tfds csp

official/vision/beta/projects/yolo/configs/experiments/csp_darknet53_tfds.yaml

 runtime:
   distribution_strategy: 'mirrored'
   mixed_precision_dtype: 'float16'
+  loss_scale: 'dynamic'
+  num_gpus: 2
 task:
   model:
     num_classes: 1001
@@ -12,28 +14,29 @@ task:
   losses:
     l2_weight_decay: 0.0005
     one_hot: True
-    label_smoothing: 0.1
   train_data:
     tfds_name: 'imagenet2012'
     tfds_split: 'train'
     tfds_data_dir: '~/tensorflow_datasets/'
     tfds_download: true
     is_training: true
-    global_batch_size: 128
+    global_batch_size: 16 #default = 128
     dtype: 'float16'
+    shuffle_buffer_size: 100
   validation_data:
     tfds_name: 'imagenet2012'
     tfds_split: 'validation'
     tfds_data_dir: '~/tensorflow_datasets/'
     tfds_download: true
     is_training: true
-    global_batch_size: 128
+    global_batch_size: 16 #default = 128
     dtype: 'float16'
     drop_remainder: false
+    shuffle_buffer_size: 100
 trainer:
-  train_steps: 1200000 # epochs: 120
-  validation_steps: 400 # size of validation data
-  validation_interval: 500 # 10000
+  train_steps: 9600000 # epochs: 120, 1200000 * 128/batchsize
+  validation_steps: 3200 # size of validation data, 400 * 128/batchsize
+  validation_interval: 10000 #10000
   steps_per_loop: 10000
   summary_interval: 10000
   checkpoint_interval: 10000
@@ -45,11 +48,11 @@ trainer:
     learning_rate:
       type: 'polynomial'
       polynomial:
-        initial_learning_rate: 0.1
-        end_learning_rate: 0.0001
+        initial_learning_rate: 0.0125 # 0.1 * batchsize/128, default = 0.1
+        end_learning_rate: 0.0000125 # 0.0001 * batchsize/128, default = 0.0001
         power: 4.0
-        decay_steps: 1190000
+        decay_steps: 9592000 # 790000 * 128/batchsize,   default =  800000 - 1000 = 799000
     warmup:
       type: 'linear'
       linear:
-        warmup_steps: 1000 #learning rate rises from 0 to 0.1 over 1000 steps
+        warmup_steps: 8000 # 0 to 0.1 over 1000 * 128/batchsize, default = 128

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

@@ -20,9 +20,6 @@ from official.vision.beta.dataloaders import decoder
 from official.vision.beta.dataloaders import parser
 from official.vision.beta.ops import preprocess_ops
 
-MEAN_RGB = (0.485 * 255, 0.456 * 255, 0.406 * 255)
-STDDEV_RGB = (0.229 * 255, 0.224 * 255, 0.225 * 255)
-
 
 class Decoder(decoder.Decoder):
   """A tf.Example decoder for classification task."""