Ahmed Ibrahim

Adding instructions to install loadcaffe for loading the VGG-16 pretrained model

_parser.add_argument('--max_token_length', default=15, type=int, help="**Set to 0 to disable filtering**")_ Setting that to zero will not disable filtering. it will discard all tokens according to _if max_token_length > 0 and len(tokens)

DataParallel should be used only with GPUs. If a user disabled Cuda in order to run the model on CPU, an error will show up.

Added SimpleITK to the requirements .As per the first line of utils.py, the SimpleITK package is required import SimpleITK as sitk

I was wondering why you are using UpSampling2D vs Conv2DTranspose for a semantic segmentation task. In keras, UpSampling2D is a simple resizing while Conv2DTranspose is the layer known as "deconvolutional"...