Open source SpineNet configs with DeepMAC Mask R-CNN.

PiperOrigin-RevId: 392534821

official/vision/beta/projects/deepmac_maskrcnn/README.md

@@ -82,6 +82,8 @@ $ python3 -m official.vision.beta.projects.deepmac_maskrcnn.train \
 ```
 
 `CONFIG_FILE` can be any file in the `configs/experiments` directory.
+When using SpineNet models, please specify
+`--experiment=deep_mask_head_rcnn_spinenet_coco`
 
 **Note:** The default eval batch size of 32 discards some samples during
 validation. For accurate vaidation statistics, launch a dedicated eval job on

official/vision/beta/projects/deepmac_maskrcnn/configs/experiments/deep_mask_head_rcnn_nonvoc_spinenet143_hg52.yaml

+# Expect to reach: box mAP: 49.3%, mask mAP: 43.4% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  allowed_mask_class_ids: [
+    8, 10, 11, 13, 14, 15, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 35, 36,
+    37, 38, 39, 40, 41, 42, 43, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56,
+    57, 58, 59, 60, 61, 65, 70, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84,
+    85, 86, 87, 88, 89, 90
+  ]
+  per_category_metrics: true
+  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:
+    mask_head:
+      class_agnostic: true
+      convnet_variant: 'hourglass52'
+      num_filters: 64
+    mask_roi_aligner:
+      crop_size: 32
+    use_gt_boxes_for_masks: true
+    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/beta/projects/deepmac_maskrcnn/configs/experiments/deep_mask_head_rcnn_voc_spinenet143_hg52.yaml

+# Expect to reach: box mAP: 49.3%, mask mAP: 43.4% on COCO
+runtime:
+  distribution_strategy: 'tpu'
+  mixed_precision_dtype: 'bfloat16'
+task:
+  allowed_mask_class_ids: [1, 2, 3, 4, 5, 6, 7, 9, 16, 17, 18, 19, 20, 21, 44, 62, 63, 64, 67, 72]
+  per_category_metrics: true
+  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:
+    mask_head:
+      class_agnostic: true
+      convnet_variant: 'hourglass52'
+      num_filters: 64
+    mask_roi_aligner:
+      crop_size: 32
+    use_gt_boxes_for_masks: true
+    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