Internal change

PiperOrigin-RevId: 431756117

official/vision/configs/experiments/retinanet/coco_spinenet49s_mobile_tpu.yaml

+# --experiment_type=retinanet_mobile_coco
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  losses:
+    l2_weight_decay: 3.0e-05
+  model:
+    anchor:
+      anchor_size: 3
+      aspect_ratios: [0.5, 1.0, 2.0]
+      num_scales: 3
+    backbone:
+      spinenet_mobile:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '49S'
+        se_ratio: 0.2
+      type: 'spinenet_mobile'
+    decoder:
+      type: 'identity'
+    head:
+      num_convs: 4
+      num_filters: 40
+      use_separable_conv: true
+    input_size: [384, 384, 3]
+    max_level: 7
+    min_level: 3
+    norm_activation:
+      activation: 'swish'
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+  train_data:
+    dtype: 'bfloat16'
+    global_batch_size: 256
+    is_training: true
+    parser:
+      aug_rand_hflip: true
+      aug_scale_max: 2.0
+      aug_scale_min: 0.5
+  validation_data:
+    dtype: 'bfloat16'
+    global_batch_size: 8
+    is_training: false
+trainer:
+  checkpoint_interval: 462
+  optimizer_config:
+    learning_rate:
+      stepwise:
+        boundaries: [263340, 272580]
+        values: [0.32, 0.032, 0.0032]
+      type: 'stepwise'
+    warmup:
+      linear:
+        warmup_learning_rate: 0.0067
+        warmup_steps: 2000
+  steps_per_loop: 462
+  train_steps: 277200
+  validation_interval: 462
+  validation_steps: 625

official/vision/configs/experiments/retinanet/coco_spinenet49xs_mobile_tpu.yaml

+# --experiment_type=retinanet_mobile_coco
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  losses:
+    l2_weight_decay: 3.0e-05
+  model:
+    anchor:
+      anchor_size: 3
+      aspect_ratios: [0.5, 1.0, 2.0]
+      num_scales: 3
+    backbone:
+      spinenet_mobile:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '49XS'
+        se_ratio: 0.2
+      type: 'spinenet_mobile'
+    decoder:
+      type: 'identity'
+    head:
+      num_convs: 4
+      num_filters: 24
+      use_separable_conv: true
+    input_size: [256, 256, 3]
+    max_level: 7
+    min_level: 3
+    norm_activation:
+      activation: 'swish'
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+  train_data:
+    dtype: 'bfloat16'
+    global_batch_size: 256
+    is_training: true
+    parser:
+      aug_rand_hflip: true
+      aug_scale_max: 2.0
+      aug_scale_min: 0.5
+  validation_data:
+    dtype: 'bfloat16'
+    global_batch_size: 8
+    is_training: false
+trainer:
+  checkpoint_interval: 462
+  optimizer_config:
+    learning_rate:
+      stepwise:
+        boundaries: [263340, 272580]
+        values: [0.32, 0.032, 0.0032]
+      type: 'stepwise'
+    warmup:
+      linear:
+        warmup_learning_rate: 0.0067
+        warmup_steps: 2000
+  steps_per_loop: 462
+  train_steps: 277200
+  validation_interval: 462
+  validation_steps: 625

official/vision/configs/experiments/retinanet/coco_spinenet96_tpu.yaml

+# SpineNet-96 COCO detection with protocol C config. Expecting 48.5% AP.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  losses:
+    l2_weight_decay: 4.0e-05
+  model:
+    anchor:
+      anchor_size: 3
+      aspect_ratios: [0.5, 1.0, 2.0]
+      num_scales: 3
+    backbone:
+      spinenet:
+        stochastic_depth_drop_rate: 0.2
+        model_id: '96'
+      type: 'spinenet'
+    decoder:
+      type: 'identity'
+    head:
+      num_convs: 4
+      num_filters: 256
+    input_size: [1024, 1024, 3]
+    max_level: 7
+    min_level: 3
+    norm_activation:
+      activation: 'swish'
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+  train_data:
+    dtype: 'bfloat16'
+    global_batch_size: 256
+    is_training: true
+    parser:
+      aug_rand_hflip: true
+      aug_scale_max: 2.0
+      aug_scale_min: 0.1
+  validation_data:
+    dtype: 'bfloat16'
+    global_batch_size: 8
+    is_training: false
+trainer:
+  checkpoint_interval: 462
+  optimizer_config:
+    learning_rate:
+      stepwise:
+        boundaries: [219450, 226380]
+        values: [0.32, 0.032, 0.0032]
+      type: 'stepwise'
+    warmup:
+      linear:
+        warmup_learning_rate: 0.0067
+        warmup_steps: 2000
+  steps_per_loop: 462
+  train_steps: 231000
+  validation_interval: 462
+  validation_steps: 625

official/vision/configs/experiments/retinanet/resnet50fpn_coco_tfds_tpu.yaml

+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  annotation_file: ''  # Can't use annotation file when tfds is used.
+  losses:
+    l2_weight_decay: 0.0001
+  model:
+    num_classes: 91
+    max_level: 7
+    min_level: 3
+    input_size: [640, 640, 3]
+    norm_activation:
+      activation: relu
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+  train_data:
+    tfds_name: 'coco/2017'
+    tfds_split: 'train'
+    drop_remainder: true
+    dtype: bfloat16
+    global_batch_size: 256
+    input_path: ''
+    is_training: true
+    shuffle_buffer_size: 1000
+  validation_data:
+    tfds_name: 'coco/2017'
+    tfds_split: 'validation'
+    drop_remainder: true
+    dtype: bfloat16
+    global_batch_size: 8
+    input_path: ''
+    is_training: false

official/vision/configs/experiments/retinanet/resnet50fpn_coco_tpu4x4_benchmark.yaml

+# Benchmarks runs on same instnace, change eval batch size to fit on 4x4 tpu
+task:
+  validation_data:
+    global_batch_size: 32
+trainer:
+  validation_interval: 1560
+  validation_steps: 156

official/vision/configs/experiments/semantic_segmentation/deeplabv3plus_resnet101_cityscapes_tfds_tpu.yaml

+# Use your own cityscapes preprocessed dataset. 79% meanIoU.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'float32'
+task:
+  model:
+    num_classes: 19
+    input_size: [null, null, 3]
+    backbone:
+      type: 'dilated_resnet'
+      dilated_resnet:
+        model_id: 101
+        output_stride: 16
+        stem_type: 'v1'
+        se_ratio: 0.25
+        stochastic_depth_drop_rate: 0.2
+        multigrid: [1, 2, 4]
+        last_stage_repeats: 1
+    decoder:
+      aspp:
+        pool_kernel_size: [512, 1024]
+    head:
+      feature_fusion: 'deeplabv3plus'
+      low_level: 2
+      low_level_num_filters: 48
+    norm_activation:
+      activation: 'swish'
+      norm_epsilon: 0.001
+      norm_momentum: 0.99
+      use_sync_bn: true
+  losses:
+    top_k_percent_pixels: 1.0  # only backpropagate loss for the topk 100% pixels.
+  train_data:
+    output_size: [1024, 2048]
+    crop_size: [512, 1024]
+    input_path: ''
+    tfds_name: 'cityscapes/semantic_segmentation'
+    tfds_split: 'train'
+    is_training: true
+    global_batch_size: 16
+    dtype: 'float32'
+    aug_rand_hflip: true
+    aug_scale_max: 2.0
+    aug_scale_min: 0.5
+  validation_data:
+    output_size: [1024, 2048]
+    input_path: ''
+    tfds_name: 'cityscapes/semantic_segmentation'
+    tfds_split: 'validation'
+    is_training: false
+    global_batch_size: 16
+    dtype: 'float32'
+    drop_remainder: false
+    resize_eval_groundtruth: true
+trainer:
+  optimizer_config:
+    learning_rate:
+      polynomial:
+        decay_steps: 90000
+        initial_learning_rate: 0.01
+        power: 0.9
+      type: polynomial
+    optimizer:
+      sgd:
+        momentum: 0.9
+      type: sgd
+    warmup:
+      linear:
+        name: linear
+        warmup_learning_rate: 0
+        warmup_steps: 925
+      type: linear
+  steps_per_loop: 185
+  summary_interval: 185
+  train_steps: 90000
+  validation_interval: 185
+  validation_steps: 31
+  checkpoint_interval: 185

official/vision/configs/experiments/video_classification/k400_3d-resnet50_tpu.yaml

