Created by: shippingwang

• support output the JSON file while doing multi-card eval in the image classification Note: To solve the problem about the framework cannot pass string variable into the network and there is a strong demand to get it, we create a reader which always generate a batch of image_id(string) and image_data(data and label) and only feed the image_data when the executor runs, the python interface: feeder.feed_parallel is workable to implement it, note that feeder.decorate_reader ask the reader should be consistent with the feeder After doing this, the exe.run will return the observation value in the fetch list as a shape of [the single card batch size, the size of fetch_list, image_shape] The JSON file will be saved at the ./ in default
• add args:exclude to load pre-trained model correctly while doing finetune
• refine some warning