def test_train_dev_loss_equal(self):
layer_dim = 512
batcher = SrcBatcher(batch_size=5, break_ties_randomly=False)
train_args = {}
train_args['src_file'] = "examples/data/head.ja"
train_args['trg_file'] = "examples/data/head.en"
train_args['loss_calculator'] = AutoRegressiveMLELoss()
train_args['model'] = DefaultTranslator(src_reader=PlainTextReader(),
trg_reader=PlainTextReader(),
src_embedder=SimpleWordEmbedder(emb_dim=layer_dim, vocab_size=100),
encoder=BiLSTMSeqTransducer(input_dim=layer_dim, hidden_dim=layer_dim),
attender=MlpAttender(input_dim=layer_dim, state_dim=layer_dim,
hidden_dim=layer_dim),
trg_embedder=SimpleWordEmbedder(emb_dim=layer_dim, vocab_size=100),
decoder=AutoRegressiveDecoder(input_dim=layer_dim,
trg_embed_dim=layer_dim,
rnn=UniLSTMSeqTransducer(input_dim=layer_dim,
hidden_dim=layer_dim,
decoder_input_dim=layer_dim,
yaml_path="model.decoder.rnn"),
transform=NonLinear(input_dim=layer_dim*2, output_dim=layer_dim),
scorer=Softmax(input_dim=layer_dim, vocab_size=100),
bridge=CopyBridge(dec_dim=layer_dim, dec_layers=1)),
)
train_args['dev_tasks'] = [LossEvalTask(model=train_args['model'],
src_file="examples/data/head.ja",
ref_file="examples/data/head.en",
batcher=batcher)]
train_args['trainer'] = DummyTrainer()
train_args['batcher'] = batcher
train_args['run_for_epochs'] = 1
training_regimen = xnmt.training_regimen.SimpleTrainingRegimen(**train_args)
training_regimen.run_training(save_fct = lambda: None)
self.assertAlmostEqual(training_regimen.train_loss_tracker.epoch_loss.sum_factors() / training_regimen.train_loss_tracker.epoch_words,
training_regimen.dev_loss_tracker.dev_score.loss, places=5)
def test_overfitting(self):
layer_dim = 16
batcher = SrcBatcher(batch_size=10, break_ties_randomly=False)
train_args = {}
train_args['src_file'] = "examples/data/head.ja"
train_args['trg_file'] = "examples/data/head.en"
train_args['loss_calculator'] = MLELoss()
train_args['model'] = DefaultTranslator(
src_reader=PlainTextReader(),
trg_reader=PlainTextReader(),
src_embedder=SimpleWordEmbedder(vocab_size=100, emb_dim=layer_dim),
encoder=BiLSTMSeqTransducer(input_dim=layer_dim,
hidden_dim=layer_dim),
attender=MlpAttender(input_dim=layer_dim,
state_dim=layer_dim,
hidden_dim=layer_dim),
trg_embedder=SimpleWordEmbedder(vocab_size=100, emb_dim=layer_dim),
decoder=MlpSoftmaxDecoder(input_dim=layer_dim,
trg_embed_dim=layer_dim,
rnn_layer=UniLSTMSeqTransducer(
input_dim=layer_dim,
hidden_dim=layer_dim,
decoder_input_dim=layer_dim,
yaml_path="model.decoder.rnn_layer"),
mlp_layer=MLP(
input_dim=layer_dim,
hidden_dim=layer_dim,
decoder_rnn_dim=layer_dim,
vocab_size=100,
yaml_path="model.decoder.rnn_layer"),
bridge=CopyBridge(dec_dim=layer_dim,
dec_layers=1)),
)
train_args['dev_tasks'] = [
LossEvalTask(model=train_args['model'],
src_file="examples/data/head.ja",
ref_file="examples/data/head.en",
batcher=batcher)
]
train_args['run_for_epochs'] = 1
train_args['trainer'] = AdamTrainer(alpha=0.1)
train_args['batcher'] = batcher
training_regimen = xnmt.training_regimen.SimpleTrainingRegimen(
**train_args)
for _ in range(50):
training_regimen.run_training(save_fct=lambda: None,
update_weights=True)
self.assertAlmostEqual(
0.0,
training_regimen.train_loss_tracker.epoch_loss.sum() /
training_regimen.train_loss_tracker.epoch_words,