+# 3D ResNet-50 video classification on Kinetics-400.
+#
+# --experiment_type=video_classification_kinetics400
+# Expected accuracy: 77.0% top-1, 93.0% top-5.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    dropout_rate: 0.5
+    norm_activation:
+      use_sync_bn: false
+    backbone:
+      resnet_3d:
+        block_specs: !!python/tuple
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        model_id: 50
+        stem_conv_temporal_kernel_size: 5
+        stem_conv_temporal_stride: 2
+        stem_pool_temporal_stride: 1
+  train_data:
+    name: kinetics400
+    feature_shape: !!python/tuple
+    - 32
+    - 224
+    - 224
+    - 3
+    temporal_stride: 2
+    global_batch_size: 1024
+    dtype: 'bfloat16'
+    shuffle_buffer_size: 1024
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.08
+    aug_min_aspect_ratio: 0.5
+  validation_data:
+    name: kinetics400
+    feature_shape: !!python/tuple
+    - 32
+    - 256
+    - 256
+    - 3
+    temporal_stride: 2
+    num_test_clips: 10
+    num_test_crops: 3
+    global_batch_size: 64
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  optimizer_config:
+    learning_rate:
+      cosine:
+        initial_learning_rate: 0.8
+        decay_steps: 42104
+    warmup:
+      linear:
+        warmup_steps: 1053
+  train_steps: 42104
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/experiments/video_classification/k400_resnet3drs_50_tpu.yaml

+# 3D ResNet-RS-50 video classification on Kinetics-400.
+#
+# --experiment_type=video_classification_kinetics400
+# Expected accuracy: 78.2% top-1 accuracy.
+runtime:
+  mixed_precision_dtype: bfloat16
+task:
+  losses:
+    l2_weight_decay: 0.00004
+    label_smoothing: 0.1
+    one_hot: true
+  model:
+    aggregate_endpoints: false
+    backbone:
+      resnet_3d_rs:
+        model_id: 50
+        stem_type: 'v1'
+        stem_conv_temporal_kernel_size: 5
+        stem_conv_temporal_stride: 2
+        stem_pool_temporal_stride: 1
+        stochastic_depth_drop_rate: 0.1
+        se_ratio: 0.25
+      type: resnet_3d_rs
+    dropout_rate: 0.5
+    model_type: video_classification
+    norm_activation:
+      activation: relu
+      norm_epsilon: 1.0e-05
+      norm_momentum: 0.0
+      use_sync_bn: false
+  train_data:
+    data_format: channels_last
+    drop_remainder: true
+    dtype: bfloat16
+    feature_shape: !!python/tuple
+    - 32
+    - 224
+    - 224
+    - 3
+    file_type: sstable
+    global_batch_size: 1024
+    is_training: true
+    min_image_size: 256
+    name: kinetics400
+    num_channels: 3
+    num_classes: 400
+    num_examples: 215570
+    num_test_clips: 1
+    num_test_crops: 1
+    one_hot: true
+    temporal_stride: 2
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.08
+    aug_min_aspect_ratio: 0.5
+  validation_data:
+    data_format: channels_last
+    drop_remainder: false
+    dtype: bfloat16
+    feature_shape: !!python/tuple
+    - 32
+    - 256
+    - 256
+    - 3
+    file_type: sstable
+    global_batch_size: 64
+    is_training: false
+    min_image_size: 256
+    name: kinetics400
+    num_channels: 3
+    num_classes: 400
+    num_examples: 17706
+    num_test_clips: 10
+    num_test_crops: 3
+    one_hot: true
+    temporal_stride: 2
+trainer:
+  checkpoint_interval: 210
+  max_to_keep: 3
+  optimizer_config:
+    ema:
+      average_decay: 0.9999
+      trainable_weights_only: false
+    learning_rate:
+      cosine:
+        decay_steps: 73682
+        initial_learning_rate: 0.8
+        name: CosineDecay
+      type: cosine
+    warmup:
+      linear:
+        name: linear
+        warmup_learning_rate: 0
+        warmup_steps: 1050
+      type: linear
+  train_steps: 73682
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/experiments/video_classification/k400_slowonly16x4_tpu.yaml

+# SlowOnly 16x4 video classification on Kinetics-400.
+#
+# --experiment_type=video_classification_kinetics400
+# Expected accuracy: 75.6% top-1, 92.1% top-5.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    dropout_rate: 0.5
+    norm_activation:
+      use_sync_bn: false
+    backbone:
+      resnet_3d:
+        block_specs: !!python/tuple
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        model_id: 50
+        stem_conv_temporal_kernel_size: 1
+        stem_conv_temporal_stride: 1
+        stem_pool_temporal_stride: 1
+  train_data:
+    name: kinetics400
+    feature_shape: !!python/tuple
+    - 16
+    - 224
+    - 224
+    - 3
+    temporal_stride: 4
+    global_batch_size: 1024
+    dtype: 'bfloat16'
+    shuffle_buffer_size: 1024
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.08
+    aug_min_aspect_ratio: 0.5
+  validation_data:
+    name: kinetics400
+    feature_shape: !!python/tuple
+    - 16
+    - 256
+    - 256
+    - 3
+    temporal_stride: 4
+    num_test_clips: 10
+    num_test_crops: 3
+    global_batch_size: 64
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  optimizer_config:
+    learning_rate:
+      cosine:
+        initial_learning_rate: 0.8
+        decay_steps: 42104
+    warmup:
+      linear:
+        warmup_steps: 1053
+  train_steps: 42104
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/experiments/video_classification/k400_slowonly8x8_tpu.yaml

+# SlowOnly 8x8 video classification on Kinetics-400.
+#
+# --experiment_type=video_classification_kinetics400
+# Expected accuracy: 74.1% top-1, 91.4% top-5.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    dropout_rate: 0.5
+    norm_activation:
+      use_sync_bn: false
+    backbone:
+      resnet_3d:
+        block_specs: !!python/tuple
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        model_id: 50
+        stem_conv_temporal_kernel_size: 1
+        stem_conv_temporal_stride: 1
+        stem_pool_temporal_stride: 1
+  train_data:
+    name: kinetics400
+    feature_shape: !!python/tuple
+    - 8
+    - 224
+    - 224
+    - 3
+    temporal_stride: 8
+    global_batch_size: 1024
+    dtype: 'bfloat16'
+    shuffle_buffer_size: 1024
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.08
+    aug_min_aspect_ratio: 0.5
+  validation_data:
+    name: kinetics400
+    feature_shape: !!python/tuple
+    - 8
+    - 256
+    - 256
+    - 3
+    temporal_stride: 8
+    num_test_clips: 10
+    num_test_crops: 3
+    global_batch_size: 64
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  optimizer_config:
+    learning_rate:
+      cosine:
+        initial_learning_rate: 0.8
+        decay_steps: 42104
+    warmup:
+      linear:
+        warmup_steps: 1053
+  train_steps: 42104
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/experiments/video_classification/k600_3d-resnet50_tpu.yaml

