"BackBone"

official/vision/beta/configs/backbones.py

@@ -64,6 +64,7 @@ class RevNet(hyperparams.Config):
   # Specifies the depth of RevNet.
   model_id: int = 56
 
+from official.vision.beta.projects.yolo.configs.backbones import DarkNet
 
 @dataclasses.dataclass
 class Backbone(hyperparams.OneOfConfig):
@@ -85,3 +86,4 @@ class Backbone(hyperparams.OneOfConfig):
   efficientnet: EfficientNet = EfficientNet()
   spinenet: SpineNet = SpineNet()
   mobilenet: MobileNet = MobileNet()
+  darknet: DarkNet = DarkNet()

official/vision/beta/dataloaders/classification_input.py

@@ -33,11 +33,13 @@ class Decoder(decoder.Decoder):
         'image/class/label': (
             tf.io.FixedLenFeature((), tf.int64, default_value=-1))
     }
-
+  '''
   def decode(self, serialized_example):
     return tf.io.parse_single_example(
         serialized_example, self._keys_to_features)
-
+  '''
+  def decode(self, data):
+      return {'image/encoded': data['image'], 'image/class/label': data['label']}
 
 class Parser(parser.Parser):
   """Parser to parse an image and its annotations into a dictionary of tensors."""
@@ -72,11 +74,11 @@ class Parser(parser.Parser):
 
   def _parse_train_data(self, decoded_tensors):
     """Parses data for training."""
+    
     label = tf.cast(decoded_tensors['image/class/label'], dtype=tf.int32)
-
+    '''
     image_bytes = decoded_tensors['image/encoded']
     image_shape = tf.image.extract_jpeg_shape(image_bytes)
-
     # Crops image.
     # TODO(pengchong): support image format other than JPEG.
     cropped_image = preprocess_ops.random_crop_image_v2(
@@ -85,7 +87,8 @@ class Parser(parser.Parser):
         tf.reduce_all(tf.equal(tf.shape(cropped_image), image_shape)),
         lambda: preprocess_ops.center_crop_image_v2(image_bytes, image_shape),
         lambda: cropped_image)
-
+    '''
+    image = tf.cast(decoded_tensors['image/encoded'], tf.float32)
     if self._aug_rand_hflip:
       image = tf.image.random_flip_left_right(image)
 
@@ -106,12 +109,14 @@ class Parser(parser.Parser):
   def _parse_eval_data(self, decoded_tensors):
     """Parses data for evaluation."""
     label = tf.cast(decoded_tensors['image/class/label'], dtype=tf.int32)
+    '''
     image_bytes = decoded_tensors['image/encoded']
     image_shape = tf.image.extract_jpeg_shape(image_bytes)
-
+    
     # Center crops and resizes image.
     image = preprocess_ops.center_crop_image_v2(image_bytes, image_shape)
-
+    '''
+    image = tf.cast(decoded_tensors['image/encoded'], tf.float32)
     image = tf.image.resize(
         image, self._output_size, method=tf.image.ResizeMethod.BILINEAR)
 

official/vision/beta/projects/yolo/common/registry_imports.py

@@ -20,3 +20,4 @@ from official.utils.testing import mock_task
 from official.vision import beta
 
 from official.vision.beta.projects import yolo
+from official.vision.beta.projects.yolo.modeling.backbones import Darknet

official/vision/beta/projects/yolo/configs/backbones.py

@@ -4,7 +4,7 @@ import dataclasses
 from typing import Optional
 from official.modeling import hyperparams
 
-from official.vision.beta.configs import backbones
+# from official.vision.beta.configs import backbones
 
 
 @dataclasses.dataclass
@@ -13,6 +13,6 @@ class DarkNet(hyperparams.Config):
   model_id: str = "darknet53"
 
 
-@dataclasses.dataclass
-class Backbone(backbones.Backbone):
-  darknet: DarkNet = DarkNet()
+# @dataclasses.dataclass
+# class Backbone(backbones.Backbone):
+#   darknet: DarkNet = DarkNet()

official/vision/beta/projects/yolo/configs/experiments/darknet53.yaml

@@ -19,23 +19,23 @@ task:
     tfds_split: 'test'
     tfds_download: True
     is_training: True
-    global_batch_size: 128
+    global_batch_size: 2
     dtype: 'float16'
   validation_data:
     tfds_name: 'imagenet_a'
     tfds_split: 'test'
     tfds_download: True
     is_training: False
-    global_batch_size: 128
+    global_batch_size: 2
     dtype: 'float16'
     drop_remainder: False
 trainer:
-  train_steps: 800000 # in the paper
-  validation_steps: 400 # size of validation data
-  validation_interval: 10000 
-  steps_per_loop: 10000
-  summary_interval: 10000
-  checkpoint_interval: 10000
+  train_steps: 51200000 # in the paper
+  validation_steps: 25600 # size of validation data
+  validation_interval: 150 
+  steps_per_loop: 150
+  summary_interval: 150
+  checkpoint_interval: 150
   optimizer_config:
     optimizer:
       type: 'sgd'
@@ -47,8 +47,8 @@ trainer:
         initial_learning_rate: 0.1
         end_learning_rate: 0.0001
         power: 4.0 
-        decay_steps: 799000
+        decay_steps: 51136000
     warmup:
       type: 'linear'
       linear:
-        warmup_steps: 1000 #lr rise from 0 to 0.1 over 1000 steps
+        warmup_steps: 64000 #lr rise from 0 to 0.1 over 1000 steps

official/vision/beta/projects/yolo/modeling/backbones/Darknet.py

@@ -157,6 +157,10 @@ BACKBONES = {
     "cspdarknettiny": CSPDARKNETTINY
 }
 
+CSPBlockConfig = collections.namedtuple('CSPBlockConfig', ['layer', 'stack', 'repetitions', 'bottleneck',
+  'filters', 'kernel_size', 'strides', 'padding', 'activation', 'route',
+  'output_name', 'is_output'])
+
 
 @ks.utils.register_keras_serializable(package='yolo')
 class Darknet(ks.Model):
@@ -177,7 +181,11 @@ class Darknet(ks.Model):
       config=None,
       **kwargs):
 
-    layer_specs, splits = Darknet.get_config(model_id)
+    # layer_specs, splits = Darknet.get_config(model_id)
+    print(len(BACKBONES[model_id]['backbone'][-1]))
+    layer_specs = [CSPBlockConfig(*config) for config in BACKBONES[model_id]['backbone']]
+    splits = BACKBONES[model_id]['splits']
+
     self._model_name = model_id
     self._splits = splits
     self._input_shape = input_shape

official/vision/beta/projects/yolo/modeling/building_blocks/_DarkConv.py

@@ -6,7 +6,7 @@ import tensorflow.keras as ks
 import tensorflow.keras.backend as K
 from ._Identity import Identity
 
-from yolo.modeling.functions.mish_activation import mish
+from official.vision.beta.projects.yolo.modeling.functions.mish_activation import mish
 
 
 @ks.utils.register_keras_serializable(package='yolo')

official/vision/beta/projects/yolo/modeling/building_blocks/_Identity.py

+"""Contains common building blocks for yolo neural networks."""
+import tensorflow as tf
+import tensorflow.keras as ks
+
+
+@ks.utils.register_keras_serializable(package='yolo')
+class Identity(ks.layers.Layer):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        return
+
+    def call(self, input):
+        return input
+