support to the mobilnetv2-yolo structure of r1.13

train.py

@@ -4,7 +4,7 @@
 import tensorflow as tf
 import datetime
 import zipfile
-from yolo3.model import darknet_yolo_body, YoloLoss, mobilenetv2_yolo_body, efficientnet_yolo_body
+from yolo3.model import darknet_yolo_body, YoloLoss, mobilenetv2_yolo_body, efficientnet_yolo_body, mobilenetv2_yolo_body_r13
 from yolo3.data import Dataset
 from yolo3.enum import OPT, BACKBONE, DATASET_MODE
 from yolo3.map import MAPCallback
@@ -115,6 +115,12 @@ def train(FLAGS):
                                   len(anchors) // 3,
                                   num_classes,
                                   alpha=FLAGS['alpha'])
+        elif backbone == BACKBONE.MOBILENETV2_R13:
+            model = factory.build(mobilenetv2_yolo_body_r13,
+                                  155,
+                                  len(anchors) // 3,
+                                  num_classes,
+                                  alpha=FLAGS['alpha'])                                  
         elif backbone == BACKBONE.DARKNET53:
             model = factory.build(darknet_yolo_body, 185,
                                   len(anchors) // 3, num_classes)

train_backbone.py

@@ -76,7 +76,7 @@ def train(FLAGS):
     batch_size = batch_size * strategy.num_replicas_in_sync
     with strategy.scope():
         factory = ModelFactory(weights_path=model_path)
-        if backbone == BACKBONE.MOBILENETV2:
+        if backbone == BACKBONE.MOBILENETV2 or backbone == BACKBONE.MOBILENETV2_R13:
             model = factory.build(mobilenetv2,
                                   0,
                                   alpha=1.4,

yolo.py

@@ -9,7 +9,7 @@
 from PIL import Image, ImageFont, ImageDraw
 import cv2
 import tensorflow as tf
-from yolo3.model import yolo_eval, darknet_yolo_body, mobilenetv2_yolo_body, efficientnet_yolo_body, YoloEval
+from yolo3.model import yolo_eval, darknet_yolo_body, mobilenetv2_yolo_body, efficientnet_yolo_body, YoloEval, mobilenetv2_yolo_body_r13
 from yolo3.utils import letterbox_image, get_anchors, get_classes
 from yolo3.enum import OPT, BACKBONE
 from yolo3.map import MAPCallback
@@ -79,6 +79,9 @@ def generate(self, FLAGS):
             if self.backbone == BACKBONE.MOBILENETV2:
                 model_body = partial(mobilenetv2_yolo_body,
                                      alpha=FLAGS['alpha'])
+            elif self.backbone == BACKBONE.MOBILENETV2_R13:
+                model_body = partial(mobilenetv2_yolo_body_r13,
+                                     alpha=FLAGS['alpha'])                                     
             elif self.backbone == BACKBONE.DARKNET53:
                 model_body = darknet_yolo_body
             elif self.backbone == BACKBONE.EFFICIENTNET:

yolo3/enum.py

@@ -26,6 +26,7 @@ class BACKBONE(Enum):
     MOBILENETV2 = 0
     EFFICIENTNET = 1
     DARKNET53 = 2
+    MOBILENETV2_R13 = 3
 
 
 @unique

yolo3/model.py

@@ -195,6 +195,48 @@ def mobilenetv2_yolo_body(inputs, num_anchors, num_classes, alpha=1.0):
     return tf.keras.models.Model(inputs, [y1, y2, y3])
 
 
+# the mobkenetv2-yolo model befroe r1.14
+def mobilenetv2_yolo_body_r13(inputs, num_anchors, num_classes, alpha=1.0):
+    MobilenetConv2D_BN_Relu = MobilenetConv2D
+    mobilenetv2 = tf.keras.applications.MobileNetV2(alpha=alpha, input_tensor=inputs, include_top=False,
+                                                    weights='imagenet')
+    x=mobilenetv2.output
+    y1 = MobilenetConv2D_BN_Relu((1,1),alpha, 1280)(x)
+    y1 = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),padding='same',kernel_regularizer=tf.keras.regularizers.l2(5e-4))(y1)
+    x = compose(
+        MobilenetConv2D_BN_Relu((1,1),alpha, 640),
+        tf.keras.layers.UpSampling2D(2))(x)
+    x = tf.keras.layers.Concatenate()(
+        [x, MobilenetConv2D_BN_Relu((1,1),alpha, 640)(mobilenetv2.get_layer('block_12_project_BN').output)])
+    y2 = MobilenetConv2D_BN_Relu((1,1),alpha, 640)(x)
+    y2 = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),padding='same',kernel_regularizer=tf.keras.regularizers.l2(5e-4))(y2)
+    x = compose(
+        MobilenetConv2D_BN_Relu((1,1),alpha, 320),
+        tf.keras.layers.UpSampling2D(2))(x)
+    x = tf.keras.layers.Concatenate()(
+        [x, MobilenetConv2D_BN_Relu((1,1),alpha, 320)(mobilenetv2.get_layer('block_5_project_BN').output)])
+    y3 = MobilenetConv2D_BN_Relu((1,1),alpha, 320)(x)
+    y3 = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1),padding='same',kernel_regularizer=tf.keras.regularizers.l2(5e-4))(y3)
+
+    # reshape to make compatible with r1.14
+    y1 = tf.keras.layers.Lambda(lambda y: tf.reshape(y, [
+                                                        -1, tf.shape(y)[1],
+                                                        tf.shape(y)[2], num_anchors, num_classes + 5
+                                                    ]),
+                                name='y1')(y1)
+    y2 = tf.keras.layers.Lambda(lambda y: tf.reshape(y, [
+                                                        -1, tf.shape(y)[1],
+                                                        tf.shape(y)[2], num_anchors, num_classes + 5
+                                                    ]),
+                                name='y2')(y2)
+    y3 = tf.keras.layers.Lambda(lambda y: tf.reshape(y, [
+                                                        -1, tf.shape(y)[1],
+                                                        tf.shape(y)[2], num_anchors, num_classes + 5
+                                                    ]),
+                                name='y3')(y3)
+    return tf.keras.models.Model(inputs, [y1, y2, y3])
+
+
 def make_last_layers_efficientnet(x, block_args, global_params):
     if global_params.data_format == 'channels_first':
         channel_axis = 1