load_gpt2() improvements

• load_gpt2() in a fresh session is much faster and uses much less memory when loaded. (for the 117M model, the system will stay under <2 GB RAM which is the critical point for cloud services)
• start_tf_sess() now accepts a threads parameter, which is useful if you know exactly how many threads will be used.

Fix CSV Finetuning

Number of CSV tokens was inadvertently doubled. (#25)

345M model support

• Support the 345M model (thanks to Neil Shepperd for the gradient checkpointing implementation!)
• Support model_name in the CLI for above support
• Support run_name in the CLI
• Support .csv files as an input dataset to finetune (will parse the CSV as if it was done via encode_csv()).
• Fix one off issues (#21)

Better restore and checkpoint behavior

• Fix one-off error where checkpoint saved a step early.
• Fix issue where restore_from='fresh uses the counter from a previously-trained checkpoint.
• If restore_from='latest , steps will now train for the specified amount of steps, instead of the training until the specified number of steps. (#13, #14)

CLI

• Added a basic CLI.
• Added a include_prefix parameter to give an option to exclude the input prefix.
• Improved regex for truncation.

More utility functions

• is_gpt2_downloaded: Check if the model is downloaded.
• encode_csv: Convert a CSV to a format suitable for GPT-2.

Initial release

v0.1

README + setup.py cleanup