Refacto architecture outputs orders

tests/conftest.py

@@ -14,13 +14,13 @@ def cspdarknet53_416():
 
 @pytest.fixture(scope="session")
 def yolov4_neck_416():
-    input_shapes = [(13, 13, 1024), (26, 26, 512), (52, 52, 256)]
+    input_shapes = [(52, 52, 256), (26, 26, 512), (13, 13, 1024)]
     return yolov4_neck(input_shapes)
 
 
 @pytest.fixture(scope="session")
 def yolov3_head_416_training(num_classes, yolo_max_boxes):
-    input_shapes = [(13, 13, 512), (26, 26, 256), (52, 52, 128)]
+    input_shapes = [(52, 52, 128), (26, 26, 256), (13, 13, 512)]
 
     return yolov3_head(
         input_shapes,
@@ -35,7 +35,7 @@ def yolov3_head_416_training(num_classes, yolo_max_boxes):
 
 @pytest.fixture(scope="session")
 def yolov3_head_416_inference(num_classes, yolo_max_boxes):
-    input_shapes = [(13, 13, 512), (26, 26, 256), (52, 52, 128)]
+    input_shapes = [(52, 52, 128), (26, 26, 256), (13, 13, 512)]
 
     return yolov3_head(
         input_shapes,

tests/test_csp_darknet53.py

@@ -1,8 +1,8 @@
 def test_backbone_should_have_valid_output_shapes(cspdarknet53_416):
     output_1, output_2, output_3 = cspdarknet53_416.outputs
-    assert output_1.shape.as_list() == [None, 13, 13, 1024]
+    assert output_1.shape.as_list() == [None, 52, 52, 256]
     assert output_2.shape.as_list() == [None, 26, 26, 512]
-    assert output_3.shape.as_list() == [None, 52, 52, 256]
+    assert output_3.shape.as_list() == [None, 13, 13, 1024]
 
 
 def test_backbone_should_have_the_right_number_of_parameters(cspdarknet53_416):

tests/test_model.py

@@ -15,9 +15,9 @@ def test_model_should_predict_valid_shapes_at_training(yolov4_training, num_clas
         tf.random.uniform((n_images, 416, 416, 3))
     )
 
-    assert output_1.shape == (n_images, 13, 13, 3, expected_head_shape)
+    assert output_1.shape == (n_images, 52, 52, 3, expected_head_shape)
     assert output_2.shape == (n_images, 26, 26, 3, expected_head_shape)
-    assert output_3.shape == (n_images, 52, 52, 3, expected_head_shape)
+    assert output_3.shape == (n_images, 13, 13, 3, expected_head_shape)
 
 
 def test_model_should_predict_valid_shapes_at_inference(

tests/test_yolov3_head.py

@@ -6,9 +6,9 @@ def test_head_should_have_valid_output_shapes_training(
     expected_head_shape = (num_classes + objectness_score_shape) + bounding_box_shape
 
     output_1, output_2, output_3 = yolov3_head_416_training.outputs
-    assert output_1.shape.as_list() == [None, 13, 13, 3, expected_head_shape]
+    assert output_1.shape.as_list() == [None, 52, 52, 3, expected_head_shape]
     assert output_2.shape.as_list() == [None, 26, 26, 3, expected_head_shape]
-    assert output_3.shape.as_list() == [None, 52, 52, 3, expected_head_shape]
+    assert output_3.shape.as_list() == [None, 13, 13, 3, expected_head_shape]
 
 
 def test_head_should_have_valid_output_shapes_inference(

tests/test_yolov4_neck.py

@@ -1,5 +1,5 @@
 def test_neck_should_have_valid_output_shapes(yolov4_neck_416):
     output_1, output_2, output_3 = yolov4_neck_416.outputs
-    assert output_1.shape.as_list() == [None, 13, 13, 512]
+    assert output_1.shape.as_list() == [None, 52, 52, 128]
     assert output_2.shape.as_list() == [None, 26, 26, 256]
-    assert output_3.shape.as_list() == [None, 52, 52, 128]
+    assert output_3.shape.as_list() == [None, 13, 13, 512]

tf2_yolov4/anchors.py

@@ -4,15 +4,15 @@
 import numpy as np
 
 YOLOV4_ANCHORS = [
-    np.array([(142, 110), (192, 243), (459, 401)], np.float32),
-    np.array([(36, 75), (76, 55), (72, 146)], np.float32),
     np.array([(12, 16), (19, 36), (40, 28)], np.float32),
+    np.array([(36, 75), (76, 55), (72, 146)], np.float32),
+    np.array([(142, 110), (192, 243), (459, 401)], np.float32),
 ]
 