places=2)
def test_train_dev_loss_equal(self):
layer_dim = 512
batcher = SrcBatcher(batch_size=5, break_ties_randomly=False)
train_args = {}
train_args['src_file'] = "examples/data/head.ja"
train_args['trg_file'] = "examples/data/head.en"
train_args['loss_calculator'] = LossCalculator()
train_args['model'] = DefaultTranslator(
src_reader=PlainTextReader(),
trg_reader=PlainTextReader(),
src_embedder=SimpleWordEmbedder(emb_dim=layer_dim, vocab_size=100),
encoder=BiLSTMSeqTransducer(input_dim=layer_dim,
hidden_dim=layer_dim),
attender=MlpAttender(input_dim=layer_dim,
state_dim=layer_dim,
hidden_dim=layer_dim),
trg_embedder=SimpleWordEmbedder(emb_dim=layer_dim, vocab_size=100),
decoder=MlpSoftmaxDecoder(input_dim=layer_dim,
lstm_dim=layer_dim,
mlp_hidden_dim=layer_dim,
trg_embed_dim=layer_dim,
vocab_size=100,
bridge=CopyBridge(dec_layers=1,
dec_dim=layer_dim)),
)
train_args['dev_tasks'] = [
LossEvalTask(model=train_args['model'],
src_file="examples/data/head.ja",
ref_file="examples/data/head.en",
batcher=batcher)
]
train_args['trainer'] = None
train_args['batcher'] = batcher
train_args['run_for_epochs'] = 1
training_regimen = xnmt.training_regimen.SimpleTrainingRegimen(
**train_args)
training_regimen.run_training(save_fct=lambda: None,
update_weights=False)
self.assertAlmostEqual(training_regimen.logger.epoch_loss.sum() /
training_regimen.logger.epoch_words,
training_regimen.logger.dev_score.loss,
places=5)
def test_overfitting(self):
self.exp_global = ExpGlobal(
dynet_param_collection=NonPersistentParamCollection(), dropout=0.0)
self.exp_global.default_layer_dim = 16
batcher = SrcBatcher(batch_size=10, break_ties_randomly=False)
train_args = {}
train_args['src_file'] = "examples/data/head.ja"
train_args['trg_file'] = "examples/data/head.en"
train_args['loss_calculator'] = LossCalculator()
train_args['model'] = DefaultTranslator(
src_reader=PlainTextReader(),
trg_reader=PlainTextReader(),
src_embedder=SimpleWordEmbedder(self.exp_global, vocab_size=100),
encoder=BiLSTMSeqTransducer(self.exp_global),
attender=MlpAttender(self.exp_global),
trg_embedder=SimpleWordEmbedder(self.exp_global, vocab_size=100),
decoder=MlpSoftmaxDecoder(self.exp_global,
vocab_size=100,
bridge=CopyBridge(
exp_global=self.exp_global,
dec_layers=1)),
)
train_args['dev_tasks'] = [
LossEvalTask(model=train_args['model'],
src_file="examples/data/head.ja",
ref_file="examples/data/head.en",
batcher=batcher)
]
train_args['run_for_epochs'] = 1
train_args['trainer'] = AdamTrainer(self.exp_global, alpha=0.1)
train_args['batcher'] = batcher
training_regimen = xnmt.training_regimen.SimpleTrainingRegimen(
exp_global=self.exp_global, **train_args)
training_regimen.exp_global = self.exp_global
for _ in range(50):
training_regimen.run_training(save_fct=lambda: None,
update_weights=True)
self.assertAlmostEqual(0.0,
training_regimen.logger.epoch_loss.sum() /
training_regimen.logger.epoch_words,
places=2)
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)
EXP_DIR = os.path.dirname(__file__)
EXP = "programmatic"
model_file = f"{EXP_DIR}/models/{EXP}.mod"
log_file = f"{EXP_DIR}/logs/{EXP}.log"
xnmt.tee.set_out_file(log_file)
ParamManager.init_param_col()
ParamManager.param_col.model_file = model_file
src_vocab = Vocab(vocab_file="examples/data/head.ja.vocab")
