coco2017_pose_dekr_w32_no_dc.yaml

#  DEKR training example with COCO dataset.
#  Reproduction and refinement of paper: Bottom-Up Human Pose Estimation Via Disentangled Keypoint Regression.
#
#  Note: Original DEKR architecture using deformable convolutions. This recipe uses standard convolutions to enable
#  model be exportable to ONNX.
#
#  Recipe runs with batch size = 24 X 8 gpus = 192.
#
#  Instructions:
#   0. Make sure that the data is stored in dataset_params.dataset_dir or add "dataset_params.data_dir=<PATH-TO-DATASET>" at the end of the command below (feel free to check ReadMe)
#   1. Move to the project root (where you will find the ReadMe and src folder)
#   2. Make sure you've downloaded pretrained backbone weights from https://1drv.ms/u/s!Aus8VCZ_C_33dYBMemi9xOUFR0w to project root (See line 55).
#   3. Run the command:
#       DEKR-W32-NO-DC: python -m super_gradients.train_from_recipe --config-name=coco2017_pose_dekr_w32_no_dc checkpoint_params.checkpoint_path=hrnetv2_w32_imagenet_pretrained.pth
#
#
#  Validation AP (Without flip augmentation and rescoring) - COCO, training time:
#      DEKR-W32-NO-DC:    input-size: [640, 640]     AP: 63.08 (Regular training)   8 X RTX A5000 - 21h
#
#  Scores with flip TTA and rescoring (Using best model from above):
#      DEKR-W32-NO-DC:    input-size: [640, 640]     AP: 64.96 (With Flip TTA)
#      DEKR-W32-NO-DC:    input-size: [640, 640]     AP: 67.34 (With Flip TTA and Rescoring)
#
#  Rescoring:
#      See `coco2017_pose_dekr_rescoring.yaml` recipe and `documentation/source/PoseEstimation.md#Rescoring` section of the documentation.
#
#  Official git repo:
#      https://github.com/HRNet/DEKR
#  Paper:
#      https://arxiv.org/abs/2104.02300
#
#
#  Comments:
#      * Pretrained backbones were used.
#      * In DEKR-W32-NO-DC A suffix "NO-DC" stands for "No deformable convolutions".

defaults:
  - training_hyperparams: coco2017_dekr_pose_train_params
  - dataset_params: coco_pose_estimation_dekr_dataset_params
  - arch_params: pose_dekr_w32_no_dc_arch_params
  - checkpoint_params: default_checkpoint_params
  - _self_
  - variable_setup

architecture: dekr_w32_no_dc

multi_gpu: DDP
num_gpus: 8

experiment_suffix: ""
experiment_name: coco2017_pose_${architecture}${experiment_suffix}

ckpt_root_dir:

train_dataloader: coco2017_pose_train
val_dataloader: coco2017_pose_val

arch_params:
  num_classes: ${dataset_params.num_joints}

checkpoint_params:
  # Original training recipe uses pretrained weights for HRNet on ImageNet.
  # You will need to download the pretrained weights from the original repo and place
  # them in `external_checkpoint_path` param.
  # Download weights from this url https://1drv.ms/u/s!Aus8VCZ_C_33dYBMemi9xOUFR0w
  checkpoint_path: # <location of the downloaded hrnetv2_w32_imagenet_pretrained.pth>
  strict_load:
    _target_: super_gradients.training.sg_trainer.StrictLoad
    value: key_matching

dataset_params:
  train_dataloader_params:
    batch_size: 24

  val_dataloader_params:
    batch_size: 32