def setUp(self):
xnmt.events.clear()
self.model_context = ModelContext()
self.model_context.dynet_param_collection = PersistentParamCollection(
"some_file", 1)
self.model = DefaultTranslator(
src_embedder=SimpleWordEmbedder(self.model_context,
vocab_size=100),
encoder=BiLSTMSeqTransducer(self.model_context),
attender=MlpAttender(self.model_context),
trg_embedder=SimpleWordEmbedder(self.model_context,
vocab_size=100),
decoder=MlpSoftmaxDecoder(self.model_context,
vocab_size=100,
bridge=CopyBridge(self.model_context,
dec_layers=1)),
)
self.model.initialize_training_strategy(TrainingStrategy())
self.model.set_train(False)
self.model.initialize_generator()
self.training_corpus = BilingualTrainingCorpus(
train_src="examples/data/head.ja",
train_trg="examples/data/head.en",
dev_src="examples/data/head.ja",
dev_trg="examples/data/head.en")
self.corpus_parser = BilingualCorpusParser(
src_reader=PlainTextReader(),
trg_reader=PlainTextReader(),
training_corpus=self.training_corpus)
def setUp(self):
xnmt.events.clear()
self.input_reader = PlainTextReader()
list(self.input_reader.read_sents('examples/data/head.ja'))
self.input_reader.freeze()
self.context = ModelContext()
self.context.dynet_param_collection = PersistentParamCollection(
None, 0)
def setUp(self):
xnmt.events.clear()
self.model_context = ModelContext()
self.model_context.dynet_param_collection = PersistentParamCollection(
"some_file", 1)
self.training_corpus = BilingualTrainingCorpus(
train_src="examples/data/head.ja",
train_trg="examples/data/head.en",
dev_src="examples/data/head.ja",
dev_trg="examples/data/head.en")
self.corpus_parser = BilingualCorpusParser(
src_reader=PlainTextReader(),
trg_reader=PlainTextReader(),
training_corpus=self.training_corpus)
def test_overfitting(self):
self.model_context = ModelContext()
self.model_context.dynet_param_collection = PersistentParamCollection(
"some_file", 1)
self.model_context.default_layer_dim = 16
train_args = {}
training_corpus = BilingualTrainingCorpus(
train_src="examples/data/head.ja",
train_trg="examples/data/head.en",
dev_src="examples/data/head.ja",
dev_trg="examples/data/head.en")
train_args['corpus_parser'] = BilingualCorpusParser(
training_corpus=training_corpus,
src_reader=PlainTextReader(),
trg_reader=PlainTextReader())
train_args['training_strategy'] = TrainingStrategy()
train_args['model'] = DefaultTranslator(
src_embedder=SimpleWordEmbedder(self.model_context,
vocab_size=100),
encoder=BiLSTMSeqTransducer(self.model_context),
attender=MlpAttender(self.model_context),
trg_embedder=SimpleWordEmbedder(self.model_context,
vocab_size=100),
decoder=MlpSoftmaxDecoder(self.model_context, vocab_size=100),
)
train_args['model_file'] = None
train_args['save_num_checkpoints'] = 0
train_args['trainer'] = AdamTrainer(self.model_context, alpha=0.1)
train_args['batcher'] = SrcBatcher(batch_size=10,
break_ties_randomly=False)
training_regimen = xnmt.train.TrainingRegimen(
yaml_context=self.model_context, **train_args)
training_regimen.model_context = self.model_context
for _ in range(50):
training_regimen.one_epoch(update_weights=True)
self.assertAlmostEqual(
0.0,
training_regimen.logger.epoch_loss.loss_values['loss'] /
training_regimen.logger.epoch_words,
places=2)
def test_train_dev_loss_equal(self):
self.model_context = ModelContext()
self.model_context.dynet_param_collection = NonPersistentParamCollection(
)
train_args = {}
training_corpus = BilingualTrainingCorpus(
train_src="examples/data/head.ja",
train_trg="examples/data/head.en",
dev_src="examples/data/head.ja",
dev_trg="examples/data/head.en")
train_args['corpus_parser'] = BilingualCorpusParser(
training_corpus=training_corpus,
src_reader=PlainTextReader(),
trg_reader=PlainTextReader())
train_args['loss_calculator'] = LossCalculator()
train_args['model'] = DefaultTranslator(
src_embedder=SimpleWordEmbedder(self.model_context,
vocab_size=100),
encoder=BiLSTMSeqTransducer(self.model_context),
attender=MlpAttender(self.model_context),
trg_embedder=SimpleWordEmbedder(self.model_context,
vocab_size=100),
decoder=MlpSoftmaxDecoder(self.model_context, vocab_size=100),
)
train_args['trainer'] = None
train_args['batcher'] = SrcBatcher(batch_size=5,
break_ties_randomly=False)
train_args['run_for_epochs'] = 1
training_regimen = xnmt.training_regimen.SimpleTrainingRegimen(
yaml_context=self.model_context, **train_args)
