backbone:
  NStageBackbone:

    stem:
      YoloNASStem:
        out_channels: 48

    stages:
      - YoloNASStage:
          out_channels: 96
          num_blocks: 2
          activation_type: relu
          hidden_channels: 96
          concat_intermediates: True

      - YoloNASStage:
          out_channels: 192
          num_blocks: 3
          activation_type: relu
          hidden_channels: 128
          concat_intermediates: True

      - YoloNASStage:
          out_channels: 384
          num_blocks: 5
          activation_type: relu
          hidden_channels: 256
          concat_intermediates: True

      - YoloNASStage:
          out_channels: 768
          num_blocks: 2
          activation_type: relu
          hidden_channels: 512
          concat_intermediates: True


    context_module:
      SPP:
        output_channels: 768
        activation_type: relu
        k: [5,9,13]

    out_layers: [stage1, stage2, stage3, context_module]

neck:
  YoloNASPANNeckWithC2:

    neck1:
      YoloNASUpStage:
        out_channels: 192
        num_blocks: 4
        hidden_channels: 128
        width_mult: 1
        depth_mult: 1
        activation_type: relu
        reduce_channels: True

    neck2:
      YoloNASUpStage:
        out_channels: 96
        num_blocks: 4
        hidden_channels: 128
        width_mult: 1
        depth_mult: 1
        activation_type: relu
        reduce_channels: True

    neck3:
      YoloNASDownStage:
        out_channels: 192
        num_blocks: 4
        hidden_channels: 128
        activation_type: relu
        width_mult: 1
        depth_mult: 1

    neck4:
      YoloNASDownStage:
        out_channels: 384
        num_blocks: 4
        hidden_channels: 256
        activation_type: relu
        width_mult: 1
        depth_mult: 1

heads:
  NDFLHeads:
    num_classes: 80
    reg_max: 16
    heads_list:
      - YoloNASDFLHead:
          inter_channels: 128
          width_mult: 1
          first_conv_group_size: 0
          stride: 8
      - YoloNASDFLHead:
          inter_channels: 256
          width_mult: 1
          first_conv_group_size: 0
          stride: 16
      - YoloNASDFLHead:
          inter_channels: 512
          width_mult: 1
          first_conv_group_size: 0
          stride: 32

bn_eps: 1e-3
bn_momentum: 0.03
inplace_act: True

_convert_: all

backbone:
  NStageBackbone:

    stem:
      YoloNASStem:
        out_channels: 48

    stages:
      - YoloNASStage:
          out_channels: 96
          num_blocks: 2
          activation_type: relu
          hidden_channels: 64
          concat_intermediates: True

      - YoloNASStage:
          out_channels: 192
          num_blocks: 3
          activation_type: relu
          hidden_channels: 128
          concat_intermediates: True

      - YoloNASStage:
          out_channels: 384
          num_blocks: 5
          activation_type: relu
          hidden_channels: 256
          concat_intermediates: True

      - YoloNASStage:
          out_channels: 768
          num_blocks: 2
          activation_type: relu
          hidden_channels: 384
          concat_intermediates: False


    context_module:
      SPP:
        output_channels: 768
        activation_type: relu
        k: [5,9,13]

    out_layers: [stage1, stage2, stage3, context_module]

neck:
  YoloNASPANNeckWithC2:

    neck1:
      YoloNASUpStage:
        out_channels: 192
        num_blocks: 2
        hidden_channels: 192
        width_mult: 1
        depth_mult: 1
        activation_type: relu
        reduce_channels: True

    neck2:
      YoloNASUpStage:
        out_channels: 96
        num_blocks: 3
        hidden_channels: 64
        width_mult: 1
        depth_mult: 1
        activation_type: relu
        reduce_channels: True

    neck3:
      YoloNASDownStage:
        out_channels: 192
        num_blocks: 2
        hidden_channels: 192
        activation_type: relu
        width_mult: 1
        depth_mult: 1

    neck4:
      YoloNASDownStage:
        out_channels: 384
        num_blocks: 3
        hidden_channels: 256
        activation_type: relu
        width_mult: 1
        depth_mult: 1