+# 3D ResNet-50 video classification on Kinetics-600.
+#
+# --experiment_type=video_classification_kinetics600
+# Expected accuracy: 79.5% top-1, 94.8% top-5.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    dropout_rate: 0.5
+    norm_activation:
+      use_sync_bn: false
+    backbone:
+      resnet_3d:
+        block_specs: !!python/tuple
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        model_id: 50
+        stem_conv_temporal_kernel_size: 5
+        stem_conv_temporal_stride: 2
+        stem_pool_temporal_stride: 1
+  train_data:
+    name: kinetics600
+    feature_shape: !!python/tuple
+    - 32
+    - 224
+    - 224
+    - 3
+    temporal_stride: 2
+    global_batch_size: 1024
+    dtype: 'bfloat16'
+    shuffle_buffer_size: 1024
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.08
+    aug_min_aspect_ratio: 0.5
+  validation_data:
+    name: kinetics600
+    feature_shape: !!python/tuple
+    - 32
+    - 256
+    - 256
+    - 3
+    temporal_stride: 2
+    num_test_clips: 10
+    num_test_crops: 3
+    global_batch_size: 64
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  optimizer_config:
+    learning_rate:
+      cosine:
+        initial_learning_rate: 0.8
+        decay_steps: 71488
+    warmup:
+      linear:
+        warmup_steps: 1787
+  train_steps: 71488
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/experiments/video_classification/k600_3d-resnet50g_tpu.yaml

+# 3D ResNet-50g video classification on Kinetics-600.
+#
+# --experiment_type=video_classification_kinetics600
+# Expected accuracy: 78.7% accuracy, 93.6% top-5.
+# Train on TPU: v3-128, eval on TPU: v3-32
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  init_checkpoint: null
+  init_checkpoint_modules: all
+  losses:
+    l2_weight_decay: 0.0001
+    label_smoothing: 0.0
+  model:
+    aggregate_endpoints: false
+    backbone:
+      resnet_3d:
+        block_specs: !!python/tuple
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: true
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 1
+          - 3
+          - 1
+          temporal_strides: 1
+          use_self_gating: true
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 1
+          - 3
+          - 1
+          - 3
+          - 1
+          temporal_strides: 1
+          use_self_gating: true
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 3
+          - 1
+          temporal_strides: 1
+          use_self_gating: true
+        model_id: 50
+        stem_conv_temporal_kernel_size: 5
+        stem_conv_temporal_stride: 2
+        stem_pool_temporal_stride: 2
+        stem_type: v0
+        stochastic_depth_drop_rate: 0.0
+      type: resnet_3d
+    dropout_rate: 0.2
+    model_type: video_classification
+    norm_activation:
+      activation: relu
+      norm_epsilon: 1.0e-05
+      norm_momentum: 0.9
+      use_sync_bn: false
+  train_data:
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.49
+    aug_min_aspect_ratio: 0.5
+    drop_remainder: true
+    dtype: 'bfloat16'
+    feature_shape: !!python/tuple
+    - 64
+    - 224
+    - 224
+    - 3
+    global_batch_size: 1024
+    min_image_size: 256
+    name: kinetics600
+    num_classes: 600
+    split: train
+  validation_data:
+    dtype: 'bfloat16'
+    feature_shape: !!python/tuple
+    - 250
+    - 224
+    - 224
+    - 3
+    global_batch_size: 64
+    min_image_size: 256
+    name: kinetics600
+    num_classes: 600
+    num_examples: 27780
+    num_test_clips: 1
+    num_test_crops: 1
+    one_hot: true
+trainer:
+  optimizer_config:
+    learning_rate:
+      cosine:
+        alpha: 0.0
+        decay_steps: 71400
+        initial_learning_rate: 1.6
+        name: CosineDecay
+      type: cosine
+    warmup:
+      linear:
+        name: linear
+        warmup_learning_rate: 0
+        warmup_steps: 1785
+      type: linear
+  train_steps: 71400
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/experiments/video_classification/k600_slowonly8x8_tpu.yaml

+# SlowOnly 8x8 video classification on Kinetics-600.
+#
+# --experiment_type=video_classification_kinetics600
+# Expected accuracy: 77.3% top-1, 93.6% top-5.
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  model:
+    dropout_rate: 0.5
+    norm_activation:
+      use_sync_bn: false
+    backbone:
+      resnet_3d:
+        block_specs: !!python/tuple
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 1
+          - 1
+          - 1
+          - 1
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        - temporal_kernel_sizes: !!python/tuple
+          - 3
+          - 3
+          - 3
+          temporal_strides: 1
+          use_self_gating: false
+        model_id: 50
+        stem_conv_temporal_kernel_size: 1
+        stem_conv_temporal_stride: 1
+        stem_pool_temporal_stride: 1
+  train_data:
+    name: kinetics600
+    feature_shape: !!python/tuple
+    - 8
+    - 224
+    - 224
+    - 3
+    temporal_stride: 8
+    global_batch_size: 1024
+    dtype: 'bfloat16'
+    shuffle_buffer_size: 1024
+    aug_max_area_ratio: 1.0
+    aug_max_aspect_ratio: 2.0
+    aug_min_area_ratio: 0.08
+    aug_min_aspect_ratio: 0.5
+  validation_data:
+    name: kinetics600
+    feature_shape: !!python/tuple
+    - 8
+    - 256
+    - 256
+    - 3
+    temporal_stride: 8
+    num_test_clips: 10
+    num_test_crops: 3
+    global_batch_size: 64
+    dtype: 'bfloat16'
+    drop_remainder: false
+trainer:
+  optimizer_config:
+    learning_rate:
+      cosine:
+        initial_learning_rate: 0.8
+        decay_steps: 71488
+    warmup:
+      linear:
+        warmup_steps: 1787
+  train_steps: 71488
+  steps_per_loop: 500
+  summary_interval: 500
+  validation_interval: 500

