def __init__(self, lm, recalculate_props=False, directory='searcher'):
self.lm = lm
# load all the variables and parameters
# I'm fixing the file locations by hand because lazy.
self.gen_config = gen_model_beam_search.Config(lm)
self.gen_config.load(directory + '/gen.parameters')
self.gen_var = gen_model_beam_search.Variables(self.gen_config)
self.gen_var.load(directory + '/gen.weights')
self.pred_config = pred_model_run.Config(lm)
self.pred_config.load(directory + '/pred.parameters')
self.pred_var = pred_model_run.Variables(self.pred_config)
self.pred_var.load(directory + '/pred.weights')
self.payout_config = payout_model_run.Config(lm)
self.payout_config.load(directory + '/payout.parameters')
self.payout_var = payout_model_run.Variables(self.payout_config)
self.payout_var.load(directory + '/payout.weights')
# beam search interface
self.bsi = gen_model_beam_search.BeamSearchInterface([self.gen_var] *
GEN_ENSEMBLE)
# remember the answer so that we don't need to constantly recalculate it
file_path = directory + '/pred_database'
if os.path.isfile(file_path) and not recalculate_props:
print 'loading proposition vectors'
with open(file_path, 'rb') as handle:
self.pred_database = pickle.load(handle)
else:
print 'using proposition vectors at ' + file_path
self.initialize_pred(file_path)
# text = file_contents()
# database = meta_math_database(text,n=2000)
with open('lm', 'rb') as handle:
database = pickle.load(handle)
database.remember_proof_steps = False
language_model = LanguageModel(database)
print()
print()
# this is the main routine for actual training
#import pred_model as model
#import gen_model_train as model
import payout_model_5_train as model
config = model.Config(language_model)
config.p.lr = 1.0e-5
config.p.r = 128
config.regularization = 1.0e-4
config.p.lr_reduction = 10
config.p.gru_depth = 2
#config.p.dropout = None
#config.p.augmentation= False
config.p.structure_data = True
#config.p.attention=True
#config.p.full_state_attention=False
config.p.bidirectional = True
config.p.out_layers = 1
config.p.max_epochs = None
train_object = trainer.Trainer(config,