Internal change

PiperOrigin-RevId: 525585347

official/projects/yolo/README.md

@@ -90,6 +90,7 @@ connected to a new, more powerful backbone if a person chose to.
 | Yolo-v4 csp      |
 | Yolo-v4 large    |
 | Yolo-v7          |
+| Yolo-v7-tiny     |
 
 ## Requirements
 [![TensorFlow 2.11](https://img.shields.io/badge/TensorFlow-2.11-FF6F00?logo=tensorflow)](https://github.com/tensorflow/tensorflow/releases/tag/v2.11.0)

official/projects/yolo/configs/yolov7.py

@@ -309,3 +309,33 @@ def coco_yolov7() -> cfg.ExperimentConfig:
   )
 
   return config
+
+
+@exp_factory.register_config_factory('coco_yolov7tiny')
+def coco_yolov7_tiny() -> cfg.ExperimentConfig:
+  """COCO object detection with YOLOv7-tiny."""
+  config = coco_yolov7()
+  config.task.model.input_size = [416, 416, 3]
+  config.task.model.backbone.yolov7.model_id = 'yolov7-tiny'
+  config.task.model.decoder.yolov7.model_id = 'yolov7-tiny'
+  config.task.model.norm_activation.activation = 'leaky'
+  config.task.model.anchor_boxes.boxes = [
+      Box(box=[10, 13]),
+      Box(box=[16, 30]),
+      Box(box=[33, 23]),
+      Box(box=[30, 61]),
+      Box(box=[62, 45]),
+      Box(box=[59, 119]),
+      Box(box=[116, 90]),
+      Box(box=[156, 198]),
+      Box(box=[373, 326]),
+  ]
+
+  config.task.model.loss.cls_weight = 0.5
+  config.task.model.loss.obj_weight = 1.0
+  config.task.train_data.parser.aug_rand_translate = 0.1
+  config.task.train_data.parser.mosaic.mixup_frequency = 0.05
+  config.task.train_data.parser.mosaic.aug_scale_min = 0.5
+  config.task.train_data.parser.mosaic.aug_scale_max = 1.5
+  config.trainer.optimizer_config.learning_rate.cosine.alpha = 0.01
+  return config

official/projects/yolo/modeling/backbones/yolov7.py

@@ -68,6 +68,44 @@ _BLOCK_SPEC_SCHEMAS = {
     ]
 }
 
+# Define YOLOv7-tiny variant.
+_YoloV7Tiny = [
+    ['convbn', -1, 3, 2, 32, False],  # 0-P1/2
+
+    ['convbn', -1, 3, 2, 64, False],  # 1-P2/4
+
+    ['convbn', -1, 1, 1, 32, False],
+    ['convbn', -2, 1, 1, 32, False],
+    ['convbn', -1, 3, 1, 32, False],
+    ['convbn', -1, 3, 1, 32, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 64, False],  # 7
+
+    ['maxpool2d', -1, 2, 2, 'same', False],  # 8-P3/8
+    ['convbn', -1, 1, 1, 64, False],
+    ['convbn', -2, 1, 1, 64, False],
+    ['convbn', -1, 3, 1, 64, False],
+    ['convbn', -1, 3, 1, 64, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 128, True],  # 14
+
+    ['maxpool2d', -1, 2, 2, 'same', False],  # 15-P4/16
+    ['convbn', -1, 1, 1, 128, False],
+    ['convbn', -2, 1, 1, 128, False],
+    ['convbn', -1, 3, 1, 128, False],
+    ['convbn', -1, 3, 1, 128, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 256, True],  # 21
+
+    ['maxpool2d', -1, 2, 2, 'same', False],  # 22-P5/32
+    ['convbn', -1, 1, 1, 256, False],
+    ['convbn', -2, 1, 1, 256, False],
+    ['convbn', -1, 3, 1, 256, False],
+    ['convbn', -1, 3, 1, 256, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 512, True],  # 28
+]
+
 # Define YOLOv7 variant.
 _YoloV7 = [
     ['convbn', -1, 3, 1, 32, False],  # 0
@@ -133,6 +171,7 @@ _YoloV7 = [
 