official/vision/configs/image_classification.py

+# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Image classification configuration definition."""
+import dataclasses
+import os
+from typing import List, Optional
+
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.modeling import hyperparams
+from official.modeling import optimization
+from official.vision.configs import common
+from official.vision.configs import backbones
+
+
+@dataclasses.dataclass
+class DataConfig(cfg.DataConfig):
+  """Input config for training."""
+  input_path: str = ''
+  global_batch_size: int = 0
+  is_training: bool = True
+  dtype: str = 'float32'
+  shuffle_buffer_size: int = 10000
+  cycle_length: int = 10
+  is_multilabel: bool = False
+  aug_rand_hflip: bool = True
+  aug_type: Optional[
+      common.Augmentation] = None  # Choose from AutoAugment and RandAugment.
+  color_jitter: float = 0.
+  random_erasing: Optional[common.RandomErasing] = None
+  file_type: str = 'tfrecord'
+  image_field_key: str = 'image/encoded'
+  label_field_key: str = 'image/class/label'
+  decode_jpeg_only: bool = True
+  mixup_and_cutmix: Optional[common.MixupAndCutmix] = None
+  decoder: Optional[common.DataDecoder] = common.DataDecoder()
+
+  # Keep for backward compatibility.
+  aug_policy: Optional[str] = None  # None, 'autoaug', or 'randaug'.
+  randaug_magnitude: Optional[int] = 10
+
+
+@dataclasses.dataclass
+class ImageClassificationModel(hyperparams.Config):
+  """The model config."""
+  num_classes: int = 0
+  input_size: List[int] = dataclasses.field(default_factory=list)
+  backbone: backbones.Backbone = backbones.Backbone(
+      type='resnet', resnet=backbones.ResNet())
+  dropout_rate: float = 0.0
+  norm_activation: common.NormActivation = common.NormActivation(
+      use_sync_bn=False)
+  # Adds a BatchNormalization layer pre-GlobalAveragePooling in classification
+  add_head_batch_norm: bool = False
+  kernel_initializer: str = 'random_uniform'
+
+
+@dataclasses.dataclass
+class Losses(hyperparams.Config):
+  loss_weight: float = 1.0
+  one_hot: bool = True
+  label_smoothing: float = 0.0
+  l2_weight_decay: float = 0.0
+  soft_labels: bool = False
+
+
+@dataclasses.dataclass
+class Evaluation(hyperparams.Config):
+  top_k: int = 5
+
+
+@dataclasses.dataclass
+class ImageClassificationTask(cfg.TaskConfig):
+  """The task config."""
+  model: ImageClassificationModel = ImageClassificationModel()
+  train_data: DataConfig = DataConfig(is_training=True)
+  validation_data: DataConfig = DataConfig(is_training=False)
+  losses: Losses = Losses()
+  evaluation: Evaluation = Evaluation()
+  init_checkpoint: Optional[str] = None
+  init_checkpoint_modules: str = 'all'  # all or backbone
+  model_output_keys: Optional[List[int]] = dataclasses.field(
+      default_factory=list)
+
+
+@exp_factory.register_config_factory('image_classification')
+def image_classification() -> cfg.ExperimentConfig:
+  """Image classification general."""
+  return cfg.ExperimentConfig(
+      task=ImageClassificationTask(),
+      trainer=cfg.TrainerConfig(),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+
+
+IMAGENET_TRAIN_EXAMPLES = 1281167
+IMAGENET_VAL_EXAMPLES = 50000
+IMAGENET_INPUT_PATH_BASE = 'imagenet-2012-tfrecord'
+
+
+@exp_factory.register_config_factory('resnet_imagenet')
+def image_classification_imagenet() -> cfg.ExperimentConfig:
+  """Image classification on imagenet with resnet."""
+  train_batch_size = 4096
+  eval_batch_size = 4096
+  steps_per_epoch = IMAGENET_TRAIN_EXAMPLES // train_batch_size
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(enable_xla=True),
+      task=ImageClassificationTask(
+          model=ImageClassificationModel(
+              num_classes=1001,
+              input_size=[224, 224, 3],
+              backbone=backbones.Backbone(
+                  type='resnet', resnet=backbones.ResNet(model_id=50)),
+              norm_activation=common.NormActivation(
+                  norm_momentum=0.9, norm_epsilon=1e-5, use_sync_bn=False)),
+          losses=Losses(l2_weight_decay=1e-4),
+          train_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size),
+          validation_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'valid*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          train_steps=90 * steps_per_epoch,
+          validation_steps=IMAGENET_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          30 * steps_per_epoch, 60 * steps_per_epoch,
+                          80 * steps_per_epoch
+                      ],
+                      'values': [
+                          0.1 * train_batch_size / 256,
+                          0.01 * train_batch_size / 256,
+                          0.001 * train_batch_size / 256,
+                          0.0001 * train_batch_size / 256,
+                      ]
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 5 * steps_per_epoch,
+                      'warmup_learning_rate': 0
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+
+  return config
+
+
+@exp_factory.register_config_factory('resnet_rs_imagenet')
+def image_classification_imagenet_resnetrs() -> cfg.ExperimentConfig:
+  """Image classification on imagenet with resnet-rs."""
+  train_batch_size = 4096
+  eval_batch_size = 4096
+  steps_per_epoch = IMAGENET_TRAIN_EXAMPLES // train_batch_size
+  config = cfg.ExperimentConfig(
+      task=ImageClassificationTask(
+          model=ImageClassificationModel(
+              num_classes=1001,
+              input_size=[160, 160, 3],
+              backbone=backbones.Backbone(
+                  type='resnet',
+                  resnet=backbones.ResNet(
+                      model_id=50,
+                      stem_type='v1',
+                      resnetd_shortcut=True,
+                      replace_stem_max_pool=True,
+                      se_ratio=0.25,
+                      stochastic_depth_drop_rate=0.0)),
+              dropout_rate=0.25,
+              norm_activation=common.NormActivation(
+                  norm_momentum=0.0,
+                  norm_epsilon=1e-5,
+                  use_sync_bn=False,
+                  activation='swish')),
+          losses=Losses(l2_weight_decay=4e-5, label_smoothing=0.1),
+          train_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              aug_type=common.Augmentation(
+                  type='randaug', randaug=common.RandAugment(magnitude=10))),
+          validation_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'valid*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          train_steps=350 * steps_per_epoch,
+          validation_steps=IMAGENET_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'ema': {
+                  'average_decay': 0.9999,
+                  'trainable_weights_only': False,
+              },
+              'learning_rate': {
+                  'type': 'cosine',
+                  'cosine': {
+                      'initial_learning_rate': 1.6,
+                      'decay_steps': 350 * steps_per_epoch
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 5 * steps_per_epoch,
+                      'warmup_learning_rate': 0
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+  return config
+
+
+@exp_factory.register_config_factory('revnet_imagenet')
+def image_classification_imagenet_revnet() -> cfg.ExperimentConfig:
+  """Returns a revnet config for image classification on imagenet."""
+  train_batch_size = 4096
+  eval_batch_size = 4096
+  steps_per_epoch = IMAGENET_TRAIN_EXAMPLES // train_batch_size
+
+  config = cfg.ExperimentConfig(
+      task=ImageClassificationTask(
+          model=ImageClassificationModel(
+              num_classes=1001,
+              input_size=[224, 224, 3],
+              backbone=backbones.Backbone(
+                  type='revnet', revnet=backbones.RevNet(model_id=56)),
+              norm_activation=common.NormActivation(
+                  norm_momentum=0.9, norm_epsilon=1e-5, use_sync_bn=False),
+              add_head_batch_norm=True),
+          losses=Losses(l2_weight_decay=1e-4),
+          train_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size),
+          validation_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'valid*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          train_steps=90 * steps_per_epoch,
+          validation_steps=IMAGENET_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          30 * steps_per_epoch, 60 * steps_per_epoch,
+                          80 * steps_per_epoch
+                      ],
+                      'values': [0.8, 0.08, 0.008, 0.0008]
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 5 * steps_per_epoch,
+                      'warmup_learning_rate': 0
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+
+  return config
+
+
+@exp_factory.register_config_factory('mobilenet_imagenet')
+def image_classification_imagenet_mobilenet() -> cfg.ExperimentConfig:
+  """Image classification on imagenet with mobilenet."""
+  train_batch_size = 4096
+  eval_batch_size = 4096
+  steps_per_epoch = IMAGENET_TRAIN_EXAMPLES // train_batch_size
+  config = cfg.ExperimentConfig(
+      task=ImageClassificationTask(
+          model=ImageClassificationModel(
+              num_classes=1001,
+              dropout_rate=0.2,
+              input_size=[224, 224, 3],
+              backbone=backbones.Backbone(
+                  type='mobilenet',
+                  mobilenet=backbones.MobileNet(
+                      model_id='MobileNetV2', filter_size_scale=1.0)),
+              norm_activation=common.NormActivation(
+                  norm_momentum=0.997, norm_epsilon=1e-3, use_sync_bn=False)),
+          losses=Losses(l2_weight_decay=1e-5, label_smoothing=0.1),
+          train_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size),
+          validation_data=DataConfig(
+              input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'valid*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          train_steps=500 * steps_per_epoch,
+          validation_steps=IMAGENET_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'rmsprop',
+                  'rmsprop': {
+                      'rho': 0.9,
+                      'momentum': 0.9,
+                      'epsilon': 0.002,
+                  }
+              },
+              'learning_rate': {
+                  'type': 'exponential',
+                  'exponential': {
+                      'initial_learning_rate':
+                          0.008 * (train_batch_size // 128),
+                      'decay_steps':
+                          int(2.5 * steps_per_epoch),
+                      'decay_rate':
+                          0.98,
+                      'staircase':
+                          True
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 5 * steps_per_epoch,
+                      'warmup_learning_rate': 0
+                  }
+              },
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+
+  return config