trg_vocab = Vocab(vocab_file="examples/data/head.en.vocab")
batcher = SrcBatcher(batch_size=64)
inference = SimpleInference(batcher=batcher)
layer_dim = 512
model = DefaultTranslator(
src_reader=PlainTextReader(vocab=src_vocab),
trg_reader=PlainTextReader(vocab=trg_vocab),
src_embedder=SimpleWordEmbedder(emb_dim=layer_dim,
vocab_size=len(src_vocab)),
encoder=BiLSTMSeqTransducer(input_dim=layer_dim,
hidden_dim=layer_dim,
layers=1),
attender=MlpAttender(hidden_dim=layer_dim,
state_dim=layer_dim,
def __init__(self,
yaml_context,
corpus_parser,
model,
glob={},
dev_every=0,
batcher=None,
loss_calculator=None,
pretrained_model_file="",
src_format="text",
run_for_epochs=None,
lr_decay=1.0,
lr_decay_times=3,
patience=1,
initial_patience=None,
dev_metrics="",
schedule_metric="loss",
restart_trainer=False,
reload_command=None,
name=None,
inference=None):
"""
:param yaml_context:
:param corpus_parser: an input.InputReader object
:param model: a generator.GeneratorModel object
: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 loss_calculator:
:param pretrained_model_file: Path of pre-trained model file
:param src_format: Format of input data: text/contvec
:param lr_decay (float):
:param lr_decay_times (int): Early stopping after decaying learning rate a certain number of times
:param patience (int): apply LR decay after dev scores haven't improved over this many checkpoints
:param initial_patience (int): if given, allows adjusting patience for the first LR decay
: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'.
:param name: will be prepended to log outputs if given
:param inference: used for inference during dev checkpoints if dev_metrics are specified
"""
assert yaml_context is not None
self.yaml_context = yaml_context
self.model_file = self.yaml_context.dynet_param_collection.model_file
self.yaml_serializer = YamlSerializer()
if lr_decay > 1.0 or lr_decay <= 0.0:
raise RuntimeError(
"illegal lr_decay, must satisfy: 0.0 < lr_decay <= 1.0")
self.lr_decay = lr_decay
self.patience = patience
self.initial_patience = initial_patience
self.lr_decay_times = lr_decay_times
self.restart_trainer = restart_trainer
self.run_for_epochs = run_for_epochs
self.early_stopping_reached = False
# training state
self.training_state = TrainingState()
self.evaluators = [
s.lower() for s in dev_metrics.split(",") if s.strip() != ""
]
if schedule_metric.lower() not in self.evaluators:
self.evaluators.append(schedule_metric.lower())
if "loss" not in self.evaluators: self.evaluators.append("loss")
if dev_metrics:
self.inference = inference or SimpleInference()
self.reload_command = reload_command
if reload_command is not None:
self._augmentation_handle = None
self._augment_data_initial()
self.model = model
self.corpus_parser = corpus_parser
self.loss_calculator = loss_calculator or LossCalculator(MLELoss())
self.pretrained_model_file = pretrained_model_file
if self.pretrained_model_file:
self.yaml_context.dynet_param_collection.load_from_data_file(
self.pretrained_model_file + '.data')
self.batcher = batcher or SrcBatcher(32)
if src_format == "contvec":
self.batcher.pad_token = np.zeros(self.model.src_embedder.emb_dim)
self.pack_batches()
self.logger = BatchLossTracker(self, dev_every, name)
self.schedule_metric = schedule_metric.lower()