Merge branch 'tensorflow:master' into panoptic-deeplab-modeling

official/vision/configs/experiments/image_classification/imagenet_resnet50_tfds_tpu.yaml

+# ResNet-50 ImageNet classification. 78.1% top-1 and 93.9% top-5 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [224, 224, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 50
+    norm_activation:
+      activation: 'swish'
+  losses:
+    l2_weight_decay: 0.0001
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: ''
+    tfds_name: 'imagenet2012'
+    tfds_split: 'train'
+    sharding: true
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+  validation_data:
+    input_path: ''
+    tfds_name: 'imagenet2012'
+    tfds_split: 'validation'
+    sharding: true
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 62400
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 62400
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnet50_tpu.yaml

+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [224, 224, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 50
+  losses:
+    l2_weight_decay: 0.0001
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 28080
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [9360, 18720, 24960]
+        values: [1.6, 0.16, 0.016, 0.0016]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs101_i160.yaml

+# ResNet-RS-101 ImageNet classification. 80.2% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [160, 160, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 101
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.0
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs101_i192.yaml

+# ResNet-RS-101 ImageNet classification. 81.3% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [192, 192, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 101
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.0
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs152_i192.yaml

+# ResNet-RS-152 ImageNet classification. 81.9% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [192, 192, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 152
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.0
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs152_i224.yaml

+# ResNet-RS-152 ImageNet classification. 82.5% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [224, 224, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 152
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.0
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs152_i256.yaml

+# ResNet-RS-152 ImageNet classification. 83.1% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [256, 256, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 152
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.0
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs200_i256.yaml

+# ResNet-RS-200 ImageNet classification. 83.5% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [256, 256, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 200
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.1
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs270_i256.yaml

+# ResNet-RS-270 ImageNet classification. 83.6% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [256, 256, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 270
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.1
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs350_i256.yaml

+# ResNet-RS-350 ImageNet classification. 83.7% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [256, 256, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 350
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.1
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs350_i320.yaml

+# ResNet-RS-350 ImageNet classification. 84.2% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [320, 320, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 350
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.1
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.4
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs420_i320.yaml

+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [320, 320, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 420
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.1
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.4
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 15
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/image_classification/imagenet_resnetrs50_i160.yaml

+# ResNet-RS-50 ImageNet classification. 79.1% top-1 accuracy.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    num_classes: 1001
+    input_size: [160, 160, 3]
+    backbone:
+      type: 'resnet'
+      resnet:
+        model_id: 50
+        replace_stem_max_pool: true
+        resnetd_shortcut: true
+        se_ratio: 0.25
+        stem_type: 'v1'
+        stochastic_depth_drop_rate: 0.0
+    norm_activation:
+      activation: 'swish'
+      norm_momentum: 0.0
+      use_sync_bn: false
+    dropout_rate: 0.25
+  losses:
+    l2_weight_decay: 0.00004
+    one_hot: true
+    label_smoothing: 0.1
+  train_data:
+    input_path: 'imagenet-2012-tfrecord/train*'
+    is_training: true
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    aug_type:
+      type: 'randaug'
+      randaug:
+        magnitude: 10
+  validation_data:
+    input_path: 'imagenet-2012-tfrecord/valid*'
+    is_training: false
+    global_batch_size: 4096
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  train_steps: 109200
+  validation_steps: 13
+  validation_interval: 312
+  steps_per_loop: 312
+  summary_interval: 312
+  checkpoint_interval: 312
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+    optimizer:
+      type: 'sgd'
+      sgd:
+        momentum: 0.9
+    learning_rate:
+      type: 'cosine'
+      cosine:
+        initial_learning_rate: 1.6
+        decay_steps: 109200
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 1560

official/vision/configs/experiments/maskrcnn/coco_spinenet143_cascadercnn_tpu.yaml