official/vision/configs/image_classification_test.py

+# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Tests for image_classification."""
+# pylint: disable=unused-import
+from absl.testing import parameterized
+import tensorflow as tf
+
+from official import vision
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.vision.configs import image_classification as exp_cfg
+
+
+class ImageClassificationConfigTest(tf.test.TestCase, parameterized.TestCase):
+
+  @parameterized.parameters(
+      ('resnet_imagenet',),
+      ('resnet_rs_imagenet',),
+      ('revnet_imagenet',),
+      ('mobilenet_imagenet'),
+  )
+  def test_image_classification_configs(self, config_name):
+    config = exp_factory.get_exp_config(config_name)
+    self.assertIsInstance(config, cfg.ExperimentConfig)
+    self.assertIsInstance(config.task, exp_cfg.ImageClassificationTask)
+    self.assertIsInstance(config.task.model,
+                          exp_cfg.ImageClassificationModel)
+    self.assertIsInstance(config.task.train_data, exp_cfg.DataConfig)
+    config.validate()
+    config.task.train_data.is_training = None
+    with self.assertRaises(KeyError):
+      config.validate()
+
+
+if __name__ == '__main__':
+  tf.test.main()

official/vision/configs/maskrcnn.py

+# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""R-CNN(-RS) configuration definition."""
+
+import dataclasses
+import os
+from typing import List, Optional, Union
+
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.modeling import hyperparams
+from official.modeling import optimization
+from official.vision.configs import common
+from official.vision.configs import decoders
+from official.vision.configs import backbones
+
+
+# pylint: disable=missing-class-docstring
+@dataclasses.dataclass
+class Parser(hyperparams.Config):
+  num_channels: int = 3
+  match_threshold: float = 0.5
+  unmatched_threshold: float = 0.5
+  aug_rand_hflip: bool = False
+  aug_scale_min: float = 1.0
+  aug_scale_max: float = 1.0
+  skip_crowd_during_training: bool = True
+  max_num_instances: int = 100
+  rpn_match_threshold: float = 0.7
+  rpn_unmatched_threshold: float = 0.3
+  rpn_batch_size_per_im: int = 256
+  rpn_fg_fraction: float = 0.5
+  mask_crop_size: int = 112
+
+
+@dataclasses.dataclass
+class DataConfig(cfg.DataConfig):
+  """Input config for training."""
+  input_path: str = ''
+  global_batch_size: int = 0
+  is_training: bool = False
+  dtype: str = 'bfloat16'
+  decoder: common.DataDecoder = common.DataDecoder()
+  parser: Parser = Parser()
+  shuffle_buffer_size: int = 10000
+  file_type: str = 'tfrecord'
+  drop_remainder: bool = True
+  # Number of examples in the data set, it's used to create the annotation file.
+  num_examples: int = -1
+
+
+@dataclasses.dataclass
+class Anchor(hyperparams.Config):
+  num_scales: int = 1
+  aspect_ratios: List[float] = dataclasses.field(
+      default_factory=lambda: [0.5, 1.0, 2.0])
+  anchor_size: float = 8.0
+
+
+@dataclasses.dataclass
+class RPNHead(hyperparams.Config):
+  num_convs: int = 1
+  num_filters: int = 256
+  use_separable_conv: bool = False
+
+
+@dataclasses.dataclass
+class DetectionHead(hyperparams.Config):
+  num_convs: int = 4
+  num_filters: int = 256
+  use_separable_conv: bool = False
+  num_fcs: int = 1
+  fc_dims: int = 1024
+  class_agnostic_bbox_pred: bool = False  # Has to be True for Cascade RCNN.
+  # If additional IoUs are passed in 'cascade_iou_thresholds'
+  # then ensemble the class probabilities from all heads.
+  cascade_class_ensemble: bool = False
+
+
+@dataclasses.dataclass
+class ROIGenerator(hyperparams.Config):
+  pre_nms_top_k: int = 2000
+  pre_nms_score_threshold: float = 0.0
+  pre_nms_min_size_threshold: float = 0.0
+  nms_iou_threshold: float = 0.7
+  num_proposals: int = 1000
+  test_pre_nms_top_k: int = 1000
+  test_pre_nms_score_threshold: float = 0.0
+  test_pre_nms_min_size_threshold: float = 0.0
+  test_nms_iou_threshold: float = 0.7
+  test_num_proposals: int = 1000
+  use_batched_nms: bool = False
+
+
+@dataclasses.dataclass
+class ROISampler(hyperparams.Config):
+  mix_gt_boxes: bool = True
+  num_sampled_rois: int = 512
+  foreground_fraction: float = 0.25
+  foreground_iou_threshold: float = 0.5
+  background_iou_high_threshold: float = 0.5
+  background_iou_low_threshold: float = 0.0
+  # IoU thresholds for additional FRCNN heads in Cascade mode.
+  # `foreground_iou_threshold` is the first threshold.
+  cascade_iou_thresholds: Optional[List[float]] = None
+
+
+@dataclasses.dataclass
+class ROIAligner(hyperparams.Config):
+  crop_size: int = 7
+  sample_offset: float = 0.5
+
+
+@dataclasses.dataclass
+class DetectionGenerator(hyperparams.Config):
+  apply_nms: bool = True
+  pre_nms_top_k: int = 5000
+  pre_nms_score_threshold: float = 0.05
+  nms_iou_threshold: float = 0.5
+  max_num_detections: int = 100
+  nms_version: str = 'v2'  # `v2`, `v1`, `batched`
+  use_cpu_nms: bool = False
+  soft_nms_sigma: Optional[float] = None  # Only works when nms_version='v1'.
+
+
+@dataclasses.dataclass
+class MaskHead(hyperparams.Config):
+  upsample_factor: int = 2
+  num_convs: int = 4
+  num_filters: int = 256
+  use_separable_conv: bool = False
+  class_agnostic: bool = False
+
+
+@dataclasses.dataclass
+class MaskSampler(hyperparams.Config):
+  num_sampled_masks: int = 128
+
+
+@dataclasses.dataclass
+class MaskROIAligner(hyperparams.Config):
+  crop_size: int = 14
+  sample_offset: float = 0.5
+
+
+@dataclasses.dataclass
+class MaskRCNN(hyperparams.Config):
+  num_classes: int = 0
+  input_size: List[int] = dataclasses.field(default_factory=list)
+  min_level: int = 2
+  max_level: int = 6
+  anchor: Anchor = Anchor()
+  include_mask: bool = True
+  backbone: backbones.Backbone = backbones.Backbone(
+      type='resnet', resnet=backbones.ResNet())
+  decoder: decoders.Decoder = decoders.Decoder(
+      type='fpn', fpn=decoders.FPN())
+  rpn_head: RPNHead = RPNHead()
+  detection_head: DetectionHead = DetectionHead()
+  roi_generator: ROIGenerator = ROIGenerator()
+  roi_sampler: ROISampler = ROISampler()
+  roi_aligner: ROIAligner = ROIAligner()
+  detection_generator: DetectionGenerator = DetectionGenerator()
+  mask_head: Optional[MaskHead] = MaskHead()
+  mask_sampler: Optional[MaskSampler] = MaskSampler()
+  mask_roi_aligner: Optional[MaskROIAligner] = MaskROIAligner()
+  norm_activation: common.NormActivation = common.NormActivation(
+      norm_momentum=0.997,
+      norm_epsilon=0.0001,
+      use_sync_bn=True)
+
+
+@dataclasses.dataclass
+class Losses(hyperparams.Config):
+  loss_weight: float = 1.0
+  rpn_huber_loss_delta: float = 1. / 9.
+  frcnn_huber_loss_delta: float = 1.
+  l2_weight_decay: float = 0.0
+  rpn_score_weight: float = 1.0
+  rpn_box_weight: float = 1.0
+  frcnn_class_weight: float = 1.0
+  frcnn_box_weight: float = 1.0
+  mask_weight: float = 1.0
+
+
+@dataclasses.dataclass
+class MaskRCNNTask(cfg.TaskConfig):
+  model: MaskRCNN = MaskRCNN()
+  train_data: DataConfig = DataConfig(is_training=True)
+  validation_data: DataConfig = DataConfig(is_training=False,
+                                           drop_remainder=False)
+  losses: Losses = Losses()
+  init_checkpoint: Optional[str] = None
+  init_checkpoint_modules: Union[
