The rate of dropout seems to be inverted in your implementation. The inception v4 paper describes the dropout layer as keeping 80% of the units, but Keras' dropout layer takes the drop rate as a parameter. If I'm not mistaken, the model is actually keeping only 20% of the units.
I'm however not 100% sure about that; Keras' implementation of dropout seems to take a different approach (which I don't fully understand), but by reading the paper and Keras' docs, I believe the mistake is there.
The rate of dropout seems to be inverted in your implementation. The inception v4 paper describes the dropout layer as keeping 80% of the units, but Keras' dropout layer takes the drop rate as a parameter. If I'm not mistaken, the model is actually keeping only 20% of the units.
I'm however not 100% sure about that; Keras' implementation of dropout seems to take a different approach (which I don't fully understand), but by reading the paper and Keras' docs, I believe the mistake is there.