def _setup_context(args, exit_stack):
if args.use_cpu:
context = mx.cpu()
else:
num_gpus = get_num_gpus()
check_condition(
num_gpus >= 1,
"No GPUs found, consider running on the CPU with --use-cpu "
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi "
"binary isn't on the path).")
check_condition(
len(args.device_ids) == 1,
"cannot run on multiple devices for now")
gpu_id = args.device_ids[0]
if args.disable_device_locking:
# without locking and a negative device id we just take the first device
gpu_id = 0
else:
if gpu_id < 0:
# get a single (!) gpu id automatically:
gpu_ids = exit_stack.enter_context(
acquire_gpus([-1], lock_dir=args.lock_dir))
gpu_id = gpu_ids[0]
context = mx.gpu(gpu_id)
return context
def main():
params = argparse.ArgumentParser(description='Translate CLI')
arguments.add_translate_cli_args(params)
arguments.add_bpe_args(params)
args = params.parse_args()
with ExitStack() as exit_stack:
if args.use_cpu:
context = mx.cpu()
else:
num_gpus = get_num_gpus()
check_condition(num_gpus >= 1,
"No GPUs found, consider running on the CPU with --use-cpu "
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi "
"binary isn't on the path).")
check_condition(len(args.device_ids) == 1, "cannot run on multiple devices for now")
gpu_id = args.device_ids[0]
if args.disable_device_locking:
if gpu_id < 0:
# without locking and a negative device id we just take the first device
gpu_id = 0
else:
gpu_ids = exit_stack.enter_context(acquire_gpus([gpu_id], lock_dir=args.lock_dir))
gpu_id = gpu_ids[0]
context = mx.gpu(gpu_id)
models, source_vocabs, target_vocab = inference.load_models(
context=context,
max_input_len=args.max_input_len,
beam_size=args.beam_size,
batch_size=args.batch_size,
model_folders=args.models,
checkpoints=args.checkpoints,
softmax_temperature=args.softmax_temperature,
max_output_length_num_stds=args.max_output_length_num_stds,
decoder_return_logit_inputs=args.restrict_lexicon is not None,
cache_output_layer_w_b=args.restrict_lexicon is not None)
translator = inference.Translator(context=context,
ensemble_mode=args.ensemble_mode,
bucket_source_width=args.bucket_width,
length_penalty=inference.LengthPenalty(args.length_penalty_alpha,
args.length_penalty_beta),
models=models,
source_vocabs=source_vocabs,
target_vocab=target_vocab,
restrict_lexicon=None,
store_beam=False,
strip_unknown_words=args.strip_unknown_words)
logger.info('Parsing vocabulary')
sys.stdin = io.TextIOWrapper(sys.stdin.buffer, encoding='utf-8')
sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding='utf-8')
sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8', write_through=True, line_buffering=True)
opened_vocab = codecs.open(args.bpe_vocabulary.name, encoding='utf-8')
bpe_filtered_vocab = read_vocabulary(opened_vocab, args.bpe_vocabulary_threshold)
bpe_merges = -1 # Apply all merge operations.
bpe_separator = '@@' # Use default BPE separator.
bpe_glossaries = None # No excluded words.
bpe = BPE(args.bpe_codes, bpe_merges, bpe_separator, bpe_filtered_vocab, bpe_glossaries)
logger.info('Starting RPC server.')
rpc_server = SockeyeRpcServer(translator, bpe)
rpc_server.serve()
def main():
params = argparse.ArgumentParser(
description='Translate from STDIN to STDOUT')
params.add_argument(
'--model-prefixes',
'-m',
required=False,
nargs='+',
help='model prefix(es). Use multiple for ensemble decoding. ' +
'Model prefix determines config, best epoch params and vocab files.')
params.add_argument(
'--epochs',
'-e',
required=False,
default=None,
type=int,
nargs='+',
help=
'If not given, chooses best epochs/checkpoints for each model. If specified, must have the same '
+ 'length as --model-prefix and be integer')
params.add_argument(
'--max-input-len',
'-n',
type=int,
default=None,
help='Maximum sentence length. Default: value from trained model(s).')
params.add_argument(
'--output-type',
default='translation',
choices=["translation", "align_plot", "align_text"],
help=
'Either print the translation or visualize the alignment. Default: translation'
)
params.add_argument('--align-plot-prefix',
default="align",
help='Prefix used when plotting the alignment.')
