def percrank_train(args):
opts, files = getopt(args, 'c:d:s:j:w:e:r:')
candgen_model = None
train_size = 1.0
parallel = False
jobs_number = 0
work_dir = None
experiment_id = None
for opt, arg in opts:
if opt == '-d':
set_debug_stream(file_stream(arg, mode='w'))
elif opt == '-s':
train_size = float(arg)
elif opt == '-c':
candgen_model = arg
elif opt == '-j':
parallel = True
jobs_number = int(arg)
elif opt == '-w':
work_dir = arg
elif opt == '-e':
experiment_id = arg
elif opt == '-r' and arg:
rnd.seed(arg)
if len(files) != 4:
sys.exit(__doc__)
fname_rank_config, fname_train_das, fname_train_ttrees, fname_rank_model = files
log_info('Training perceptron ranker...')
rank_config = Config(fname_rank_config)
if candgen_model:
rank_config['candgen_model'] = candgen_model
if rank_config.get('nn'):
from tgen.rank_nn import SimpleNNRanker, EmbNNRanker
if rank_config['nn'] in ['emb', 'emb_trees', 'emb_prev']:
ranker_class = EmbNNRanker
else:
ranker_class = SimpleNNRanker
else:
ranker_class = PerceptronRanker
log_info('Using %s for ranking' % ranker_class.__name__)
if not parallel:
ranker = ranker_class(rank_config)
else:
rank_config['jobs_number'] = jobs_number
if work_dir is None:
work_dir, _ = os.path.split(fname_rank_config)
ranker = ParallelRanker(rank_config, work_dir, experiment_id,
ranker_class)
ranker.train(fname_train_das, fname_train_ttrees, data_portion=train_size)
# avoid the "maximum recursion depth exceeded" error
sys.setrecursionlimit(100000)
ranker.save_to_file(fname_rank_model)
def percrank_train(args):
opts, files = getopt(args, 'c:d:s:j:w:e:r:')
candgen_model = None
train_size = 1.0
parallel = False
jobs_number = 0
work_dir = None
experiment_id = None
for opt, arg in opts:
if opt == '-d':
set_debug_stream(file_stream(arg, mode='w'))
elif opt == '-s':
train_size = float(arg)
elif opt == '-c':
candgen_model = arg
elif opt == '-j':
parallel = True
jobs_number = int(arg)
elif opt == '-w':
work_dir = arg
elif opt == '-e':
experiment_id = arg
elif opt == '-r' and arg:
rnd.seed(arg)
if len(files) != 4:
sys.exit(__doc__)
fname_rank_config, fname_train_das, fname_train_ttrees, fname_rank_model = files
log_info('Training perceptron ranker...')
rank_config = Config(fname_rank_config)
if candgen_model:
rank_config['candgen_model'] = candgen_model
if rank_config.get('nn'):
from tgen.rank_nn import SimpleNNRanker, EmbNNRanker
if rank_config['nn'] in ['emb', 'emb_trees', 'emb_prev']:
ranker_class = EmbNNRanker
else:
ranker_class = SimpleNNRanker
else:
ranker_class = PerceptronRanker
log_info('Using %s for ranking' % ranker_class.__name__)
if not parallel:
ranker = ranker_class(rank_config)
else:
rank_config['jobs_number'] = jobs_number
if work_dir is None:
work_dir, _ = os.path.split(fname_rank_config)
ranker = ParallelRanker(rank_config, work_dir, experiment_id, ranker_class)
ranker.train(fname_train_das, fname_train_ttrees, data_portion=train_size)
# avoid the "maximum recursion depth exceeded" error
sys.setrecursionlimit(100000)
ranker.save_to_file(fname_rank_model)
def seq2seq_train(args):
ap = ArgumentParser(prog=' '.join(sys.argv[0:2]))
ap.add_argument('-s', '--train-size', type=float,
help='Portion of the training data to use (default: 1.0)', default=1.0)
ap.add_argument('-d', '--debug-logfile', type=str, help='Debug output file name')
ap.add_argument('-j', '--jobs', type=int, help='Number of parallel jobs to use')
ap.add_argument('-w', '--work-dir', type=str, help='Main working directory for parallel jobs')
ap.add_argument('-e', '--experiment-id', type=str,
help='Experiment ID for parallel jobs (used as job name prefix)')
ap.add_argument('-r', '--random-seed', type=str,
help='Initial random seed (used as string).')
