model builds

official/vision/beta/projects/yolo/configs/experiments/yolov3/tpu/512-letter.yaml

-runtime:
-  distribution_strategy: 'tpu'
-  mixed_precision_dtype: 'bfloat16'
-task:
-  smart_bias_lr: 0.0
-  model:
-    darknet_based_model: True
-    input_size: [512, 512, 3]
-    backbone:
-      type: 'darknet'
-      darknet:
-        model_id: 'darknet53'
-        max_level: 5
-        min_level: 3
-    decoder:
-      type: yolo_decoder
-      yolo_decoder:
-        version: v3
-        type: regular
-    head:
-      smart_bias: true
-    detection_generator:
-      box_type:
-        'all': original
-      scale_xy:
-        '5': 1.05
-        '4': 1.1
-        '3': 1.2
-      max_boxes: 200
-      nms_type: greedy
-      iou_thresh: 0.001
-      nms_thresh: 0.60
-    loss:
-      use_scaled_loss: False
-      box_loss_type:  
-        'all': ciou
-      ignore_thresh:
-        'all': 0.7
-      iou_normalizer: 
-        'all': 0.07
-      cls_normalizer: 
-        'all': 1.0
-      obj_normalizer: 
-        'all': 1.0
-      objectness_smooth: 
-        'all': 0.0
-      max_delta:
-        'all': 5.0
-    norm_activation:
-      activation: mish
-      norm_epsilon: 0.0001
-      norm_momentum: 0.99
-      use_sync_bn: true
-    num_classes: 80
-    anchor_boxes:
-      anchors_per_scale: 3
-      boxes: [box: [12, 16],   box: [19, 36],   box: [40, 28], 
-              box: [36, 75],   box: [76, 55],   box: [72, 146], 
-              box: [142, 110], box: [192, 243], box: [459, 401]]
-  train_data:
-    global_batch_size: 64
-    dtype: float32
-    input_path: 'gs://cam2-datasets/coco/train*'
-    is_training: true
-    drop_remainder: true
-    seed: 1000
-    parser:
-      mosaic: 
-        mosaic_frequency: 1.0
-        mixup_frequency: 0.0
-        mosaic_crop_mode: 'scale'
-        mosaic_center: 0.25
-        aug_scale_min: 0.1
-        aug_scale_max: 1.9
-      max_num_instances: 200
-      letter_box: True
-      random_flip: True
-      aug_rand_saturation: 0.7
-      aug_rand_brightness: 0.4
-      aug_rand_hue: 0.015
-      aug_rand_translate: 0.1
-      area_thresh: 0.1
-      random_pad: False
-      use_tie_breaker: True
-      anchor_thresh: 0.213
-  validation_data:
-    global_batch_size: 8
-    dtype: float32
-    input_path: 'gs://cam2-datasets/coco/val*'
-    is_training: false
-    drop_remainder: true
-    parser:
-      max_num_instances: 300
-      letter_box: True
-      use_tie_breaker: True
-      anchor_thresh: 0.213
-  weight_decay: 0.000
-  init_checkpoint: 'gs://tensorflow2/darknet/cspdarknet53-golden'
-  init_checkpoint_modules: 'backbone'
-  annotation_file: null
-trainer:
-  train_steps: 500500 # 160 epochs at 64 batchsize -> 500500 * 64/2
-  validation_steps: 625
-  steps_per_loop: 1850
-  summary_interval: 1850
-  validation_interval: 9250 
-  checkpoint_interval: 1850
-  optimizer_config:
-    ema:
-      average_decay: 0.9998
-      trainable_weights_only: False
-      dynamic_decay: True
-    learning_rate:
-      type: stepwise
-      stepwise:
-        boundaries: [400000, 450000]
-        name: PiecewiseConstantDecay
-        values: [0.00131, 0.000131, 0.0000131] 
-    optimizer:
-      type: sgd_torch
-      sgd_torch:
-        momentum: 0.949
-        momentum_start: 0.949
-        nesterov: True
-        warmup_steps: 1000
-        weight_decay: 0.0005
-        sim_torch: true
-        name: SGD
-    warmup:
-      type: 'linear'
-      linear:
-        warmup_steps: 1000 #learning rate rises from 0 to 0.0013 over 1000 steps

official/vision/beta/projects/yolo/configs/experiments/yolov4-csp/tpu/640-ign.yaml