+      str, List[str]] = 'all'  # all, backbone, and/or decoder
+  annotation_file: Optional[str] = None
+  per_category_metrics: bool = False
+  # If set, we only use masks for the specified class IDs.
+  allowed_mask_class_ids: Optional[List[int]] = None
+  # If set, the COCO metrics will be computed.
+  use_coco_metrics: bool = True
+  # If set, the Waymo Open Dataset evaluator would be used.
+  use_wod_metrics: bool = False
+
+
+COCO_INPUT_PATH_BASE = 'coco'
+
+
+@exp_factory.register_config_factory('fasterrcnn_resnetfpn_coco')
+def fasterrcnn_resnetfpn_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with Faster R-CNN."""
+  steps_per_epoch = 500
+  coco_val_samples = 5000
+  train_batch_size = 64
+  eval_batch_size = 8
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(mixed_precision_dtype='bfloat16'),
+      task=MaskRCNNTask(
+          init_checkpoint='gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080',
+          init_checkpoint_modules='backbone',
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=MaskRCNN(
+              num_classes=91,
+              input_size=[1024, 1024, 3],
+              include_mask=False,
+              mask_head=None,
+              mask_sampler=None,
+              mask_roi_aligner=None),
+          losses=Losses(l2_weight_decay=0.00004),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.8, aug_scale_max=1.25)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size,
+              drop_remainder=False)),
+      trainer=cfg.TrainerConfig(
+          train_steps=22500,
+          validation_steps=coco_val_samples // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [15000, 20000],
+                      'values': [0.12, 0.012, 0.0012],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 500,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+  return config
+
+
+@exp_factory.register_config_factory('maskrcnn_resnetfpn_coco')
+def maskrcnn_resnetfpn_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with Mask R-CNN."""
+  steps_per_epoch = 500
+  coco_val_samples = 5000
+  train_batch_size = 64
+  eval_batch_size = 8
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(
+          mixed_precision_dtype='bfloat16', enable_xla=True),
+      task=MaskRCNNTask(
+          init_checkpoint='gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080',
+          init_checkpoint_modules='backbone',
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=MaskRCNN(
+              num_classes=91, input_size=[1024, 1024, 3], include_mask=True),
+          losses=Losses(l2_weight_decay=0.00004),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.8, aug_scale_max=1.25)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size,
+              drop_remainder=False)),
+      trainer=cfg.TrainerConfig(
+          train_steps=22500,
+          validation_steps=coco_val_samples // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [15000, 20000],
+                      'values': [0.12, 0.012, 0.0012],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 500,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+  return config
+
+
+@exp_factory.register_config_factory('maskrcnn_spinenet_coco')
+def maskrcnn_spinenet_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with Mask R-CNN with SpineNet backbone."""
+  steps_per_epoch = 463
+  coco_val_samples = 5000
+  train_batch_size = 256
+  eval_batch_size = 8
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(mixed_precision_dtype='bfloat16'),
+      task=MaskRCNNTask(
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=MaskRCNN(
+              backbone=backbones.Backbone(
+                  type='spinenet',
+                  spinenet=backbones.SpineNet(
+                      model_id='49',
+                      min_level=3,
+                      max_level=7,
+                  )),
+              decoder=decoders.Decoder(
+                  type='identity', identity=decoders.Identity()),
+              anchor=Anchor(anchor_size=3),
+              norm_activation=common.NormActivation(use_sync_bn=True),
+              num_classes=91,
+              input_size=[640, 640, 3],
+              min_level=3,
+              max_level=7,
+              include_mask=True),
+          losses=Losses(l2_weight_decay=0.00004),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.5, aug_scale_max=2.0)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size,
+              drop_remainder=False)),
+      trainer=cfg.TrainerConfig(
+          train_steps=steps_per_epoch * 350,
+          validation_steps=coco_val_samples // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          steps_per_epoch * 320, steps_per_epoch * 340
+                      ],
+                      'values': [0.32, 0.032, 0.0032],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 2000,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None',
+          'task.model.min_level == task.model.backbone.spinenet.min_level',
+          'task.model.max_level == task.model.backbone.spinenet.max_level',
+      ])
+  return config
+
+
+@exp_factory.register_config_factory('cascadercnn_spinenet_coco')
+def cascadercnn_spinenet_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with Cascade RCNN-RS with SpineNet backbone."""
+  steps_per_epoch = 463
+  coco_val_samples = 5000
+  train_batch_size = 256
+  eval_batch_size = 8
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(mixed_precision_dtype='bfloat16'),
+      task=MaskRCNNTask(
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=MaskRCNN(
+              backbone=backbones.Backbone(
+                  type='spinenet',
+                  spinenet=backbones.SpineNet(
+                      model_id='49',
+                      min_level=3,
+                      max_level=7,
+                  )),
+              decoder=decoders.Decoder(
+                  type='identity', identity=decoders.Identity()),
+              roi_sampler=ROISampler(cascade_iou_thresholds=[0.6, 0.7]),
+              detection_head=DetectionHead(
+                  class_agnostic_bbox_pred=True, cascade_class_ensemble=True),
+              anchor=Anchor(anchor_size=3),
+              norm_activation=common.NormActivation(
+                  use_sync_bn=True, activation='swish'),
+              num_classes=91,
+              input_size=[640, 640, 3],
+              min_level=3,
+              max_level=7,
+              include_mask=True),
+          losses=Losses(l2_weight_decay=0.00004),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.1, aug_scale_max=2.5)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size,
+              drop_remainder=False)),
+      trainer=cfg.TrainerConfig(
+          train_steps=steps_per_epoch * 500,
+          validation_steps=coco_val_samples // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          steps_per_epoch * 475, steps_per_epoch * 490
+                      ],
+                      'values': [0.32, 0.032, 0.0032],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 2000,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None',
+          'task.model.min_level == task.model.backbone.spinenet.min_level',
+          'task.model.max_level == task.model.backbone.spinenet.max_level',
+      ])
+  return config