training_regimen.model_context = self.model_context
training_regimen.run_training(update_weights=False)
self.assertAlmostEqual(
training_regimen.logger.epoch_loss.loss_values['loss'] /
training_regimen.logger.epoch_words,
training_regimen.logger.dev_score.loss)
def main(overwrite_args=None):
argparser = argparse.ArgumentParser()
argparser.add_argument("--dynet-mem", type=int)
argparser.add_argument("--dynet-seed", type=int)
argparser.add_argument("--dynet-autobatch", type=int)
argparser.add_argument("--dynet-devices", type=str)
argparser.add_argument("--dynet-viz", action='store_true', help="use visualization")
argparser.add_argument("--dynet-gpu", action='store_true', help="use GPU acceleration")
argparser.add_argument("--dynet-gpu-ids", type=int)
argparser.add_argument("--dynet-gpus", type=int)
argparser.add_argument("--dynet-weight-decay", type=float)
argparser.add_argument("--generate-doc", action='store_true', help="Do not run, output documentation instead")
argparser.add_argument("experiments_file")
argparser.add_argument("experiment_name", nargs='*', help="Run only the specified experiments")
argparser.set_defaults(generate_doc=False)
args = argparser.parse_args(overwrite_args)
config_parser = OptionParser()
if args.generate_doc:
print(config_parser.generate_options_table())
exit(0)
if args.dynet_seed:
random.seed(args.dynet_seed)
np.random.seed(args.dynet_seed)
config_experiment_names = config_parser.experiment_names_from_file(args.experiments_file)
results = []
# Check ahead of time that all experiments exist, to avoid bad surprises
experiment_names = args.experiment_name or config_experiment_names
if args.experiment_name:
nonexistent = set(experiment_names).difference(config_experiment_names)
if len(nonexistent) != 0:
raise Exception("Experiments {} do not exist".format(",".join(list(nonexistent))))
for experiment_name in experiment_names:
exp_tasks = config_parser.parse_experiment(args.experiments_file, experiment_name)
print("=> Running {}".format(experiment_name))
exp_args = exp_tasks.get("experiment", {})
# TODO: refactor
if not "model_file" in exp_args: exp_args["model_file"] = "<EXP>.mod"
if not "hyp_file" in exp_args: exp_args["hyp_file"] = "<EXP>.hyp"
if not "out_file" in exp_args: exp_args["out_file"] = "<EXP>.out"
if not "err_file" in exp_args: exp_args["model_file"] = "<EXP>.err"
if not "cfg_file" in exp_args: exp_args["cfg_file"] = None
if not "eval_only" in exp_args: exp_args["eval_only"] = False
if not "eval_metrics" in exp_args: exp_args["eval_metrics"] = "bleu"
if "cfg_file" in exp_args and exp_args["cfg_file"] != None:
shutil.copyfile(args.experiments_file, exp_args["cfg_file"])
preproc_args = exp_tasks.get("preproc", {})
# Do preprocessing
print("> Preprocessing")
xnmt.xnmt_preproc.xnmt_preproc(**preproc_args)
print("> Initializing TrainingRegimen")
train_args = exp_tasks["train"]
train_args.model_file = exp_args["model_file"] # TODO: can we use param sharing for this?
model_context = ModelContext()
model_context.dynet_param_collection = PersistentParamCollection(exp_args["model_file"], 1)
if hasattr(train_args, "glob"):
for k in train_args.glob:
setattr(model_context, k, train_args.glob[k])
train_args = YamlSerializer().initialize_if_needed(UninitializedYamlObject(train_args), model_context)
xnmt_decoder = exp_tasks.get("decode", {})
xnmt_decoder.trg_file = exp_args["hyp_file"] # TODO: can we use param sharing for this?
xnmt_decoder.model_file = None # The model is passed to the decoder directly
xnmt_decoder = YamlSerializer().initialize_if_needed(UninitializedYamlObject(xnmt_decoder), model_context)
evaluate_args = exp_tasks.get("evaluate", {})
evaluate_args["hyp_file"] = exp_args["hyp_file"]
evaluators = map(lambda s: s.lower(), exp_args["eval_metrics"].split(","))
output = Tee(exp_args["out_file"], 3)
err_output = Tee(exp_args["err_file"], 3, error=True)
# Do training
if "random_search_report" in exp_tasks:
print("> instantiated random parameter search: %s" % exp_tasks["random_search_report"])
print("> Training")
training_regimen = train_args
training_regimen.xnmt_decoder = copy.copy(xnmt_decoder)
training_regimen.evaluate_args = copy.copy(evaluate_args)
eval_scores = "Not evaluated"
if not exp_args["eval_only"]:
training_regimen.run_epochs(exp_args["run_for_epochs"])
if not exp_args["eval_only"]:
print('reverting learned weights to best checkpoint..')
