def init_train_from_config(self, config, train_data, dev_data=None, eval_data=None):
"""
:type config: Config.Config
:type train_data: Dataset.Dataset
:type dev_data: Dataset.Dataset | None
:type eval_data: Dataset.Dataset | None
"""
self.train_data = train_data
self.dev_data = dev_data
self.eval_data = eval_data
self.start_epoch, self.start_batch = self.get_train_start_epoch_batch(config)
self.batch_size = config.int('batch_size', 1)
self.shuffle_batches = config.bool('shuffle_batches', True)
self.update_batch_size = config.int('update_batch_size', 0)
self.model_filename = config.value('model', None)
self.save_model_epoch_interval = config.int('save_interval', 1)
self.save_epoch1_initial_model = config.bool('save_epoch1_initial_model', False)
self.learning_rate_control = loadLearningRateControlFromConfig(config)
self.learning_rate = self.learning_rate_control.defaultLearningRate
self.initial_learning_rate = self.learning_rate
self.pretrain_learning_rate = config.float('pretrain_learning_rate', self.learning_rate)
self.final_epoch = self.config_get_final_epoch(config) # Inclusive.
self.max_seqs = config.int('max_seqs', -1)
self.updater = Updater.initFromConfig(config)
self.ctc_prior_file = config.value('ctc_prior_file', None)
self.exclude = config.int_list('exclude', [])
self.init_train_epoch_posthook = config.value('init_train_epoch_posthook', None)
self.share_batches = config.bool('share_batches', False)
self.batch_variance = config.float('batch_variance', 0.0)
self.max_seq_length = config.float('max_seq_length', 0)
self.inc_seq_length = config.float('inc_seq_length', 0)
if self.max_seq_length == 0:
self.max_seq_length = sys.maxint
# And also initialize the network. That depends on some vars here such as pretrain.
self.init_network_from_config(config)
def init_train_from_config(self, config, train_data, dev_data, eval_data):
"""
:type config: Config.Config
:type train_data: Dataset.Dataset
:type dev_data: Dataset.Dataset | None
:type eval_data: Dataset.Dataset | None
"""
self.use_dynamic_train_flag = True
self.train_data = train_data
self.dev_data = dev_data
self.eval_data = eval_data
self.start_epoch, self.start_batch = self.get_train_start_epoch_batch(config)
self.batch_size = config.int('batch_size', 1)
self.shuffle_batches = config.bool('shuffle_batches', True)
self.update_batch_size = config.int('update_batch_size', 0)
self.save_model_epoch_interval = config.int('save_interval', 1)
self.save_epoch1_initial_model = config.bool('save_epoch1_initial_model', False)
self.learning_rate_control = loadLearningRateControlFromConfig(config)
self.learning_rate = self.learning_rate_control.defaultLearningRate
self.initial_learning_rate = self.learning_rate
self.pretrain_learning_rate = config.float('pretrain_learning_rate', self.learning_rate)
self.final_epoch = self.config_get_final_epoch(config) # Inclusive.
self.max_seqs = config.int('max_seqs', -1)
self.ctc_prior_file = config.value('ctc_prior_file', None)
self.exclude = config.int_list('exclude', [])
self.init_train_epoch_posthook = config.value('init_train_epoch_posthook', None)
self.share_batches = config.bool('share_batches', False)
self.seq_drop = config.float('seq_drop', 0.0)
self.seq_drop_freq = config.float('seq_drop_freq', 10)
self.max_seq_length = config.float('max_seq_length', 0)
self.inc_seq_length = config.float('inc_seq_length', 0)
if self.max_seq_length == 0:
self.max_seq_length = sys.maxsize
# And also initialize the network. That depends on some vars here such as pretrain.
self.init_network_from_config(config)
def test_init_error_new():
config = Config()
config.update({
"learning_rate_control": "newbob",
"learning_rate_control_error_measure": "dev_score"
})
lrc = loadLearningRateControlFromConfig(config)
assert isinstance(lrc, NewbobRelative)
lrc.getLearningRateForEpoch(1)
lrc.setEpochError(1, {"train_score": {'cost:output': 1.9344199658230012}})
lrc.setEpochError(1, {
"dev_score": {
'cost:output': 1.99
},
"dev_error": {
'error:output': 0.6
}
})
error = lrc.getEpochErrorDict(1)
assert "train_score" in error
assert "dev_score" in error
assert "dev_error" in error
assert_equal(lrc.getErrorKey(1), "dev_score")
lrc.getLearningRateForEpoch(2)
lrc.setEpochError(2, {"train_score": {'cost:output': 1.8}})
lrc.setEpochError(2, {
"dev_score": {
'cost:output': 1.9
},
"dev_error": {
'error:output': 0.5
}
})
lrc.getLearningRateForEpoch(3)
def init_train_from_config(self,
config,
train_data,
dev_data=None,
eval_data=None):
"""
:type config: Config.Config
:type train_data: Dataset.Dataset
:type dev_data: Dataset.Dataset | None
:type eval_data: Dataset.Dataset | None
"""
self.train_data = train_data
self.dev_data = dev_data
self.eval_data = eval_data
self.start_epoch, self.start_batch = self.get_train_start_epoch_batch(
config)
self.batch_size = config.int('batch_size', 1)
self.shuffle_batches = config.bool('shuffle_batches', False)
self.update_batch_size = config.int('update_batch_size', 0)
self.batch_size_eval = config.int('batch_size_eval',
self.update_batch_size)
self.model_filename = config.value('model', None)
self.save_model_epoch_interval = config.int('save_interval', 1)
self.save_epoch1_initial_model = config.bool(
'save_epoch1_initial_model', False)
self.learning_rate_control = loadLearningRateControlFromConfig(config)
self.learning_rate = self.learning_rate_control.defaultLearningRate
self.initial_learning_rate = self.learning_rate
self.pretrain_learning_rate = config.float('pretrain_learning_rate',
self.learning_rate)
self.final_epoch = self.config_get_final_epoch(config) # Inclusive.
self.max_seqs = config.int('max_seqs', -1)
self.max_seqs_eval = config.int('max_seqs_eval', self.max_seqs)
self.updater = Updater.initFromConfig(config)
self.ctc_prior_file = config.value('ctc_prior_file', None)
self.exclude = config.int_list('exclude', [])
self.init_train_epoch_posthook = config.value(
'init_train_epoch_posthook', None)
self.share_batches = config.bool('share_batches', False)
self.seq_drop = config.float('seq_drop', 0.0)
self.seq_drop_freq = config.float('seq_drop_freq', 10)
self.max_seq_length = config.float('max_seq_length', 0)
self.inc_seq_length = config.float('inc_seq_length', 0)
self.max_seq_length_eval = config.int('max_seq_length_eval', 2e31)
self.output_precision = config.int('output_precision', 12)
self.reduction_rate = config.float('reduction_rate', 1.0)
self.batch_pruning = config.float('batch_pruning', 0.0)
if self.max_seq_length == 0:
self.max_seq_length = sys.maxsize
if config.is_typed("seq_train_parallel"):
self.seq_train_parallel = SeqTrainParallelControl(
engine=self,
config=config,
**config.typed_value("seq_train_parallel"))
else:
self.seq_train_parallel = None
# And also initialize the network. That depends on some vars here such as pretrain.
self.init_network_from_config(config)
def test_init_error_old():
config = Config()
config.update({"learning_rate_control": "newbob", "learning_rate_control_error_measure": "dev_score"})
lrc = loadLearningRateControlFromConfig(config)
assert isinstance(lrc, NewbobRelative)
lrc.getLearningRateForEpoch(1)
lrc.setEpochError(1, {"train_score": 1.9344199658230012})
lrc.setEpochError(1, {"dev_score": 1.99, "dev_error": 0.6})
error = lrc.getEpochErrorDict(1)
assert "train_score" in error
assert "dev_score" in error
assert "dev_error" in error
assert_equal(lrc.getErrorKey(1), "dev_score")
lrc.getLearningRateForEpoch(2)
lrc.setEpochError(2, {"train_score": 1.8})
lrc.setEpochError(2, {"dev_score": 1.9, "dev_error": 0.5})
lrc.getLearningRateForEpoch(3)
def test_init_error_muliple_out():
config = Config()
config.update({"learning_rate_control": "newbob", "learning_rate_control_error_measure": "dev_score"})
lrc = loadLearningRateControlFromConfig(config)
assert isinstance(lrc, NewbobRelative)
lrc.getLearningRateForEpoch(1)
lrc.setEpochError(1, {"train_score": {'cost:output': 1.95, "cost:out2": 2.95}})
lrc.setEpochError(1, {"dev_score": {'cost:output': 1.99, "cost:out2": 2.99},
"dev_error": {'error:output': 0.6, "error:out2": 0.7}})
error = lrc.getEpochErrorDict(1)
assert "train_score_output" in error
assert "train_score_out2" in error
assert "dev_score_output" in error
assert "dev_score_out2" in error
assert "dev_error_output" in error
assert "dev_error_out2" in error
assert_equal(lrc.getErrorKey(1), "dev_score_output")
lrc.getLearningRateForEpoch(2)
lrc.setEpochError(2, {"train_score": {'cost:output': 1.8, "cost:out2": 2.8}})
lrc.setEpochError(2, {"dev_score": {'cost:output': 1.9, "cost:out2": 2.9},
"dev_error": {'error:output': 0.5, "error:out2": 0.6}})
lrc.getLearningRateForEpoch(3)