-runtime:
-  distribution_strategy: 'tpu'
-  mixed_precision_dtype: 'float32'
-  tpu_enable_xla_dynamic_padder: false
-task:
-  smart_bias_lr: 0.1
-  model:
-    darknet_based_model: False
-    input_size: [640, 640, 3]
-    backbone:
-      type: 'darknet'
-      darknet:
-        model_id: 'altered_cspdarknet53'
-        max_level: 5
-        min_level: 3
-    decoder:
-      type: yolo_decoder
-      yolo_decoder:
-        version: v4
-        type: csp
-    head:
-      smart_bias: true
-    detection_generator:
-      box_type:
-        'all': scaled
-      scale_xy:
-        'all': 2.0
-      max_boxes: 300
-      nms_type: greedy
-      iou_thresh: 0.001
-      nms_thresh: 0.60
-    loss:
-      use_scaled_loss: true
-      update_on_repeat: true
-      box_loss_type:  
-        'all': ciou
-      ignore_thresh:
-        'all': 0.7
-      iou_normalizer: 
-        'all': 0.05
-      cls_normalizer: 
-        'all': 0.3
-      obj_normalizer: 
-        '5': 0.28
-        '4': 0.70
-        '3': 2.80
-      objectness_smooth: 
-        'all': 1.0
-    norm_activation:
-      activation: mish
-      norm_epsilon: 0.0001
-      norm_momentum: 0.97
-      use_sync_bn: true
-    num_classes: 80
-    anchor_boxes:
-      anchors_per_scale: 3
-      boxes: [box: [12, 16],   box: [19, 36],   box: [40, 28], 
-              box: [36, 75],   box: [76, 55],   box: [72, 146], 
-              box: [142, 110], box: [192, 243], box: [459, 401]]
-  train_data:
-    global_batch_size: 64
-    dtype: float32
-    input_path: 'gs://cam2-datasets/coco/train*'
-    is_training: true
-    shuffle_buffer_size: 10000
-    drop_remainder: true
-    seed: 1000
-    parser:
-      mosaic: 
-        mosaic_frequency: 1.0
-        mixup_frequency: 0.0
-        mosaic_crop_mode: 'scale'
-        mosaic_center: 0.25
-        aug_scale_min: 0.1
-        aug_scale_max: 1.9
-      max_num_instances: 300
-      letter_box: True
-      random_flip: True
-      aug_rand_saturation: 0.7
-      aug_rand_brightness: 0.4
-      aug_rand_hue: 0.015
-      aug_rand_translate: 0.1
-      area_thresh: 0.1
-      random_pad: False
-      use_tie_breaker: True
-      anchor_thresh: 4.0
-      best_match_only: True
-  validation_data:
-    global_batch_size: 8
-    dtype: float32
-    input_path: 'gs://cam2-datasets/coco/val*'
-    is_training: false
-    shuffle_buffer_size: 10
-    drop_remainder: true
-    parser:
-      max_num_instances: 300
-      letter_box: True
-      use_tie_breaker: True
-      anchor_thresh: 4.0
-      best_match_only: True
-  weight_decay: 0.000
-  annotation_file: null
-trainer:
-  train_steps: 555000 # 160 epochs at 64 batchsize -> 500500 * 64/2
-  validation_steps: 625
-  steps_per_loop: 1850
-  summary_interval: 1850
-  validation_interval: 1850
-  checkpoint_interval: 1850
-  optimizer_config:
-    ema:
-      average_decay: 0.9999
-      trainable_weights_only: False
-      dynamic_decay: True
-    learning_rate:
-      type: cosine
-      cosine:
-        initial_learning_rate: 0.01
-        name: Cosine
-        alpha: 0.2
-        decay_steps: 555000
-    optimizer:
-      type: sgd_torch
-      sgd_torch:
-        momentum: 0.937
-        momentum_start: 0.8
-        nesterov: True
-        warmup_steps: 5550
-        weight_decay: 0.0005
-        sim_torch: true
-        name: SGD
-    warmup:
-      type: 'linear'
-      linear:
-        warmup_steps: 5550 #learning rate rises from 0 to 0.0013 over 1000 steps

official/vision/beta/projects/yolo/configs/experiments/yolov4/inference/512-swin.yaml