official/vision/configs/maskrcnn_test.py

+# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for maskrcnn."""
+# pylint: disable=unused-import
+from absl.testing import parameterized
+import tensorflow as tf
+
+from official import vision
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.vision.configs import maskrcnn as exp_cfg
+
+
+class MaskRCNNConfigTest(tf.test.TestCase, parameterized.TestCase):
+
+  @parameterized.parameters(
+      ('fasterrcnn_resnetfpn_coco',),
+      ('maskrcnn_resnetfpn_coco',),
+      ('maskrcnn_spinenet_coco',),
+      ('cascadercnn_spinenet_coco',),
+  )
+  def test_maskrcnn_configs(self, config_name):
+    config = exp_factory.get_exp_config(config_name)
+    self.assertIsInstance(config, cfg.ExperimentConfig)
+    self.assertIsInstance(config.task, exp_cfg.MaskRCNNTask)
+    self.assertIsInstance(config.task.model, exp_cfg.MaskRCNN)
+    self.assertIsInstance(config.task.train_data, exp_cfg.DataConfig)
+    config.validate()
+    config.task.train_data.is_training = None
+    with self.assertRaisesRegex(KeyError, 'Found inconsistncy between key'):
+      config.validate()
+
+
+if __name__ == '__main__':
+  tf.test.main()