heads:
  NDFLHeads:
    num_classes: 80
    reg_max: 16
    heads_list:
      - YoloNASDFLHead:
          inter_channels: 128
          width_mult: 0.75
          first_conv_group_size: 0
          stride: 8
      - YoloNASDFLHead:
          inter_channels: 256
          width_mult: 0.75
          first_conv_group_size: 0
          stride: 16
      - YoloNASDFLHead:
          inter_channels: 512
          width_mult: 0.75
          first_conv_group_size: 0
          stride: 32

bn_eps: 1e-3
bn_momentum: 0.03
inplace_act: True

_convert_: all

backbone:
  NStageBackbone:

    stem:
      YoloNASStem:
        out_channels: 48

    stages:
      - YoloNASStage:
          out_channels: 96
          num_blocks: 2
          activation_type: relu
          hidden_channels: 32
          concat_intermediates: False

      - YoloNASStage:
          out_channels: 192
          num_blocks: 3
          activation_type: relu
          hidden_channels: 64
          concat_intermediates: False

      - YoloNASStage:
          out_channels: 384
          num_blocks: 5
          activation_type: relu
          hidden_channels: 96
          concat_intermediates: False

      - YoloNASStage:
          out_channels: 768
          num_blocks: 2
          activation_type: relu
          hidden_channels: 192
          concat_intermediates: False


    context_module:
      SPP:
        output_channels: 768
        activation_type: relu
        k: [5,9,13]

    out_layers: [stage1, stage2, stage3, context_module]

neck:
  YoloNASPANNeckWithC2:

    neck1:
      YoloNASUpStage:
        out_channels: 192
        num_blocks: 2
        hidden_channels: 64
        width_mult: 1
        depth_mult: 1
        activation_type: relu
        reduce_channels: True

    neck2:
      YoloNASUpStage:
        out_channels: 96
        num_blocks: 2
        hidden_channels: 48
        width_mult: 1
        depth_mult: 1
        activation_type: relu
        reduce_channels: True

    neck3:
      YoloNASDownStage:
        out_channels: 192
        num_blocks: 2
        hidden_channels: 64
        activation_type: relu
        width_mult: 1
        depth_mult: 1

    neck4:
      YoloNASDownStage:
        out_channels: 384
        num_blocks: 2
        hidden_channels: 64
        activation_type: relu
        width_mult: 1
        depth_mult: 1

heads:
  NDFLHeads:
    num_classes: 80
    reg_max: 16
    heads_list:
      - YoloNASDFLHead:
          inter_channels: 128
          width_mult: 0.5
          first_conv_group_size: 0
          stride: 8
      - YoloNASDFLHead:
          inter_channels: 256
          width_mult: 0.5
          first_conv_group_size: 0
          stride: 16
      - YoloNASDFLHead:
          inter_channels: 512
          width_mult: 0.5
          first_conv_group_size: 0
          stride: 32

bn_eps: 1e-3
bn_momentum: 0.03
inplace_act: True

_convert_: all

# YoloNAS-S Detection training on COCO2017 Dataset:
# This training recipe is for demonstration purposes only. Pretrained models were trained using a different recipe.
# So it will not be possible to reproduce the results of the pretrained models using this recipe.

# 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. Run the command you want:
#         yolo_nas_s: python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=coco2017_yolo_nas_s
#

defaults:
  - training_hyperparams: coco2017_yolo_nas_train_params
  - dataset_params: coco_detection_yolo_nas_dataset_params
  - arch_params: yolo_nas_s_arch_params
  - checkpoint_params: default_checkpoint_params
  - _self_
  - variable_setup

train_dataloader: coco2017_train_yolo_nas
val_dataloader: coco2017_val_yolo_nas

load_checkpoint: False
resume: False

dataset_params:
  train_dataloader_params:
    batch_size: 32

arch_params:
  num_classes: 80

training_hyperparams:
  resume: ${resume}
  mixed_precision: True

architecture: yolo_nas_s

multi_gpu: DDP
num_gpus: 8

experiment_suffix: ""
experiment_name: coco2017_${architecture}${experiment_suffix}