params.add_argument('--log-level',
default=logging.INFO,
type=int,
choices=[logging.INFO, logging.WARN, logging.DEBUG])
params.add_argument('--beam-size',
'-b',
type=int,
default=1,
help='beam size. If == 1, greedy decode')
params.add_argument(
'--ensemble-mode',
type=str,
default='linear',
choices=['linear', 'log_linear'],
help=
'Ensemble mode: linear or log-linear interpolation of model predictions. Default: linear'
)
params.add_argument(
'--softmax-temperature',
type=float,
default=None,
required=False,
help=
'Controls peakiness of model predictions. Values < 1.0 produce peaked predictions, '
+ 'values > 1.0 produce smoothed distributions.')
params = arguments.add_device_args(params)
args = params.parse_args()
args.model_prefixes = ['model/']
assert args.beam_size > 0, "Beam size must be 1 or greater."
if args.epochs is not None:
assert len(args.epochs) == len(
args.model_prefixes), "must provide epochs for each model"
#sockeye.utils.setup_logging(args.log_level)
logging.basicConfig(filename='test.log', level=logging.INFO)
logging.info("Command: %s", " ".join(sys.argv))
logging.info("Arguments: %s", args)
with ExitStack() as exit_stack:
if args.use_cpu:
context = mx.cpu()
else:
num_gpus = get_num_gpus()
assert num_gpus > 0, "No GPUs found, consider running on the CPU with --use-cpu " \
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi " \
"binary isn't on the path)."
assert len(
args.device_ids) == 1, "cannot run on multiple devices for now"
gpu_id = args.device_ids[0]
if gpu_id < 0:
# get an automatic gpu id:
gpu_id = exit_stack.enter_context(acquire_gpu())
context = mx.gpu(gpu_id)
translator = sockeye.inference.Translator(
context, args.ensemble_mode,
*sockeye.inference.load_models(context, args.max_input_len,
args.beam_size, args.model_prefixes,
args.epochs,
args.softmax_temperature))
############ CHANGE HERE ########################
sample_file = open('train_question_token.txt', 'r')
encoder_file = open('train_question_encoder.txt', "w")
#################################################
for i, line in enumerate(sample_file, 1):
trans_input = translator.make_input(i, line)
source, source_length, bucket_key = translator._get_inference_input(
trans_input.tokens)
encoded_source, _, _, _, _ = translator.models[0].run_encoder(
source, source_length, bucket_key)
last_slice_source = mx.ndarray.mean(encoded_source,
axis=1,
keepdims=True)
last_slice_source = last_slice_source.reshape((-1, ))
encoder_file.write(
" ".join(map(str, last_slice_source.asnumpy())) + "\n")
encoder_file.close()
def main():
params = argparse.ArgumentParser(description='CLI to train sockeye sequence-to-sequence models.')
arguments.add_train_cli_args(params)
args = params.parse_args()
# seed the RNGs
np.random.seed(args.seed)
random.seed(args.seed)
mx.random.seed(args.seed)
if args.use_fused_rnn:
check_condition(not args.use_cpu, "GPU required for FusedRNN cells")
check_condition(args.optimized_metric == C.BLEU or args.optimized_metric in args.metrics,
"Must optimize either BLEU or one of tracked metrics (--metrics)")
# Checking status of output folder, resumption, etc.
# Create temporary logger to console only
logger = setup_main_logger(__name__, file_logging=False, console=not args.quiet)
output_folder = os.path.abspath(args.output)
resume_training = False
training_state_dir = os.path.join(output_folder, C.TRAINING_STATE_DIRNAME)
if os.path.exists(output_folder):
if args.overwrite_output:
logger.info("Removing existing output folder %s.", output_folder)
shutil.rmtree(output_folder)
os.makedirs(output_folder)
elif os.path.exists(training_state_dir):
with open(os.path.join(output_folder, C.ARGS_STATE_NAME), "r") as fp:
old_args = json.load(fp)
arg_diffs = _dict_difference(vars(args), old_args) | _dict_difference(old_args, vars(args))
# Remove args that may differ without affecting the training.
arg_diffs -= set(C.ARGS_MAY_DIFFER)
# allow different device-ids provided their total count is the same
if 'device_ids' in arg_diffs and len(old_args['device_ids']) == len(vars(args)['device_ids']):
arg_diffs.discard('device_ids')
if not arg_diffs:
resume_training = True
else:
# We do not have the logger yet
logger.error("Mismatch in arguments for training continuation.")