official/vision/configs/retinanet.py

+# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""RetinaNet configuration definition."""
+
+import dataclasses
+import os
+from typing import List, Optional, Union
+
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.modeling import hyperparams
+from official.modeling import optimization
+from official.vision.configs import common
+from official.vision.configs import decoders
+from official.vision.configs import backbones
+
+
+# pylint: disable=missing-class-docstring
+# Keep for backward compatibility.
+@dataclasses.dataclass
+class TfExampleDecoder(common.TfExampleDecoder):
+  """A simple TF Example decoder config."""
+
+
+# Keep for backward compatibility.
+@dataclasses.dataclass
+class TfExampleDecoderLabelMap(common.TfExampleDecoderLabelMap):
+  """TF Example decoder with label map config."""
+
+
+# Keep for backward compatibility.
+@dataclasses.dataclass
+class DataDecoder(common.DataDecoder):
+  """Data decoder config."""
+
+
+@dataclasses.dataclass
+class Parser(hyperparams.Config):
+  num_channels: int = 3
+  match_threshold: float = 0.5
+  unmatched_threshold: float = 0.5
+  aug_rand_hflip: bool = False
+  aug_scale_min: float = 1.0
+  aug_scale_max: float = 1.0
+  skip_crowd_during_training: bool = True
+  max_num_instances: int = 100
+  # Can choose AutoAugment and RandAugment.
+  aug_type: Optional[common.Augmentation] = None
+
+  # Keep for backward compatibility. Not used.
+  aug_policy: Optional[str] = None
+
+
+@dataclasses.dataclass
+class DataConfig(cfg.DataConfig):
+  """Input config for training."""
+  input_path: str = ''
+  global_batch_size: int = 0
+  is_training: bool = False
+  dtype: str = 'bfloat16'
+  decoder: common.DataDecoder = common.DataDecoder()
+  parser: Parser = Parser()
+  shuffle_buffer_size: int = 10000
+  file_type: str = 'tfrecord'
+
+
+@dataclasses.dataclass
+class Anchor(hyperparams.Config):
+  num_scales: int = 3
+  aspect_ratios: List[float] = dataclasses.field(
+      default_factory=lambda: [0.5, 1.0, 2.0])
+  anchor_size: float = 4.0
+
+
+@dataclasses.dataclass
+class Losses(hyperparams.Config):
+  loss_weight: float = 1.0
+  focal_loss_alpha: float = 0.25
+  focal_loss_gamma: float = 1.5
+  huber_loss_delta: float = 0.1
+  box_loss_weight: int = 50
+  l2_weight_decay: float = 0.0
+
+
+@dataclasses.dataclass
+class AttributeHead(hyperparams.Config):
+  name: str = ''
+  type: str = 'regression'
+  size: int = 1
+
+
+@dataclasses.dataclass
+class RetinaNetHead(hyperparams.Config):
+  num_convs: int = 4
+  num_filters: int = 256
+  use_separable_conv: bool = False
+  attribute_heads: List[AttributeHead] = dataclasses.field(default_factory=list)
+
+
+@dataclasses.dataclass
+class DetectionGenerator(hyperparams.Config):
+  apply_nms: bool = True
+  pre_nms_top_k: int = 5000
+  pre_nms_score_threshold: float = 0.05
+  nms_iou_threshold: float = 0.5
+  max_num_detections: int = 100
+  nms_version: str = 'v2'  # `v2`, `v1`, `batched`, or `tflite`.
+  use_cpu_nms: bool = False
+  soft_nms_sigma: Optional[float] = None  # Only works when nms_version='v1'.
+
+  # When nms_version = `tflite`, values from tflite_post_processing need to be
+  # specified. They are compatible with the input arguments used by TFLite
+  # custom NMS op and override above parameters.
+  tflite_post_processing: common.TFLitePostProcessingConfig = common.TFLitePostProcessingConfig(
+  )
+
+  max_detections: int = 200
+  max_classes_per_detection: int = 5
+  # Regular NMS run in a multi-class fashion and is slow. Setting it to False
+  # uses class-agnostic NMS, which is faster.
+  use_regular_nms: bool = False
+  nms_score_threshold: float = 0.1
+
+
+@dataclasses.dataclass
+class RetinaNet(hyperparams.Config):
+  num_classes: int = 0
+  input_size: List[int] = dataclasses.field(default_factory=list)
+  min_level: int = 3
+  max_level: int = 7
+  anchor: Anchor = Anchor()
+  backbone: backbones.Backbone = backbones.Backbone(
+      type='resnet', resnet=backbones.ResNet())
+  decoder: decoders.Decoder = decoders.Decoder(
+      type='fpn', fpn=decoders.FPN())
+  head: RetinaNetHead = RetinaNetHead()
+  detection_generator: DetectionGenerator = DetectionGenerator()
+  norm_activation: common.NormActivation = common.NormActivation()
+
+
+@dataclasses.dataclass
+class ExportConfig(hyperparams.Config):
+  output_normalized_coordinates: bool = False
+  cast_num_detections_to_float: bool = False
+  cast_detection_classes_to_float: bool = False
+
+
+@dataclasses.dataclass
+class RetinaNetTask(cfg.TaskConfig):
+  model: RetinaNet = RetinaNet()
+  train_data: DataConfig = DataConfig(is_training=True)
+  validation_data: DataConfig = DataConfig(is_training=False)
+  losses: Losses = Losses()
+  init_checkpoint: Optional[str] = None
+  init_checkpoint_modules: Union[
+      str, List[str]] = 'all'  # all, backbone, and/or decoder
+  annotation_file: Optional[str] = None
+  per_category_metrics: bool = False
+  export_config: ExportConfig = ExportConfig()
+  # If set, the COCO metrics will be computed.
+  use_coco_metrics: bool = True
+  # If set, the Waymo Open Dataset evaluator would be used.
+  use_wod_metrics: bool = False
+
+
+@exp_factory.register_config_factory('retinanet')
+def retinanet() -> cfg.ExperimentConfig:
+  """RetinaNet general config."""
+  return cfg.ExperimentConfig(
+      task=RetinaNetTask(),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+
+
+COCO_INPUT_PATH_BASE = 'coco'
+COCO_TRAIN_EXAMPLES = 118287
+COCO_VAL_EXAMPLES = 5000
+
+
+@exp_factory.register_config_factory('retinanet_resnetfpn_coco')
+def retinanet_resnetfpn_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with RetinaNet."""
+  train_batch_size = 256
+  eval_batch_size = 8
+  steps_per_epoch = COCO_TRAIN_EXAMPLES // train_batch_size
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(mixed_precision_dtype='bfloat16'),
+      task=RetinaNetTask(
+          init_checkpoint='gs://cloud-tpu-checkpoints/vision-2.0/resnet50_imagenet/ckpt-28080',
+          init_checkpoint_modules='backbone',
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=RetinaNet(
+              num_classes=91,
+              input_size=[640, 640, 3],
+              norm_activation=common.NormActivation(use_sync_bn=False),
+              min_level=3,
+              max_level=7),
+          losses=Losses(l2_weight_decay=1e-4),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.8, aug_scale_max=1.2)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          train_steps=72 * steps_per_epoch,
+          validation_steps=COCO_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          57 * steps_per_epoch, 67 * steps_per_epoch
+                      ],
+                      'values': [
+                          0.32 * train_batch_size / 256.0,
+                          0.032 * train_batch_size / 256.0,
+                          0.0032 * train_batch_size / 256.0
+                      ],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 500,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None'
+      ])
+
+  return config
+
+
+@exp_factory.register_config_factory('retinanet_spinenet_coco')
+def retinanet_spinenet_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with RetinaNet using SpineNet backbone."""
+  train_batch_size = 256
+  eval_batch_size = 8
+  steps_per_epoch = COCO_TRAIN_EXAMPLES // train_batch_size
+  input_size = 640
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(mixed_precision_dtype='float32'),
+      task=RetinaNetTask(
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=RetinaNet(
+              backbone=backbones.Backbone(
+                  type='spinenet',
+                  spinenet=backbones.SpineNet(
+                      model_id='49',
+                      stochastic_depth_drop_rate=0.2,
+                      min_level=3,
+                      max_level=7)),
+              decoder=decoders.Decoder(
+                  type='identity', identity=decoders.Identity()),
+              anchor=Anchor(anchor_size=3),
+              norm_activation=common.NormActivation(
+                  use_sync_bn=True, activation='swish'),
+              num_classes=91,
+              input_size=[input_size, input_size, 3],
+              min_level=3,
+              max_level=7),
+          losses=Losses(l2_weight_decay=4e-5),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.1, aug_scale_max=2.0)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          train_steps=500 * steps_per_epoch,
+          validation_steps=COCO_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          475 * steps_per_epoch, 490 * steps_per_epoch
+                      ],
+                      'values': [
+                          0.32 * train_batch_size / 256.0,
+                          0.032 * train_batch_size / 256.0,
+                          0.0032 * train_batch_size / 256.0
+                      ],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 2000,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None',
+          'task.model.min_level == task.model.backbone.spinenet.min_level',
+          'task.model.max_level == task.model.backbone.spinenet.max_level',
+      ])
+
+  return config
+
+
+@exp_factory.register_config_factory('retinanet_mobile_coco')
+def retinanet_spinenet_mobile_coco() -> cfg.ExperimentConfig:
+  """COCO object detection with mobile RetinaNet."""
+  train_batch_size = 256
+  eval_batch_size = 8
+  steps_per_epoch = COCO_TRAIN_EXAMPLES // train_batch_size
+  input_size = 384
+
+  config = cfg.ExperimentConfig(
+      runtime=cfg.RuntimeConfig(mixed_precision_dtype='float32'),
+      task=RetinaNetTask(
+          annotation_file=os.path.join(COCO_INPUT_PATH_BASE,
+                                       'instances_val2017.json'),
+          model=RetinaNet(
+              backbone=backbones.Backbone(
+                  type='spinenet_mobile',
+                  spinenet_mobile=backbones.SpineNetMobile(
+                      model_id='49',
+                      stochastic_depth_drop_rate=0.2,
+                      min_level=3,
+                      max_level=7,
+                      use_keras_upsampling_2d=False)),
+              decoder=decoders.Decoder(
+                  type='identity', identity=decoders.Identity()),
+              head=RetinaNetHead(num_filters=48, use_separable_conv=True),
+              anchor=Anchor(anchor_size=3),
+              norm_activation=common.NormActivation(
+                  use_sync_bn=True, activation='swish'),
+              num_classes=91,
+              input_size=[input_size, input_size, 3],
+              min_level=3,
+              max_level=7),
+          losses=Losses(l2_weight_decay=3e-5),
+          train_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'train*'),
+              is_training=True,
+              global_batch_size=train_batch_size,
+              parser=Parser(
+                  aug_rand_hflip=True, aug_scale_min=0.1, aug_scale_max=2.0)),
+          validation_data=DataConfig(
+              input_path=os.path.join(COCO_INPUT_PATH_BASE, 'val*'),
+              is_training=False,
+              global_batch_size=eval_batch_size)),
+      trainer=cfg.TrainerConfig(
+          train_steps=600 * steps_per_epoch,
+          validation_steps=COCO_VAL_EXAMPLES // eval_batch_size,
+          validation_interval=steps_per_epoch,
+          steps_per_loop=steps_per_epoch,
+          summary_interval=steps_per_epoch,
+          checkpoint_interval=steps_per_epoch,
+          optimizer_config=optimization.OptimizationConfig({
+              'optimizer': {
+                  'type': 'sgd',
+                  'sgd': {
+                      'momentum': 0.9
+                  }
+              },
+              'learning_rate': {
+                  'type': 'stepwise',
+                  'stepwise': {
+                      'boundaries': [
+                          575 * steps_per_epoch, 590 * steps_per_epoch
+                      ],
+                      'values': [
+                          0.32 * train_batch_size / 256.0,
+                          0.032 * train_batch_size / 256.0,
+                          0.0032 * train_batch_size / 256.0
+                      ],
+                  }
+              },
+              'warmup': {
+                  'type': 'linear',
+                  'linear': {
+                      'warmup_steps': 2000,
+                      'warmup_learning_rate': 0.0067
+                  }
+              }
+          })),
+      restrictions=[
+          'task.train_data.is_training != None',
+          'task.validation_data.is_training != None',
+      ])
+
+  return config

official/vision/configs/retinanet_test.py

+# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for retinanet."""
+# pylint: disable=unused-import
+from absl.testing import parameterized
+import tensorflow as tf
+
+from official import vision
+from official.core import config_definitions as cfg
+from official.core import exp_factory
+from official.vision.configs import retinanet as exp_cfg
+
+
+class RetinaNetConfigTest(tf.test.TestCase, parameterized.TestCase):
+
+  @parameterized.parameters(
+      ('retinanet_resnetfpn_coco',),
+      ('retinanet_spinenet_coco',),
+      ('retinanet_mobile_coco',),
+  )
+  def test_retinanet_configs(self, config_name):
+    config = exp_factory.get_exp_config(config_name)
+    self.assertIsInstance(config, cfg.ExperimentConfig)
+    self.assertIsInstance(config.task, exp_cfg.RetinaNetTask)
+    self.assertIsInstance(config.task.model, exp_cfg.RetinaNet)
+    self.assertIsInstance(config.task.train_data, exp_cfg.DataConfig)
+    config.validate()
+    config.task.train_data.is_training = None
+    with self.assertRaisesRegex(KeyError, 'Found inconsistncy between key'):
+      config.validate()
+
+
+if __name__ == '__main__':
+  tf.test.main()