostProcess #the post_process to use
flip_perm: *flip_perm #same to the upper "flip_perm"
num_joints: *num_joints #the joint number（the number of output channels）
width: &width 40 #backbone output channels
loss: KeyPointMSELoss #loss funciton
use_dark: true #whther to use DarkPose in postprocess
LiteHRNet: #LiteHRNet configs
network_type: wider_naive #the network type of backbone
freeze_at: -1 #the branch match this id doesn't backward，-1 means all branch backward
freeze_norm: false #whether to freeze normalize weights
return_idx: [0] #the branch id to fetch features
KeyPointMSELoss: #Loss configs
use_target_weight: true #whether to use target weights
loss_scale: 1.0 #loss weights，finalloss = loss*loss_scale
\#####optimizer
LearningRate: #LearningRate configs
base_lr: 0.002 #the original base learning rate
schedulers:
\- !PiecewiseDecay #the scheduler to adjust learning rate
milestones: [380, 410] #the milestones(epochs) to adjust learning rate
gamma: 0.1 #the ratio to adjust learning rate, new_lr = lr*gamma
\- !LinearWarmup #Warmup configs
start_factor: 0.001 #the original ratio with respect to base_lr
steps: 500 #iters used to warmup
OptimizerBuilder: #Optimizer type configs
optimizer:
type: Adam #optimizer type: Adam
regularizer:
factor: 0.0 #the regularizer weight
type: L2 #regularizer type: L2/L1
\#####data
TrainDataset: #Train Dataset configs
!KeypointTopDownCocoDataset #the dataset class to load data
image_dir: "" #the image directory, relative to dataset_dir
anno_path: aic_coco_train_cocoformat.json #the train datalist，coco format, relative to dataset_dir
dataset_dir: dataset #the dataset directory, the image_dir and anno_path based on this directory
num_joints: *num_joints #joint numbers
trainsize: *trainsize #the input size of model
pixel_std: *pixel_std #same to the upper "pixel_std"
use_gt_bbox: True #whether to use gt bbox, commonly used in eval
EvalDataset: #Eval Dataset configs
!KeypointTopDownCocoDataset #the dataset class to load data
image_dir: val2017 #the image directory, relative to dataset_dir
anno_path: annotations/person_keypoints_val2017.json #the eval datalist，coco format, relative to dataset_dir
dataset_dir: dataset/coco #the dataset directory, the image_dir and anno_path based on this directory
num_joints: *num_joints #joint numbers
trainsize: *trainsize #the input size of model
pixel_std: *pixel_std #same to the upper "pixel_std"
use_gt_bbox: True #whether to use gt bbox, commonly used in eval
image_thre: 0.5 #the threshold of detected rect, used while use_gt_bbox is False
TestDataset: #the test dataset without label
!ImageFolder #the class to load data, find images by folder
anno_path: dataset/coco/keypoint_imagelist.txt #the image list file
worker_num: 2 #the workers to load Dataset
global_mean: &global_mean [0.485, 0.456, 0.406] #means used to nomalize image
global_std: &global_std [0.229, 0.224, 0.225] #stds used to nomalize image
TrainReader: #TrainReader configs
sample_transforms: #transform configs
\- RandomFlipHalfBodyTransform: #random flip & random HalfBodyTransform
scale: 0.25 #the maximum scale for size transform
rot: 30 #the maximum rotation to transoform
num_joints_half_body: 8 #the HalfBodyTransform is skiped while joints found is less than this number
prob_half_body: 0.3 #the ratio of halfbody transform
pixel_std: *pixel_std #same to upper "pixel_std"
trainsize: *trainsize #the input size of model
upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] #the joint id which is belong to upper body
flip_pairs: *flip_perm #same to the upper "flip_perm"
\- AugmentationbyInformantionDropping:
prob_cutout: 0.5 #the probability to cutout keypoint
offset_factor: 0.05 #the jitter offset of cutout position, expressed as a percentage of trainwidth
num_patch: 1 #the numbers of area to cutout
trainsize: *trainsize #same to upper "trainsize"
\- TopDownAffine:
trainsize: *trainsize #same to upper "trainsize"
use_udp: true #whether to use udp_unbias（just for flip eval）
\- ToHeatmapsTopDown_DARK: #generate gt heatmaps
hmsize: *hmsize #the size of output heatmaps
sigma: 2 #the sigma of gaussin kernel which used to generate gt heatmaps
batch_transforms:
\- NormalizeImage: #image normalize class
mean: *global_mean #mean of normalize
std: *global_std #std of normalize
is_scale: true #whether scale by 1/255 to every image pixels，transform pixel from [0,255] to [0,1]
\- Permute: {} #channel transform from HWC to CHW
batch_size: 128 #batchsize used for train
shuffle: true #whether to shuffle the images before train
drop_last: false #whether drop the last images which is not enogh for batchsize
EvalReader:
sample_transforms: #transform configs
\- TopDownAffine: #Affine configs
trainsize: *trainsize #same to upper "trainsize"
use_udp: true #whether to use udp_unbias（just for flip eval）
batch_transforms:
\- NormalizeImage: #image normalize, the values should be same to values in TrainReader
mean: *global_mean
std: *global_std
is_scale: true
\- Permute: {} #channel transform from HWC to CHW
batch_size: 16 #batchsize used for test
TestReader:
inputs_def:
image_shape: [3, *train_height, *train_width] #the input dimensions used in model，CHW
sample_transforms:
\- Decode: {} #load image
\- TopDownEvalAffine: #Affine class used in Eval
trainsize: *trainsize #the input size of model
\- NormalizeImage: #image normalize, the values should be same to values in TrainReader
mean: *global_mean #mean of normalize
std: *global_std #std of normalize
is_scale: true #whether scale by 1/255 to every image pixels，transform pixel from [0,255] to [0,1]
\- Permute: {} #channel transform from HWC to CHW
batch_size: 1 #Test batchsize
fuse_normalize: false #whether fuse the normalize into model while export model, this speedup the model infer