 YOLOV3_ANCHORS = [
-    np.array([(116, 90), (156, 198), (373, 326)], np.float32),
-    np.array([(30, 61), (62, 45), (59, 119)], np.float32),
     np.array([(10, 13), (16, 30), (33, 23)], np.float32),
+    np.array([(30, 61), (62, 45), (59, 119)], np.float32),
+    np.array([(116, 90), (156, 198), (373, 326)], np.float32),
 ]
 
 

tf2_yolov4/backbones/csp_darknet53.py

@@ -111,7 +111,7 @@ def csp_darknet53(input_shape):
     # Fifth downsampling: L616 -> L744
     output_3 = csp_block(output_2, filters=1024, num_blocks=4)
 
-    return tf.keras.Model(inputs, [output_3, output_2, output_1], name="CSPDarknet53")
+    return tf.keras.Model(inputs, [output_1, output_2, output_3], name="CSPDarknet53")
 
 
 if __name__ == "__main__":

tf2_yolov4/heads/yolov3_head.py

@@ -42,13 +42,13 @@ def yolov3_head(
     input_2 = tf.keras.Input(shape=filter(None, input_shapes[1]))
     input_3 = tf.keras.Input(shape=filter(None, input_shapes[2]))
 
-    x = conv_bn(input_3, filters=256, kernel_size=3, strides=1, activation="leaky_relu")
-    output_3 = conv_classes_anchors(
+    x = conv_bn(input_1, filters=256, kernel_size=3, strides=1, activation="leaky_relu")
+    output_1 = conv_classes_anchors(
         x, num_anchors_stage=len(anchors[0]), num_classes=num_classes
     )
 
     x = conv_bn(
-        input_3,
+        input_1,
         filters=256,
         kernel_size=3,
         strides=2,
@@ -80,14 +80,14 @@ def yolov3_head(
         padding="valid",
         activation="leaky_relu",
     )
-    x = tf.keras.layers.Concatenate()([x, input_1])
+    x = tf.keras.layers.Concatenate()([x, input_3])
     x = conv_bn(x, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=1024, kernel_size=3, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=1024, kernel_size=3, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=1024, kernel_size=3, strides=1, activation="leaky_relu")
-    output_1 = conv_classes_anchors(
+    output_3 = conv_classes_anchors(
         x, num_anchors_stage=len(anchors[2]), num_classes=num_classes
     )
 

tf2_yolov4/necks/yolov4_neck.py

@@ -19,7 +19,7 @@ def yolov4_neck(input_shapes):
     input_2 = tf.keras.Input(shape=filter(None, input_shapes[1]))
     input_3 = tf.keras.Input(shape=filter(None, input_shapes[2]))
 
-    x = conv_bn(input_1, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
+    x = conv_bn(input_3, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=1024, kernel_size=3, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
 
@@ -31,11 +31,11 @@ def yolov4_neck(input_shapes):
 
     x = conv_bn(spp, filters=512, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=1024, kernel_size=3, strides=1, activation="leaky_relu")
-    output_1 = conv_bn(
+    output_3 = conv_bn(
         x, filters=512, kernel_size=1, strides=1, activation="leaky_relu"
     )
     x = conv_bn(
-        output_1, filters=256, kernel_size=1, strides=1, activation="leaky_relu"
+        output_3, filters=256, kernel_size=1, strides=1, activation="leaky_relu"
     )
 
     upsampled = tf.keras.layers.UpSampling2D()(x)
@@ -56,14 +56,14 @@ def yolov4_neck(input_shapes):
 
     upsampled = tf.keras.layers.UpSampling2D()(x)
 
-    x = conv_bn(input_3, filters=128, kernel_size=1, strides=1, activation="leaky_relu")
+    x = conv_bn(input_1, filters=128, kernel_size=1, strides=1, activation="leaky_relu")
     x = tf.keras.layers.Concatenate()([x, upsampled])
 
     x = conv_bn(x, filters=128, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=256, kernel_size=3, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=128, kernel_size=1, strides=1, activation="leaky_relu")
     x = conv_bn(x, filters=256, kernel_size=3, strides=1, activation="leaky_relu")
-    output_3 = conv_bn(
+    output_1 = conv_bn(
         x, filters=128, kernel_size=1, strides=1, activation="leaky_relu"
     )