modify default hyperparameters

official/vision/beta/configs/experiments/image_classification/imagenet_mobilenetv2_1.0_gpu.yaml

@@ -13,7 +13,7 @@ task:
         width_multiplier: 1.0
     dropout_rate: 0.2
   losses:
-    l2_weight_decay: 0.0001
+    l2_weight_decay: 0.00002
     one_hot: True
     label_smoothing: 0.1
   train_data:
@@ -28,7 +28,7 @@ task:
     dtype: 'float32'
     drop_remainder: False
 trainer:
-  train_steps: 150120  # 90 * steps_per_epoch
+  train_steps: 700000  # 700K
   validation_steps: 65
   validation_interval: 1668
   steps_per_loop: 1668  # NUM_EXAMPLES (1281167) // global_batch_size
@@ -46,7 +46,7 @@ trainer:
       exponential:
         initial_learning_rate: 0.36  # 0.045 * NUM_GPUS
         decay_steps: 4170  # 2.5 * steps_per_epoch
-        decay_rate: 0.97
+        decay_rate: 0.98
         staircase: True
     warmup:
       type: 'linear'

official/vision/beta/configs/image_classification.py

@@ -233,7 +233,7 @@ def image_classification_imagenet_mobilenet() -> cfg.ExperimentConfig:
                       model_id='MobileNetV2', width_multiplier=1.0)),
               norm_activation=common.NormActivation(
                   norm_momentum=0.9997, norm_epsilon=1e-3)),
-          losses=Losses(l2_weight_decay=1e-4, label_smoothing=0.1),
+          losses=Losses(l2_weight_decay=2e-5, label_smoothing=0.1),
           train_data=DataConfig(
               input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
               is_training=True,
@@ -264,8 +264,7 @@ def image_classification_imagenet_mobilenet() -> cfg.ExperimentConfig:
                       # 0.045 * NUM_GPUS
                       'initial_learning_rate': 0.045 * (train_batch_size // 96),
                       # (2.5 / NUM_GPUS) epochs
-                      'decay_steps': int((2.5 / (train_batch_size // 96))
-                                         * steps_per_epoch),
+                      'decay_steps': int(2.5 * steps_per_epoch),
                       'decay_rate': 0.98,
                       'staircase': True
                   }