@@ -30,15 +30,15 @@ train_dataset_params:
         mixup_scale: [ 0.5, 1.5 ]         # random rescale range for the additional sample in mixup
         prob: 0.5                       # probability to apply per-sample mixup
         flip_prob: 0.5                  # probability to apply horizontal flip
-    - DetectionStandardizeImage:
-        max_value: 255.
     - DetectionPaddedRescale:
         input_dim: [640, 640]
         max_targets: 120
         pad_value: 114
+    - DetectionStandardize:
+        max_value: 255.
     - DetectionTargetsFormatTransform:
         max_targets: 256
-        output_format: LABEL_NORMALIZED_CXCYWH
+        output_format: LABEL_CXCYWH
 
   tight_box_rotation: False
   class_inclusion_list:
@@ -67,13 +67,13 @@ val_dataset_params:
     - DetectionPadToSize:
         output_size: [640, 640]
         pad_value: 114
-    - DetectionStandardizeImage:
+    - DetectionStandardize:
         max_value: 255.
     - DetectionImagePermute
     - DetectionTargetsFormatTransform:
         max_targets: 50
         input_dim: [640, 640]
-        output_format: LABEL_NORMALIZED_CXCYWH
+        output_format: LABEL_CXCYWH
   tight_box_rotation: False
   class_inclusion_list:
   max_num_samples:
@@ -83,6 +83,7 @@ val_dataloader_params:
   batch_size: 25
   num_workers: 8
   drop_last: False
+  shuffle: False
   pin_memory: True
   collate_fn:
     _target_: super_gradients.training.utils.detection_utils.CrowdDetectionCollateFN

@@ -9,18 +9,27 @@ train_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  ignore_empty_annotations: False
   transforms:
+    - DetectionMosaic:
+        input_dim: ${dataset_params.train_dataset_params.input_dim}
+        prob: 1.
     - DetectionRandomAffine:
         degrees: 0.                  # rotation degrees, randomly sampled from [-degrees, degrees]
         translate: 0.1                # image translation fraction
         scales: [ 0.5, 1.5 ]              # random rescale range (keeps size by padding/cropping) after mosaic transform.
         shear: 0.0                    # shear degrees, randomly sampled from [-degrees, degrees]
         target_size: ${dataset_params.train_dataset_params.input_dim}
-        filter_box_candidates: True   # whether to filter out transformed bboxes by edge size, area ratio, and aspect ratio.
+        filter_box_candidates: False  # whether to filter out transformed bboxes by edge size, area ratio, and aspect ratio.
         wh_thr: 2                     # edge size threshold when filter_box_candidates = True (pixels)
         area_thr: 0.1                 # threshold for area ratio between original image and the transformed one, when filter_box_candidates = True
         ar_thr: 20                    # aspect ratio threshold when filter_box_candidates = True
         border_value: 128
+#    - DetectionMixup:
+#        input_dim: ${dataset_params.train_dataset_params.input_dim}
+#        mixup_scale: [ 0.5, 1.5 ]         # random rescale range for the additional sample in mixup
+#        prob: 1.0                       # probability to apply per-sample mixup
+#        flip_prob: 0.5                  # probability to apply horizontal flip
     - DetectionHSV:
         prob: 1.0                       # probability to apply HSV transform
         hgain: 5                        # HSV transform hue gain (randomly sampled from [-hgain, hgain])
@@ -30,8 +39,11 @@ train_dataset_params:
         prob: 0.5                       # probability to apply horizontal flip
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 120
+        max_targets: 300
+    - DetectionStandardize:
+        max_value: 255.
     - DetectionTargetsFormatTransform:
+        max_targets: 300
         input_dim: ${dataset_params.train_dataset_params.input_dim}
         output_format: LABEL_CXCYWH
   tight_box_rotation: False
@@ -43,8 +55,8 @@ train_dataset_params:
 train_dataloader_params:
   shuffle: True
   batch_size: 16
-  num_workers: 0
-  sampler:
+  min_samples: 512
+  num_workers: 4
   drop_last: False
   pin_memory: True
   worker_init_fn:
@@ -60,11 +72,16 @@ val_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  ignore_empty_annotations: False
   transforms:
   - DetectionPaddedRescale:
       input_dim: ${dataset_params.val_dataset_params.input_dim}
+      max_targets: 300
+      pad_value: 114
+  - DetectionStandardize:
+      max_value: 255.
   - DetectionTargetsFormatTransform:
-      max_targets: 50
+      max_targets: 300
       input_dim: ${dataset_params.val_dataset_params.input_dim}
       output_format: LABEL_CXCYWH
   tight_box_rotation: False
@@ -74,10 +91,10 @@ val_dataset_params:
   verbose: 0
 