-runtime:
-  distribution_strategy: 'mirrored'
-  mixed_precision_dtype: 'float16'
-  num_gpus: 1
-task:
-  smart_bias_lr: 0.0
-  model:
-    darknet_based_model: True
-    input_size: [512, 512, 3]
-    backbone:
-      type: 'swin'
-      swin:
-        min_level: 3
-        max_level: 5
-        patch_size: 4
-        embed_dims: 96 
-        window_size: [7, 7, 7, 7] 
-        depths: [2, 2, 6, 2]
-        num_heads: [3, 6, 12, 24]
-        drop_path: 0.0
-        absolute_positional_embed: False
-    decoder:
-      type: yolo_decoder
-      yolo_decoder:
-        version: v4
-        type: csp
-        activation: leaky
-    head:
-      smart_bias: true
-    detection_generator:
-      box_type:
-        'all': original
-      scale_xy:
-        '5': 1.05
-        '4': 1.1
-        '3': 1.2
-      max_boxes: 200
-      nms_type: greedy
-      iou_thresh: 0.25
-      nms_thresh: 0.45
-      pre_nms_points: 500
-    loss:
-      use_scaled_loss: False
-      box_loss_type:  
-        'all': ciou
-      ignore_thresh:
-        'all': 0.7
-      iou_normalizer: 
-        'all': 0.07
-      cls_normalizer: 
-        'all': 1.0
-      obj_normalizer: 
-        'all': 1.0
-      objectness_smooth: 
-        'all': 0.0
-      max_delta:
-        'all': 5.0
-    norm_activation:
-      activation: gelu
-      norm_epsilon: 0.0001
-      norm_momentum: 0.99
-      use_sync_bn: false
-    num_classes: 80
-    anchor_boxes:
-      anchors_per_scale: 3
-      boxes: [box: [12, 16],   box: [19, 36],   box: [40, 28], 
-              box: [36, 75],   box: [76, 55],   box: [72, 146], 
-              box: [142, 110], box: [192, 243], box: [459, 401]]
-  train_data:
-    global_batch_size: 4
-    dtype: float16
-    input_path: '/media/vbanna/DATA_SHARE/CV/datasets/COCO_raw/records/train*'
-    is_training: true
-    drop_remainder: true
-    seed: 1000
-    parser:
-      mosaic:
-        mosaic_frequency: 0.6
-        mixup_frequency: 0.0
-        mosaic_crop_mode: 'crop'
-        mosaic_center: 0.2
-        aug_scale_min: 0.2
-        aug_scale_max: 1.6
-        jitter: 0.3
-      max_num_instances: 200
-      letter_box: True
-      random_flip: True
-      aug_rand_saturation: 1.5
-      aug_rand_brightness: 1.5
-      aug_rand_hue: 0.1
-      aug_scale_min: 1.0
-      aug_scale_max: 1.0
-      aug_rand_translate: 0.0 
-      jitter: 0.3
-      area_thresh: 0.1
-      random_pad: True
-      use_tie_breaker: True
-      anchor_thresh: 0.213
-  validation_data:
-    global_batch_size: 8
-    dtype: float16
-    input_path: '/media/vbanna/DATA_SHARE/CV/datasets/COCO_raw/records/val*'
-    is_training: false
-    drop_remainder: true
-    parser:
-      max_num_instances: 200
-      letter_box: True
-      use_tie_breaker: True
-      anchor_thresh: 0.213
-  weight_decay: 0.000
-  init_checkpoint: '../checkpoints/swin-baseline-3'
-  init_checkpoint_modules: 'backbone'
-  annotation_file: null
-trainer:
-  train_steps: 500500 # 160 epochs at 64 batchsize -> 500500 * 64/2
-  validation_steps: 625
-  steps_per_loop: 10
-  summary_interval: 10
-  validation_interval: 9250 
-  checkpoint_interval: 1850
-  optimizer_config:
-    ema: null
-    learning_rate:
-      type: stepwise
-      stepwise:
-        boundaries: [400000, 450000]
-        name: PiecewiseConstantDecay
-        values: [0.00131, 0.000131, 0.0000131] 
-    optimizer:
-      type: sgd_torch
-      sgd_torch:
-        momentum: 0.949
-        momentum_start: 0.949
-        nesterov: True
-        warmup_steps: 1000
-        weight_decay: 0.0005
-        sim_torch: true
-        name: SGD
-    warmup:
-      type: 'linear'
-      linear:
-        warmup_steps: 1000 #learning rate rises from 0 to 0.0013 over 1000 steps

