datapipeline fixes

official/vision/beta/projects/yolo/dataloaders/yolo_input.py

@@ -147,6 +147,7 @@ class Parser(parser.Parser):
     # Set the per level values needed for operation
     self._darknet = darknet
     self._area_thresh = area_thresh
+    self._level_limits = level_limits
 
     self._seed = seed
     self._dtype = dtype
@@ -259,7 +260,7 @@ class Parser(parser.Parser):
         self._aug_rand_saturation,
         self._aug_rand_brightness,
         seed=self._seed,
-        darknet=self._darknet)
+        darknet=self._darknet or self._level_limits is not None)
 
     # Cast the image to the selcted datatype.
     image, labels = self._build_label(

official/vision/beta/projects/yolo/ops/mosaic.py

@@ -14,7 +14,6 @@
 
 """Mosaic op."""
 import random
-
 import tensorflow as tf
 import tensorflow_addons as tfa
 
@@ -22,7 +21,6 @@ from official.vision.beta.ops import box_ops
 from official.vision.beta.ops import preprocess_ops
 from official.vision.beta.projects.yolo.ops import preprocessing_ops
 
-
 class Mosaic:
   """Stitch together sets of 4 images to generate samples with more boxes."""
 
@@ -211,6 +209,7 @@ class Mosaic:
       boxes, classes, is_crowd, area = self._select_ind(inds, boxes, classes,  # pylint:disable=unbalanced-tuple-unpacking
                                                         is_crowd, area)
 
+
     # warp and scale the fully stitched sample
     image, _, affine = preprocessing_ops.affine_warp_image(
         image, [self._output_size[0], self._output_size[1]],
@@ -325,6 +324,11 @@ class Mosaic:
     else:
       return self._add_param(noop)
 
+  def _beta(self, alpha, beta):
+    a = tf.random.gamma([], alpha)
+    b = tf.random.gamma([], beta)
+    return b / (a + b)
+
   def _mixup(self, one, two):
     """Blend together 2 images for the mixup data augmentation."""
     if self._mixup_frequency >= 1.0:
@@ -337,8 +341,8 @@ class Mosaic:
     if domo >= (1 - self._mixup_frequency):
       sample = one
       otype = one['image'].dtype
-      r = preprocessing_ops.random_uniform_strong(
-          0.4, 0.6, tf.float32, seed=self._seed)
+
+      r = self._beta(8.0, 8.0)
       sample['image'] = (
           r * tf.cast(one['image'], tf.float32) +
           (1 - r) * tf.cast(two['image'], tf.float32))

official/vision/beta/projects/yolo/ops/preprocessing_ops.py

@@ -170,14 +170,14 @@ def get_image_shape(image):
 def _augment_hsv_darknet(image, rh, rs, rv, seed=None):
   """Randomize the hue, saturation, and brightness via the darknet method."""
   if rh > 0.0:
-    delta = random_uniform_strong(-rh, rh, seed=seed)
-    image = tf.image.adjust_hue(image, delta)
+    deltah = random_uniform_strong(-rh, rh, seed=seed)
+    image = tf.image.adjust_hue(image, deltah)
   if rs > 0.0:
-    delta = random_scale(rs, seed=seed)
-    image = tf.image.adjust_saturation(image, delta)
+    deltas = random_scale(rs, seed=seed)
+    image = tf.image.adjust_saturation(image, deltas)
   if rv > 0.0:
-    delta = random_scale(rv, seed=seed)
-    image *= delta
+    deltav = random_scale(rv, seed=seed)
+    image *= tf.cast(deltav, image.dtype)
 
   # clip the values of the image between 0.0 and 1.0
   image = tf.clip_by_value(image, 0.0, 1.0)
@@ -719,7 +719,7 @@ def affine_warp_boxes(affine, boxes, output_size, box_history):
     return tf.stack([y_min, x_min, y_max, x_max], axis=-1)
 
   def _aug_boxes(affine_matrix, box):
-    """Apply an affine transformation matrix M to the boxes augmente boxes."""
+    """Apply an affine transformation matrix M to the boxes augment boxes."""
     corners = _get_corners(box)
     corners = tf.reshape(corners, [-1, 4, 2])
     z = tf.expand_dims(tf.ones_like(corners[..., 1]), axis=-1)