ap.add_argument('-c', '--context-file', type=str,
help='Input ttree/text file with context utterances')
ap.add_argument('-v', '--valid-data', type=str,
help='Validation data paths (2-3 comma-separated files: DAs, trees/sentences, contexts)')
ap.add_argument('-l', '--lexic-data', type=str,
help='Lexicalization data paths (1-2 comma-separated files: surface forms,' +
'training lexic. instructions)')
ap.add_argument('-t', '--tb-summary-dir', '--tensorboard-summary-dir', '--tensorboard', type=str,
help='Directory where Tensorboard summaries are saved during training')
ap.add_argument('seq2seq_config_file', type=str, help='Seq2Seq generator configuration file')
ap.add_argument('da_train_file', type=str, help='Input training DAs')
ap.add_argument('tree_train_file', type=str, help='Input training trees/sentences')
ap.add_argument('seq2seq_model_file', type=str,
help='File name where to save the trained Seq2Seq generator model')
args = ap.parse_args(args)
if args.debug_logfile:
set_debug_stream(file_stream(args.debug_logfile, mode='w'))
if args.random_seed:
rnd.seed(args.random_seed)
log_info('Training sequence-to-sequence generator...')
config = Config(args.seq2seq_config_file)
if args.tb_summary_dir: # override Tensorboard setting
config['tb_summary_dir'] = args.tb_summary_dir
if args.jobs: # parallelize when training
config['jobs_number'] = args.jobs
if not args.work_dir:
work_dir, _ = os.path.split(args.seq2seq_config_file)
generator = ParallelSeq2SeqTraining(config, args.work_dir or work_dir, args.experiment_id)
else: # just a single training instance
generator = Seq2SeqGen(config)
generator.train(args.da_train_file, args.tree_train_file,
data_portion=args.train_size, context_file=args.context_file,
validation_files=args.valid_data, lexic_files=args.lexic_data)
sys.setrecursionlimit(100000)
generator.save_to_file(args.seq2seq_model_file)
def seq2seq_train(args):
ap = ArgumentParser(prog=' '.join(sys.argv[0:2]))
ap.add_argument('-s', '--train-size', type=float,
help='Portion of the training data to use (default: 1.0)', default=1.0)
ap.add_argument('-d', '--debug-logfile', type=str, help='Debug output file name')
ap.add_argument('-j', '--jobs', type=int, help='Number of parallel jobs to use')
ap.add_argument('-w', '--work-dir', type=str, help='Main working directory for parallel jobs')
ap.add_argument('-e', '--experiment-id', type=str,
help='Experiment ID for parallel jobs (used as job name prefix)')
ap.add_argument('-r', '--random-seed', type=str,
help='Initial random seed (used as string).')
ap.add_argument('-c', '--context-file', type=str,
help='Input ttree/text file with context utterances')
ap.add_argument('-v', '--valid-data', type=str,
help='Validation data paths (2-3 comma-separated files: DAs, trees/sentences, contexts)')
ap.add_argument('-l', '--lexic-data', type=str,
help='Lexicalization data paths (1-2 comma-separated files: surface forms,' +
'training lexic. instructions)')
ap.add_argument('-t', '--tb-summary-dir', '--tensorboard-summary-dir', '--tensorboard', type=str,
help='Directory where Tensorboard summaries are saved during training')
ap.add_argument('seq2seq_config_file', type=str, help='Seq2Seq generator configuration file')
ap.add_argument('da_train_file', type=str, help='Input training DAs')
ap.add_argument('tree_train_file', type=str, help='Input training trees/sentences')
ap.add_argument('seq2seq_model_file', type=str,
help='File name where to save the trained Seq2Seq generator model')
args = ap.parse_args(args)
if args.debug_logfile:
set_debug_stream(file_stream(args.debug_logfile, mode='w'))
if args.random_seed:
rnd.seed(args.random_seed)
log_info('Training sequence-to-sequence generator...')
config = Config(args.seq2seq_config_file)
if args.tb_summary_dir: # override Tensorboard setting
config['tb_summary_dir'] = args.tb_summary_dir
if args.jobs: # parallelize when training
config['jobs_number'] = args.jobs
if not args.work_dir:
work_dir, _ = os.path.split(args.seq2seq_config_file)
generator = ParallelSeq2SeqTraining(config, args.work_dir or work_dir, args.experiment_id)
else: # just a single training instance
generator = Seq2SeqGen(config)
generator.train(args.da_train_file, args.tree_train_file,
data_portion=args.train_size, context_file=args.context_file,
validation_files=args.valid_data, lexic_files=args.lexic_data)
sys.setrecursionlimit(100000)
generator.save_to_file(args.seq2seq_model_file)
def exposed_train(self, rnd_seed, das_file, ttree_file, data_portion, context_file, validation_files):
"""Run the whole training.