 # Aggregates all variants for YOLOv7 backbones.
 BACKBONES = {
+    'yolov7-tiny': _YoloV7Tiny,
     'yolov7': _YoloV7,
 }
 

official/projects/yolo/modeling/decoders/yolov7.py

@@ -91,6 +91,69 @@ _BLOCK_SPEC_SCHEMAS = {
     ],
 }
 
+# Define specs for YOLOv7-tiny variant. It is recommended to use together with
+# YOLOv7-tiny backbone.
+_YoloV7Tiny = [
+    ['convbn', -1, 1, 1, 256, False],
+    ['convbn', -2, 1, 1, 256, False],
+    ['maxpool2d', -1, 5, 1, 'same', False],
+    ['maxpool2d', -2, 9, 1, 'same', False],
+    ['maxpool2d', -3, 13, 1, 'same', False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 256, False],
+    ['concat', [-1, -7], -1, False],
+    ['convbn', -1, 1, 1, 256, False],  # 8
+
+    ['convbn', -1, 1, 1, 128, False],
+    ['upsample2d', -1, 2, 'nearest', False],
+    ['convbn', '4', 1, 1, 128, False],  # route from backbone P4
+    ['concat', [-1, -2], -1, False],
+
+    ['convbn', -1, 1, 1, 64, False],
+    ['convbn', -2, 1, 1, 64, False],
+    ['convbn', -1, 3, 1, 64, False],
+    ['convbn', -1, 3, 1, 64, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 128, False],  # 18
+
+    ['convbn', -1, 1, 1, 64, False],
+    ['upsample2d', -1, 2, 'nearest', False],
+    ['convbn', '3', 1, 1, 64, False],  # route from backbone P3
+    ['concat', [-1, -2], -1, False],
+
+    ['convbn', -1, 1, 1, 32, False],
+    ['convbn', -2, 1, 1, 32, False],
+    ['convbn', -1, 3, 1, 32, False],
+    ['convbn', -1, 3, 1, 32, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 64, False],  # 28
+
+    ['convbn', -1, 3, 2, 128, False],
+    ['concat', [-1, 18], -1, False],
+
+    ['convbn', -1, 1, 1, 64, False],
+    ['convbn', -2, 1, 1, 64, False],
+    ['convbn', -1, 3, 1, 64, False],
+    ['convbn', -1, 3, 1, 64, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 128, False],  # 36
+
+    ['convbn', -1, 3, 2, 256, False],
+    ['concat', [-1, 8], -1, False],
+
+    ['convbn', -1, 1, 1, 128, False],
+    ['convbn', -2, 1, 1, 128, False],
+    ['convbn', -1, 3, 1, 128, False],
+    ['convbn', -1, 3, 1, 128, False],
+    ['concat', [-1, -2, -3, -4], -1, False],
+    ['convbn', -1, 1, 1, 256, False],  # 44
+
+    ['convbn', 28, 1, 1, 128, True],
+    ['convbn', 36, 1, 1, 256, True],
+    ['convbn', 44, 1, 1, 512, True],
+]
+
+
 # Define specs YOLOv7 variant. The spec schema is defined above.
 # It is recommended to use together with YOLOv7 backbone.
 _YoloV7 = [
@@ -161,6 +224,7 @@ _YoloV7 = [
 
 # Aggregates all variants for YOLOv7 decoders.
 DECODERS = {
+    'yolov7-tiny': _YoloV7Tiny,
     'yolov7': _YoloV7,
 }
 
@@ -257,13 +321,13 @@ class YoloV7(tf.keras.Model):
     return inputs
 
   def _group_layer_inputs(self, from_index, inputs, outputs):
-    if not outputs:
-      return inputs[max(inputs.keys())]
-
     if isinstance(from_index, list):
       return [self._group_layer_inputs(i, inputs, outputs) for i in from_index]
 
     if isinstance(from_index, int):
+      # Need last layer output from backbone.
+      if len(outputs) + from_index == -1:
+        return inputs[max(inputs.keys())]
       return outputs[from_index]
     return inputs[from_index]  # from_index is a string.