fix converter, fix model load and model save

[yuantailing]

converter.py:

1. To run converter.py, we have to compile caffe at deeplab version (an old version), which doesn't contain caffe.TRAIN.

train_v1.py, train_v2.py:

1. biases are [1, 1, 1, 1024] sized in converted init file, while [1024] is expected in Net. I use squeeze_() as workaround.
2. snapshot_dir may be an empty string, we should use os.path.join to join path robustly.