"""
rnd.seed(rnd_seed)
log_info('Random seed: %f' % rnd_seed)
tstart = time.time()
log_info('Starting training...')
self.seq2seq.train(das_file, ttree_file, data_portion, context_file, validation_files)
log_info('Training finished -- time taken: %f secs.' % (time.time() - tstart))
top_cost = self.seq2seq.top_k_costs[0]
log_info('Best cost: %f' % top_cost)
return top_cost
def exposed_train(self, rnd_seed, das_file, ttree_file, data_portion, context_file, validation_files):
"""Run the whole training.
"""
rnd.seed(rnd_seed)
log_info('Random seed: %f' % rnd_seed)
tstart = time.time()
log_info('Starting training...')
self.seq2seq.train(das_file, ttree_file, data_portion, context_file, validation_files)
log_info('Training finished -- time taken: %f secs.' % (time.time() - tstart))
top_cost = self.seq2seq.top_k_costs[0]
log_info('Best cost: %f' % top_cost)
return top_cost
def exposed_training_pass(self, w, pass_no, rnd_seed, data_offset,
data_len):
"""(Worker) Run one pass over a part of the training data.
@param w: initial perceptron weights (pickled)
@param pass_no: pass number (for logging purposes)
@param rnd_seed: random generator seed for shuffling training examples
@param data_offset: training data portion start
@param data_len: training data portion size
@return: updated perceptron weights after passing the selected data portion (pickled)
"""
log_info('Training pass %d with data portion %d + %d' %
(pass_no, data_offset, data_len))
# use the local ranker instance
ranker = self.ranker_inst
# import current feature weights
tstart = time.time()
ranker.set_weights(pickle.loads(w))
log_info('Weights loading: %f secs.' % (time.time() - tstart))
# save rest of the training data to temporary variables, set just the
# required portion for computation
all_train_das = ranker.train_das
ranker.train_das = ranker.train_das[data_offset:data_offset + data_len]
all_train_trees = ranker.train_trees
ranker.train_trees = ranker.train_trees[data_offset:data_offset +
data_len]
all_train_feats = ranker.train_feats
ranker.train_feats = ranker.train_feats[data_offset:data_offset +
data_len]
all_train_sents = ranker.train_sents
ranker.train_sents = ranker.train_sents[data_offset:data_offset +
data_len]
all_train_order = ranker.train_order
ranker.train_order = range(len(ranker.train_trees))
if ranker.randomize:
rnd.seed(rnd_seed)
rnd.shuffle(ranker.train_order)
# do the actual computation (update w)
ranker._training_pass(pass_no)
# return the rest of the training data to member variables
ranker.train_das = all_train_das
ranker.train_trees = all_train_trees
ranker.train_feats = all_train_feats
ranker.train_sents = all_train_sents
ranker.train_order = all_train_order
# return the result of the computation
log_info('Training pass %d / %d / %d done.' %
(pass_no, data_offset, data_len))
tstart = time.time()
dump = pickle.dumps((ranker.get_weights(), ranker.get_diagnostics()),
pickle.HIGHEST_PROTOCOL)
log_info('Weights saving: %f secs.' % (time.time() - tstart))
return dump
def seq2seq_train(args):
ap = ArgumentParser()
ap.add_argument('-s', '--train-size', type=float,
help='Portion of the training data to use (default: 1.0)', default=1.0)
ap.add_argument('-d', '--debug-logfile', type=str, help='Debug output file name')
ap.add_argument('-j', '--jobs', type=int, help='Number of parallel jobs to use')
ap.add_argument('-w', '--work-dir', type=str, help='Main working for parallel jobs')
ap.add_argument('-e', '--experiment-id', type=str,
help='Experiment ID for parallel jobs (used as job name prefix)')
ap.add_argument('-r', '--random-seed', type=str,
help='Initial random seed (used as string).')