official/vision/beta/projects/yolo/configs/experiments/yolov4/tpu/512-mb.yaml

-runtime:
-  distribution_strategy: 'tpu'
-  mixed_precision_dtype: 'bfloat16'
-task:
-  smart_bias_lr: 0.0
-  model:
-    darknet_based_model: True
-    input_size: [512, 512, 3]
-    backbone:
-      type: 'darknet'
-      darknet:
-        model_id: 'cspdarknet53'
-        max_level: 5
-        min_level: 3
-    decoder:
-      type: yolo_decoder
-      yolo_decoder:
-        version: v4
-        type: regular
-        activation: leaky
-    head:
-      smart_bias: true
-    detection_generator:
-      box_type:
-        'all': original
-      scale_xy:
-        '5': 1.05
-        '4': 1.1
-        '3': 1.2
-      max_boxes: 200
-      nms_type: iou
-      iou_thresh: 0.001
-      nms_thresh: 0.60
-    loss:
-      use_scaled_loss: False
-      box_loss_type:  
-        'all': ciou
-      ignore_thresh:
-        'all': 0.7
-      iou_normalizer: 
-        'all': 0.07
-      cls_normalizer: 
-        'all': 1.0
-      obj_normalizer: 
-        'all': 1.0
-      objectness_smooth: 
-        'all': 0.0
-      max_delta:
-        'all': 5.0
-    norm_activation:
-      activation: mish
-      norm_epsilon: 0.0001
-      norm_momentum: 0.99
-      use_sync_bn: true
-    num_classes: 80
-    anchor_boxes:
-      anchors_per_scale: 3
-      boxes: [box: [12, 16],   box: [19, 36],   box: [40, 28], 
-              box: [36, 75],   box: [76, 55],   box: [72, 146], 
-              box: [142, 110], box: [192, 243], box: [459, 401]]
-  train_data:
-    input_path: 'gs://cam2-datasets/coco/train*'
-    parser:
-      mosaic: 
-        mosaic_frequency: 1.0
-        mixup_frequency: 0.0
-        mosaic_crop_mode: 'scale'
-        mosaic_center: 0.25
-        aug_scale_min: 0.1
-        aug_scale_max: 1.9
-        jitter: 0.3
-      max_num_instances: 200
-      letter_box: False
-      random_flip: True
-      aug_rand_translate: 0.1
-      random_pad: False
-  validation_data:
-    input_path: 'gs://cam2-datasets/coco/val*'
-    parser:
-      letter_box: False
-  weight_decay: 0.000
-  init_checkpoint: 'gs://tensorflow2/darknet/cspdarknet53-golden'
-  init_checkpoint_modules: 'backbone'
-  annotation_file: null
-# trainer:
-#   train_steps: 500500 # 160 epochs at 64 batchsize -> 500500 * 64/2
-#   validation_steps: 625
-#   steps_per_loop: 1850
-#   summary_interval: 1850
-#   validation_interval: 9250 
-#   checkpoint_interval: 1850
-#   optimizer_config:
-#     ema:
-#       average_decay: 0.9998
-#       trainable_weights_only: False
-#       dynamic_decay: True
-#     learning_rate:
-#       type: stepwise
-#       stepwise:
-#         boundaries: [400000, 450000]
-#         values: [0.00131, 0.000131, 0.0000131] 
-#     optimizer:
-#       type: sgd_torch
-#       sgd_torch:
-#         momentum: 0.949
-#         momentum_start: 0.949
-#         nesterov: True
-#         warmup_steps: 1000
-#         weight_decay: 0.0005
-#         sim_torch: true
-#         name: SGD
-#     warmup:
-#       type: 'linear'
-#       linear:
-#         warmup_steps: 1000 #learning rate rises from 0 to 0.0013 over 1000 steps

