Merge branch 'panoptic-segmentation' into panoptic-deeplab-modeling

research/object_detection/protos/center_net.proto

@@ -403,7 +403,7 @@ message CenterNet {
 
 
   // Mask prediction support using DeepMAC. See https://arxiv.org/abs/2104.00613
-  // Next ID 24
+  // Next ID 25
   message DeepMACMaskEstimation {
     // The loss used for penalizing mask predictions.
     optional ClassificationLoss classification_loss = 1;
@@ -485,6 +485,14 @@ message CenterNet {
 
     optional int32 color_consistency_warmup_start = 23 [default=0];
 
+    // DeepMAC has been refactored to process the entire batch at once,
+    // instead of the previous (simple) approach of processing one sample at
+    // a time. Because of this, the memory consumption has increased and
+    // it's crucial to only feed the mask head the last stage outputs
+    // from the hourglass. Doing so halves the memory requirement of the
+    // mask head and does not cause a drop in evaluation metrics.
+    optional bool use_only_last_stage = 24 [default=false];
+
   }
 
   optional DeepMACMaskEstimation deepmac_mask_estimation = 14;

research/seq_flow_lite/models/sgnn/BUILD

@@ -90,6 +90,7 @@ py_binary(
     main = "run_tflite.py",
     python_version = "PY3",
     deps = [
+        ":sgnn_projection_op_resolver",
         # Expect numpy installed
         # package TFLite flex delegate
         # package TFLite interpreter

research/seq_flow_lite/utils/tflite_utils.py

@@ -80,6 +80,7 @@ def set_output_quantized_for_custom_ops(graph_def, use_mlir=True):
       'ExpectedValueOp': [tf.float32.as_datatype_enum],
       'LayerNorm': [tf.float32.as_datatype_enum],
       'UniformCausalAttn': [tf.float32.as_datatype_enum],
+      'DynamicUniformCausalAttn': [tf.float32.as_datatype_enum],
       'RnnDecoderReadState': [tf.float32.as_datatype_enum],
       'RnnDecoderWriteState': [tf.float32.as_datatype_enum],
   }