Update dropout value

Dropout prob in the transformers model layers is defined as the probability of an activation being set to 0, in the original MobileNet repo it's defined as the keep probability.

See:

https://github.com/huggingface/transformers/issues/22767
https://github.com/tensorflow/models/blob/ad32e81e31232675319d5572b78bc196216fd52e/research/slim/nets/mobilenet_v1.py#L306

config.json

@@ -2,7 +2,7 @@
   "architectures": [
     "MobileNetV2ForImageClassification"
   ],
-  "classifier_dropout_prob": 0.8,
+  "classifier_dropout_prob": 0.2,
   "depth_divisible_by": 8,
   "depth_multiplier": 0.75,
   "expand_ratio": 6,