 val_dataloader_params:
-  batch_size: 64
-  num_workers: 0
-  sampler:
+  batch_size: 32
+  num_workers: 4
   drop_last: False
+  shuffle: False
   pin_memory: True
   collate_fn: # collate function for valset
     _target_: super_gradients.training.utils.detection_utils.CrowdDetectionCollateFN

# A recipe to fine-tune YoloNAS on Roboflow datasets.
# Checkout the datasets at https://universe.roboflow.com/roboflow-100?ref=blog.roboflow.com
#
# `dataset_name` refers to the official name of the dataset.
# You can find it in the url of the dataset: https://universe.roboflow.com/roboflow-100/digits-t2eg6 -> digits-t2eg6
#
# Example: python -m super_gradients.train_from_recipe --config-name=roboflow_yolo_nas_m dataset_name=digits-t2eg6

defaults:
  - training_hyperparams: coco2017_yolo_nas_train_params
  - dataset_params: roboflow_detection_dataset_params
  - checkpoint_params: default_checkpoint_params
  - arch_params: yolo_nas_m_arch_params
  - _self_
  - variable_setup

train_dataloader: roboflow_train_yolox
val_dataloader: roboflow_val_yolox

dataset_name: ??? # Placeholder for the name of the dataset you want to use (e.g. "digits-t2eg6")
dataset_params:
  dataset_name: ${dataset_name}

  train_dataloader_params:
    batch_size: 12

  val_dataloader_params:
    batch_size: 16

num_classes: ${roboflow_dataset_num_classes:${dataset_name}}

architecture: yolo_nas_m
arch_params:
  num_classes: ${num_classes}


load_checkpoint: False
checkpoint_params:
  pretrained_weights: coco


result_path: # By defaults saves results in checkpoints directory
resume: False
training_hyperparams:
  resume: ${resume}
  zero_weight_decay_on_bias_and_bn: True

  lr_warmup_epochs: 3
  warmup_mode: linear_epoch_step

  initial_lr: 4e-4
  cosine_final_lr_ratio: 0.1

  optimizer_params:
    weight_decay: 0.0001

  ema: True
  ema_params:
    decay: 0.9

  max_epochs: 100
  mixed_precision: True
  criterion_params:
    num_classes: ${num_classes}


  phase_callbacks: []
  loss:
    ppyoloe_loss:
      num_classes: ${num_classes}
      reg_max: 16

  valid_metrics_list:
    - DetectionMetrics_050:
        score_thres: 0.1
        top_k_predictions: 300
        num_cls: ${num_classes}
        normalize_targets: True
        post_prediction_callback:
          _target_: super_gradients.training.models.detection_models.pp_yolo_e.PPYoloEPostPredictionCallback
          score_threshold: 0.01
          nms_top_k: 1000
          max_predictions: 300
          nms_threshold: 0.7

  metric_to_watch: 'mAP@0.50'

multi_gpu: Off
num_gpus: 1

experiment_suffix: ""
experiment_name: ${architecture}_roboflow_${dataset_name}${experiment_suffix}