+# --experiment_type=cascadercnn_spinenet_coco
+# Expect to reach: box mAP: 51.9%, mask mAP: 45.0% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.1
+      aug_scale_max: 2.5
+  losses:
+    l2_weight_decay: 0.00004
+  model:
+    anchor:
+      anchor_size: 4.0
+      num_scales: 3
+    min_level: 3
+    max_level: 7
+    input_size: [1280, 1280, 3]
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '143'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    detection_head:
+      cascade_class_ensemble: true
+      class_agnostic_bbox_pred: true
+    rpn_head:
+      num_convs: 2
+      num_filters: 256
+    roi_sampler:
+      cascade_iou_thresholds: [0.7]
+      foreground_iou_threshold: 0.6
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+      activation: 'swish'
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 231000
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067

official/vision/configs/experiments/maskrcnn/coco_spinenet143_mrcnn_tpu.yaml

+# Expect to reach: box mAP: 49.3%, mask mAP: 43.4% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.1
+      aug_scale_max: 2.0
+  losses:
+    l2_weight_decay: 0.00004
+  model:
+    anchor:
+      anchor_size: 4.0
+      num_scales: 3
+    min_level: 3
+    max_level: 7
+    input_size: [1280, 1280, 3]
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '143'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 231000
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067

official/vision/configs/experiments/maskrcnn/coco_spinenet49_cascadercnn_tpu.yaml

+# --experiment_type=cascadercnn_spinenet_coco
+# Expect to reach: box mAP: 46.4%, mask mAP: 40.0% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.1
+      aug_scale_max: 2.0
+  losses:
+    l2_weight_decay: 0.00004
+  model:
+    anchor:
+      anchor_size: 3.0
+      num_scales: 3
+    min_level: 3
+    max_level: 7
+    input_size: [640, 640, 3]
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '49'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    detection_head:
+      cascade_class_ensemble: true
+      class_agnostic_bbox_pred: true
+    rpn_head:
+      num_convs: 2
+      num_filters: 256
+    roi_sampler:
+      cascade_iou_thresholds: [0.7]
+      foreground_iou_threshold: 0.6
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+      activation: 'swish'
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 231000
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067

official/vision/configs/experiments/maskrcnn/coco_spinenet49_mrcnn_tpu.yaml

+# Expect to reach: box mAP: 43.2%, mask mAP: 38.3% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.1
+      aug_scale_max: 2.0
+  losses:
+    l2_weight_decay: 0.00004
+  model:
+    anchor:
+      anchor_size: 3.0
+      num_scales: 3
+    min_level: 3
+    max_level: 7
+    input_size: [640, 640, 3]
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '49'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 231000
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067

official/vision/configs/experiments/maskrcnn/coco_spinenet96_cascadercnn_tpu.yaml

+# --experiment_type=cascadercnn_spinenet_coco
+# Expect to reach: box mAP: 51.9%, mask mAP: 45.0% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.1
+      aug_scale_max: 2.5
+  losses:
+    l2_weight_decay: 0.00004
+  model:
+    anchor:
+      anchor_size: 4.0
+      num_scales: 3
+    min_level: 3
+    max_level: 7
+    input_size: [1024, 1024, 3]
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '96'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    detection_head:
+      cascade_class_ensemble: true
+      class_agnostic_bbox_pred: true
+    rpn_head:
+      num_convs: 2
+      num_filters: 256
+    roi_sampler:
+      cascade_iou_thresholds: [0.7]
+      foreground_iou_threshold: 0.6
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+      activation: 'swish'
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 231000
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067

official/vision/configs/experiments/maskrcnn/coco_spinenet96_mrcnn_tpu.yaml

+# Expect to reach: box mAP: 48.1%, mask mAP: 42.4% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.1
+      aug_scale_max: 2.0
+  losses:
+    l2_weight_decay: 0.00004
+  model:
+    anchor:
+      anchor_size: 3.0
+      num_scales: 3
+    min_level: 3
+    max_level: 7
+    input_size: [1024, 1024, 3]
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '96'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 231000
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067

official/vision/configs/experiments/maskrcnn/r50fpn_640_coco_scratch_tpu4x4.yaml

+# Expect to reach: box mAP: 42.3%, mask mAP: 37.6% on COCO
+task:
+  init_checkpoint: null
+  train_data:
+    global_batch_size: 256
+    parser:
+      aug_rand_hflip: true
+      aug_scale_min: 0.5
+      aug_scale_max: 2.0
+  losses:
+    l2_weight_decay: 0.00008
+  model:
+    anchor:
+      anchor_size: 3.0
+    min_level: 3
+    max_level: 7
+    input_size: [640, 640, 3]
+    norm_activation:
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+    detection_generator:
+      pre_nms_top_k: 1000
+trainer:
+  train_steps: 162050
+  optimizer_config:
+    learning_rate:
+      type: 'stepwise'
+      stepwise:
+        boundaries: [148160, 157420]
+        values: [0.32, 0.032, 0.0032]
+    warmup:
+      type: 'linear'
+      linear:
+        warmup_steps: 2000
+        warmup_learning_rate: 0.0067