Refine OCR CTC model.

Replace cross_entropy with softmax_with_cross_entropy in Transformer

data load for inference

Fix the softmax in Transformer.

Add decoder init & verify output

Expose acoustic scale in decoder

Revert AsyncDataReader to avoid using shared memory.

1. Add download of data into ctc_read.py
2. Use 'clone' instead of 'get_inference_program'
3. Add time to train log

SE-ResNeXt python script update

Refine train.py for MobileNet-SSD.

Make buffer queues be objects of class AsyncReader.

Add model average option for OCR CTC model

Add examples for caffe2fluid

Add decoder for deep asr model

Solve the problem of fetching prediction and make data dim configurable