ap.add_argument('-c', '--context-file', type=str,
help='Input ttree/text file with context utterances')
ap.add_argument('-v', '--valid-data', type=str,
help='Validation data paths (2-3 comma-separated files: DAs, trees/sentences, contexts)')
ap.add_argument('seq2seq_config_file', type=str, help='Seq2Seq generator configuration file')
ap.add_argument('da_train_file', type=str, help='Input training DAs')
ap.add_argument('tree_train_file', type=str, help='Input training trees/sentences')
ap.add_argument('seq2seq_model_file', type=str,
help='File name where to save the trained Seq2Seq generator model')
args = ap.parse_args(args)
if args.debug_logfile:
set_debug_stream(file_stream(args.debug_logfile, mode='w'))
if args.random_seed:
rnd.seed(rnd.seed(args.random_seed))
log_info('Training sequence-to-sequence generator...')
config = Config(args.seq2seq_config_file)
if args.jobs:
config['jobs_number'] = args.jobs
if not args.work_dir:
work_dir, _ = os.path.split(args.seq2seq_config_file)
generator = ParallelSeq2SeqTraining(config, args.work_dir or work_dir, args.experiment_id)
else:
generator = Seq2SeqGen(config)
generator.train(args.da_train_file, args.tree_train_file,
data_portion=args.train_size, context_file=args.context_file,
validation_files=args.valid_data)
sys.setrecursionlimit(100000)
generator.save_to_file(args.seq2seq_model_file)
def exposed_training_pass(self, w, pass_no, rnd_seed, data_offset, data_len):
"""(Worker) Run one pass over a part of the training data.
@param w: initial perceptron weights (pickled)
@param pass_no: pass number (for logging purposes)
@param rnd_seed: random generator seed for shuffling training examples
@param data_offset: training data portion start
@param data_len: training data portion size
@return: updated perceptron weights after passing the selected data portion (pickled)
"""
log_info('Training pass %d with data portion %d + %d' %
(pass_no, data_offset, data_len))
# use the local ranker instance
ranker = self.ranker_inst
# import current feature weights
tstart = time.time()
ranker.set_weights(pickle.loads(w))
log_info('Weights loading: %f secs.' % (time.time() - tstart))
# save rest of the training data to temporary variables, set just the
# required portion for computation
all_train_das = ranker.train_das
ranker.train_das = ranker.train_das[data_offset:data_offset + data_len]
all_train_trees = ranker.train_trees
ranker.train_trees = ranker.train_trees[data_offset:data_offset + data_len]
all_train_feats = ranker.train_feats
ranker.train_feats = ranker.train_feats[data_offset:data_offset + data_len]
all_train_sents = ranker.train_sents
ranker.train_sents = ranker.train_sents[data_offset:data_offset + data_len]
all_train_order = ranker.train_order
ranker.train_order = range(len(ranker.train_trees))
if ranker.randomize:
rnd.seed(rnd_seed)
rnd.shuffle(ranker.train_order)
# do the actual computation (update w)
ranker._training_pass(pass_no)
# return the rest of the training data to member variables
ranker.train_das = all_train_das
ranker.train_trees = all_train_trees
ranker.train_feats = all_train_feats
ranker.train_sents = all_train_sents
ranker.train_order = all_train_order
# return the result of the computation
log_info('Training pass %d / %d / %d done.' % (pass_no, data_offset, data_len))
tstart = time.time()
dump = pickle.dumps((ranker.get_weights(), ranker.get_diagnostics()), pickle.HIGHEST_PROTOCOL)
log_info('Weights saving: %f secs.' % (time.time() - tstart))
return dump
def train(args):
if args.random_seed: # set random seed if needed
rnd.seed(args.random_seed)
log_info("Loading configuration from %s..." % args.config_file)
with codecs.open(args.config_file, 'r', 'UTF-8') as fh:
cfg = yaml.load(fh)
log_info("Initializing...")
rp = RatingPredictor(cfg)
if args.tensorboard_dir_id is not None:
tb_dir, run_id = args.tensorboard_dir_id.split(':', 1)
rp.set_tensorboard_logging(tb_dir, run_id)
log_info("Training...")
rp.train(args.train_data,
valid_data_file=args.valid_data,
data_portion=args.training_portion,
model_fname=args.model_file)
log_info("Saving model to %s..." % args.model_file)
rp.save_to_file(args.model_file)
