ai4reason / enigma-gpu-server

Tensorflow GPU server for fast evaluation with ENIGMA E Prover

ENIGMA GPU Server

Tensorflow GPU server for fast evaluation with ENIGMA E Prover.

Running the Server

The server implementation is in the directory tf-server. The main launch file is tf-server/tf_server_workers_thread.py. Example models are in the directory models. Update the values of variables SERVER_IP and SERVER_PORT in the file tf-server/tf_server_workers_thread.py if the default values (127.0.0.1 and 8888) need to be changed in your scenario.

Running the E Client

Download and compile an E client with GPU server support from

• https://github.com/ai4reason/eprover/tree/enigma

To instruct the client to connect to the GPU server, you must use EnigmaticTfs clause weigth function in your E Prover strategy. Its syntax is as follows:

EnigmaticTfs(prio_fun, server_ip, server_port, context_size, weight_type, threshold)

where

• prio_fun is E's standard priority function (we use ConstPrio),
• server_ip is server's IP address as a string in the format nn.nn.nn.nn,
• server_port is server's port at integer,
• context_size is the size of the context query,
• weight_type should be set to 2 (to use a sigmoid-like function to translate neural network's logits to the interval (0,1))
• threshold is a threshold to decide which clause weights should be consider positives/negatives (use 0.5 as a default).

An example E strategy with EnigmaticTfs is as follows:

eprover -s --print-statistics --resources-info --definitional-cnf=24 --split-aggressive 
  --simul-paramod --forward-context-sr --destructive-er-aggressive --destructive-er
  --prefer-initial-clauses -tKBO6 -winvfreqrank -c1 -Ginvfreq -F1 --delete-bad-limit=150000000 
  -WSelectMaxLComplexAvoidPosPred --delayed-eval-cache=128 
  -H'(1*EnigmaticTfs(ConstPrio,127.0.0.1,8888,1024,2,0.5))' problem.p

Note that the argument --delayed-eval-cache=128 is used to select the query size.

Running the 2-phase ENIGMA: GNN with LGB pre-filtering

In order to use a fast and cheap pre-filtering of the clauses sent to the GPU server, use the same clause weight function EnigmaticTfs as above, but with additional parameters.

EnigmaticTfs(..., threshold, lgb_model_dir, lgb_weight_type, lgb_threshold)

The additional parameters are as follows:

• lgb_model_dir is the directory with LGB model,
• lgb_weight_type should be set to 1 (to use the threshold below)
• lgb_threshold is the threshold to decide which clauses to evaluate on the GPU server (a real number between 0 and 1).

Running E with parental guidance.

Download and compile an E client from

• https://github.com/zariuq/eprover/tree/parentalguidance_frozen

Parental guidance is used via a commandline argument to E:

--filter-generated-clauses=parental_lgb_model_dir
--filter-generated-threshold=parental_threshold

An LGB model for clause selection can be used via the clause weight function EnigmaticLgb:

EnigmaticLgb(prio_fun, selection_lgb_model_dir, weight_type, threshold)

where

• parental_lgb_model_dir is the directory with the LGB model for parental guidance,
• parental_threshold is the threshold below which a pair of inference parents will be filtered (a real number between 0 and 1).
• selection_lgb_model_dir is the directory with the LGB model for clause selection,
• weight_type can be set to 1 to simply rate clauses as positive or negative.

Repositories to help with training models can be found at:

• https://github.com/zariuq/pyprove/tree/parentalguidance_frozen
• https://github.com/zariuq/enigmatic/tree/parentalguidance_frozen

Requirements

The following Python packages are required (installable through pip or otherwise):

• tensorflow (1.5)
• xopen
• asyncio