Add learning rate scheduling for Transformer.

Make gradient attack method support norm (L1/L2/L∞, etc.).

add Paddle version requirements to all models.

move paddle.v2.fluid to paddle.fluid

Fix the problem that padding index contribute to the final loss.

Add Transformer by using Fluid.

Enable checkpoints saving and training resuming

Merge from upstream

Add multiGPU in SE-ResNeXt

adversarial example attack method -- jsma

add fluid mobilenet config

Add the demo script for inference

Remove uncessary print