official/vision/beta/projects/yolo/configs/experiments/yolov4/tpu/512-swin.yaml

-runtime:
-  distribution_strategy: 'tpu'
-  mixed_precision_dtype: 'bfloat16'
-task:
-  smart_bias_lr: 0.0
-  model:
-    darknet_based_model: True
-    input_size: [512, 512, 3]
-    backbone:
-      type: 'swin'
-      swin:
-        min_level: 3
-        max_level: 5
-        patch_size: 4
-        embed_dims: 96 
-        window_size: [7, 7, 7, 7] 
-        depths: [2, 2, 6, 2]
-        num_heads: [3, 6, 12, 24]
-        drop_path: 0.0
-        absolute_positional_embed: False
-    decoder:
-      type: yolo_decoder
-      yolo_decoder:
-        version: v4
-        type: csp
-        activation: leaky
-    head:
-      smart_bias: true
-    detection_generator:
-      box_type:
-        'all': original
-      scale_xy:
-        '5': 1.05
-        '4': 1.1
-        '3': 1.2
-      max_boxes: 200
-      nms_type: greedy
-      iou_thresh: 0.001
-      nms_thresh: 0.60
-    loss:
-      use_scaled_loss: False
-      box_loss_type:  
-        'all': ciou
-      ignore_thresh:
-        'all': 0.7
-      iou_normalizer: 
-        'all': 0.07
-      cls_normalizer: 
-        'all': 1.0
-      obj_normalizer: 
-        'all': 1.0
-      objectness_smooth: 
-        'all': 0.0
-      max_delta:
-        'all': 5.0
-    norm_activation:
-      activation: mish
-      norm_epsilon: 0.0001
-      norm_momentum: 0.99
-      use_sync_bn: true
-    num_classes: 80
-    anchor_boxes:
-      anchors_per_scale: 3
-      boxes: [box: [12, 16],   box: [19, 36],   box: [40, 28], 
-              box: [36, 75],   box: [76, 55],   box: [72, 146], 
-              box: [142, 110], box: [192, 243], box: [459, 401]]
-  train_data:
-    global_batch_size: 64
-    dtype: float32
-    input_path: 'gs://cam2-datasets/coco/train*'
-    is_training: true
-    drop_remainder: true
-    seed: 1000
-    parser:
-      mosaic: 
-        mosaic_frequency: 1.0
-        mixup_frequency: 0.0
-        mosaic_crop_mode: 'scale'
-        mosaic_center: 0.25
-        aug_scale_min: 0.1
-        aug_scale_max: 1.9
-      max_num_instances: 200
-      letter_box: True
-      random_flip: True
-      aug_rand_saturation: 0.7
-      aug_rand_brightness: 0.4
-      aug_rand_hue: 0.015
-      aug_rand_translate: 0.1
-      area_thresh: 0.1
-      random_pad: False
-      use_tie_breaker: True
-      anchor_thresh: 0.213
-  validation_data:
-    global_batch_size: 8
-    dtype: float32
-    input_path: 'gs://cam2-datasets/coco/val*'
-    is_training: false
-    drop_remainder: true
-    parser:
-      max_num_instances: 300
-      letter_box: True
-      use_tie_breaker: True
-      anchor_thresh: 0.213
-  weight_decay: 0.000
-  init_checkpoint: 'gs://tensorflow2/darknet/cspdarknet53-golden'
-  init_checkpoint_modules: 'backbone'
-  annotation_file: null
-trainer:
-  train_steps: 500500 # 160 epochs at 64 batchsize -> 500500 * 64/2
-  validation_steps: 625
-  steps_per_loop: 1850
-  summary_interval: 1850
-  validation_interval: 9250 
-  checkpoint_interval: 1850
-  optimizer_config:
-    ema: null
-    learning_rate:
-      type: stepwise
-      stepwise:
-        boundaries: [400000, 450000]
-        name: PiecewiseConstantDecay
-        values: [0.00131, 0.000131, 0.0000131] 
-    optimizer:
-      type: sgd_torch
-      sgd_torch:
-        momentum: 0.949
-        momentum_start: 0.949
-        nesterov: True
-        warmup_steps: 1000
-        weight_decay: 0.0005
-        sim_torch: true
-        name: SGD
-    warmup:
-      type: 'linear'
-      linear:
-        warmup_steps: 1000 #learning rate rises from 0 to 0.0013 over 1000 steps