# A recipe to fine-tune YoloNAS on Roboflow datasets.
# Checkout the datasets at https://universe.roboflow.com/roboflow-100?ref=blog.roboflow.com
#
# `dataset_name` refers to the official name of the dataset.
# You can find it in the url of the dataset: https://universe.roboflow.com/roboflow-100/digits-t2eg6 -> digits-t2eg6
#
# Example: python -m super_gradients.train_from_recipe --config-name=roboflow_yolo_nas_s dataset_name=digits-t2eg6

defaults:
  - training_hyperparams: coco2017_yolo_nas_train_params
  - dataset_params: roboflow_detection_dataset_params
  - checkpoint_params: default_checkpoint_params
  - arch_params: yolo_nas_s_arch_params
  - _self_
  - variable_setup

train_dataloader: roboflow_train_yolox
val_dataloader: roboflow_val_yolox

dataset_name: ??? # Placeholder for the name of the dataset you want to use (e.g. "digits-t2eg6")
dataset_params:
  dataset_name: ${dataset_name}

  train_dataloader_params:
    batch_size: 16

  val_dataloader_params:
    batch_size: 16

num_classes: ${roboflow_dataset_num_classes:${dataset_name}}

architecture: yolo_nas_s
arch_params:
  num_classes: ${num_classes}


load_checkpoint: False
checkpoint_params:
  pretrained_weights: coco


result_path: # By defaults saves results in checkpoints directory
resume: False
training_hyperparams:
  resume: ${resume}
  zero_weight_decay_on_bias_and_bn: True

  lr_warmup_epochs: 3
  warmup_mode: linear_epoch_step

  initial_lr: 5e-4
  cosine_final_lr_ratio: 0.1

  optimizer_params:
    weight_decay: 0.0001

  ema: True
  ema_params:
    decay: 0.9

  max_epochs: 100
  mixed_precision: True
  criterion_params:
    num_classes: ${num_classes}


  phase_callbacks: []
  loss:
    ppyoloe_loss:
      num_classes: ${num_classes}
      reg_max: 16

  valid_metrics_list:
    - DetectionMetrics_050:
        score_thres: 0.1
        top_k_predictions: 300
        num_cls: ${num_classes}
        normalize_targets: True
        post_prediction_callback:
          _target_: super_gradients.training.models.detection_models.pp_yolo_e.PPYoloEPostPredictionCallback
          score_threshold: 0.01
          nms_top_k: 1000
          max_predictions: 300
          nms_threshold: 0.7

  metric_to_watch: 'mAP@0.50'

multi_gpu: Off
num_gpus: 1

experiment_suffix: ""
experiment_name: ${architecture}_roboflow_${dataset_name}${experiment_suffix}

defaults:
  - roboflow_yolo_nas_s
  - quantization_params: default_quantization_params
  - _self_

checkpoint_params:
  checkpoint_path: ???
  strict_load: no_key_matching

pre_launch_callbacks_list:
    - QATRecipeModificationCallback:
        batch_size_divisor: 2
        max_epochs_divisor: 10
        lr_decay_factor: 0.01
        warmup_epochs_divisor: 10
        cosine_final_lr_ratio: 0.01
        disable_phase_callbacks: True
        disable_augmentations: False

defaults:
  - default_train_params

max_epochs: 300

warmup_mode: "linear_batch_step"
warmup_initial_lr:  1e-6
lr_warmup_steps: 1000
lr_warmup_epochs: 0

initial_lr:  2e-4
lr_mode: cosine
cosine_final_lr_ratio: 0.1

zero_weight_decay_on_bias_and_bn: True
batch_accumulate: 1

save_ckpt_epoch_list: [100, 200, 250]

loss:
  ppyoloe_loss:
    use_static_assigner: False
    num_classes: ${arch_params.num_classes}
    reg_max: 16

optimizer: AdamW
optimizer_params:
  weight_decay: 0.00001

ema: True
ema_params:
  decay: 0.9997
  decay_type: threshold

mixed_precision: False
sync_bn: True

valid_metrics_list:
  - DetectionMetrics:
      score_thres: 0.1
      top_k_predictions: 300
      num_cls: ${arch_params.num_classes}
      normalize_targets: True
      post_prediction_callback:
        _target_: super_gradients.training.models.detection_models.pp_yolo_e.PPYoloEPostPredictionCallback
        score_threshold: 0.01
        nms_top_k: 1000
        max_predictions: 300
        nms_threshold: 0.7

pre_prediction_callback:

metric_to_watch: 'mAP@0.50:0.95'
greater_metric_to_watch_is_better: True

_convert_: all