training_regimen.model_context.dynet_param_collection.revert_to_best_model()
if evaluators:
print("> Evaluating test set")
output.indent += 2
xnmt_decoder(model_elements=(training_regimen.corpus_parser, training_regimen.model))
eval_scores = []
for evaluator in evaluators:
evaluate_args["evaluator"] = evaluator
eval_score = xnmt.xnmt_evaluate.xnmt_evaluate(**evaluate_args)
print(eval_score)
eval_scores.append(eval_score)
output.indent -= 2
results.append((experiment_name, eval_scores))
output.close()
err_output.close()
print("")
print("{:<30}|{:<40}".format("Experiment", " Final Scores"))
print("-" * (70 + 1))
for line in results:
experiment_name, eval_scores = line
for i in range(len(eval_scores)):
print("{:<30}| {:<40}".format((experiment_name if i==0 else ""), str(eval_scores[i])))
def __init__(self,
corpus_parser,
model_file,
model,
yaml_context=None,
glob={},
dev_every=0,
batcher=None,
training_strategy=None,
save_num_checkpoints=1,
pretrained_model_file="",
src_format="text",
trainer=None,
lr_decay=1.0,
lr_decay_times=3,
attempts_before_lr_decay=1,
dev_metrics="",
schedule_metric="loss",
restart_trainer=False,
reload_command=None):
"""
:param corpus_parser:
:param model_file:
:param model:
:param yaml_context: (TODO: remove default value)
:param dev_every (int): dev checkpoints every n sentences (0 for only after epoch)
:param batcher: Type of batcher. Defaults to SrcBatcher of batch size 32.
:param training_strategy:
:param save_num_checkpoints (int): Save recent n best checkpoints
:param pretrained_model_file: Path of pre-trained model file
:param src_format: Format of input data: text/contvec
:param trainer: Trainer object, default is SGD with learning rate 0.1
:param lr_decay (float):
:param lr_decay_times (int): Early stopping after decaying learning rate a certain number of times
:param attempts_before_lr_decay (int): apply LR decay after dev scores haven't improved over this many checkpoints
:param dev_metrics: Comma-separated list of evaluation metrics (bleu/wer/cer)
:param schedule_metric: determine learning schedule based on this dev_metric (loss/bleu/wer/cer)
:param restart_trainer: Restart trainer (useful for Adam) and revert weights to best dev checkpoint when applying LR decay (https://arxiv.org/pdf/1706.09733.pdf)
:param reload_command: Command to change the input data after each epoch.
--epoch EPOCH_NUM will be appended to the command.
To just reload the data after each epoch set the command to 'true'.
"""
dy.renew_cg()
# TODO: don't need to keep a dedicated args object any longer
args = dict(dev_every=dev_every,
batcher=batcher,
corpus_parser=corpus_parser,
training_strategy=training_strategy,
model_file=model_file,
save_num_checkpoints=save_num_checkpoints,
pretrained_model_file=pretrained_model_file,
src_format=src_format,
default_layer_dim=glob.get("default_layer_dim", 512),
trainer=trainer,
lr_decay=lr_decay,
lr_decay_times=lr_decay_times,
attempts_before_lr_decay=attempts_before_lr_decay,
dev_metrics=dev_metrics,
schedule_metric=schedule_metric,
restart_trainer=restart_trainer,
reload_command=reload_command,
dropout=glob.get("dropout", 0.0),
weight_noise=glob.get("weight_noise", 0.0),
model=model)
self.args = args
if yaml_context:
self.model_context = yaml_context
else:
self.model_context = ModelContext()
self.model_context.dynet_param_collection = PersistentParamCollection(
self.args["model_file"], self.args["save_num_checkpoints"])
if args["lr_decay"] > 1.0 or args["lr_decay"] <= 0.0:
raise RuntimeError(
"illegal lr_decay, must satisfy: 0.0 < lr_decay <= 1.0")
self.num_times_lr_decayed = 0
self.early_stopping_reached = False
self.cur_attempt = 0
self.evaluators = [
s.lower() for s in self.args["dev_metrics"].split(",")
if s.strip() != ""
]
if self.args["schedule_metric"].lower() not in self.evaluators:
self.evaluators.append(self.args["schedule_metric"].lower())
if "loss" not in self.evaluators: self.evaluators.append("loss")
if args["reload_command"] is not None:
self._augmentation_handle = None
self._augment_data_initial()
# Initialize the serializer
self.model_serializer = YamlSerializer()
self.create_corpus_and_model()
self.model.initialize_training_strategy(self.training_strategy)
if self.args["batcher"] is None:
self.batcher = SrcBatcher(32)
else:
self.batcher = self.args["batcher"]
if args["src_format"] == "contvec":
self.batcher.pad_token = np.zeros(self.model.src_embedder.emb_dim)
self.pack_batches()
self.logger = BatchLossTracker(args["dev_every"],
self.total_train_sent)
if args["trainer"] is None:
self.trainer = xnmt.optimizer.SimpleSGDTrainer(
self.model_context, 0.1)
else:
self.trainer = args["trainer"]