logger.error("Differing arguments: %s.", ", ".join(arg_diffs))
sys.exit(1)
elif os.path.exists(os.path.join(output_folder, C.PARAMS_BEST_NAME)):
logger.error("Refusing to overwrite model folder %s as it seems to contain a trained model.", output_folder)
sys.exit(1)
else:
logger.info("The output folder %s already exists, but no training state or parameter file was found. "
"Will start training from scratch.", output_folder)
else:
os.makedirs(output_folder)
logger = setup_main_logger(__name__,
file_logging=True,
console=not args.quiet, path=os.path.join(output_folder, C.LOG_NAME))
log_sockeye_version(logger)
log_mxnet_version(logger)
logger.info("Command: %s", " ".join(sys.argv))
logger.info("Arguments: %s", args)
with open(os.path.join(output_folder, C.ARGS_STATE_NAME), "w") as fp:
json.dump(vars(args), fp)
with ExitStack() as exit_stack:
# context
if args.use_cpu:
logger.info("Device: CPU")
context = [mx.cpu()]
else:
num_gpus = get_num_gpus()
check_condition(num_gpus >= 1,
"No GPUs found, consider running on the CPU with --use-cpu "
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi "
"binary isn't on the path).")
if args.disable_device_locking:
context = expand_requested_device_ids(args.device_ids)
else:
context = exit_stack.enter_context(acquire_gpus(args.device_ids, lock_dir=args.lock_dir))
logger.info("Device(s): GPU %s", context)
context = [mx.gpu(gpu_id) for gpu_id in context]
# load existing or create vocabs
if resume_training:
vocab_source = vocab.vocab_from_json_or_pickle(os.path.join(output_folder, C.VOCAB_SRC_NAME))
vocab_target = vocab.vocab_from_json_or_pickle(os.path.join(output_folder, C.VOCAB_TRG_NAME))
else:
num_words_source, num_words_target = args.num_words
word_min_count_source, word_min_count_target = args.word_min_count
# if the source and target embeddings are tied we build a joint vocabulary:
if args.weight_tying and C.WEIGHT_TYING_SRC in args.weight_tying_type \
and C.WEIGHT_TYING_TRG in args.weight_tying_type:
vocab_source = vocab_target = _build_or_load_vocab(args.source_vocab,
[args.source, args.target],
num_words_source,
word_min_count_source)
else:
vocab_source = _build_or_load_vocab(args.source_vocab, [args.source],
num_words_source, word_min_count_source)
vocab_target = _build_or_load_vocab(args.target_vocab, [args.target],
num_words_target, word_min_count_target)
# write vocabularies
vocab.vocab_to_json(vocab_source, os.path.join(output_folder, C.VOCAB_SRC_NAME) + C.JSON_SUFFIX)
vocab.vocab_to_json(vocab_target, os.path.join(output_folder, C.VOCAB_TRG_NAME) + C.JSON_SUFFIX)
vocab_source_size = len(vocab_source)
vocab_target_size = len(vocab_target)
logger.info("Vocabulary sizes: source=%d target=%d", vocab_source_size, vocab_target_size)
# create data iterators
max_seq_len_source, max_seq_len_target = args.max_seq_len
batch_num_devices = 1 if args.use_cpu else sum(-di if di < 0 else 1 for di in args.device_ids)
train_iter, eval_iter, config_data = data_io.get_training_data_iters(source=os.path.abspath(args.source),
target=os.path.abspath(args.target),
validation_source=os.path.abspath(
args.validation_source),
validation_target=os.path.abspath(
args.validation_target),
vocab_source=vocab_source,
vocab_target=vocab_target,
vocab_source_path=args.source_vocab,
vocab_target_path=args.target_vocab,
batch_size=args.batch_size,
batch_by_words=args.batch_type == C.BATCH_TYPE_WORD,
batch_num_devices=batch_num_devices,
fill_up=args.fill_up,
max_seq_len_source=max_seq_len_source,
max_seq_len_target=max_seq_len_target,
bucketing=not args.no_bucketing,
bucket_width=args.bucket_width)
# learning rate scheduling
learning_rate_half_life = none_if_negative(args.learning_rate_half_life)
# TODO: The loading for continuation of the scheduler is done separately from the other parts
if not resume_training:
lr_scheduler_instance = lr_scheduler.get_lr_scheduler(args.learning_rate_scheduler_type,
args.checkpoint_frequency,
learning_rate_half_life,
args.learning_rate_reduce_factor,
args.learning_rate_reduce_num_not_improved,
args.learning_rate_schedule,
args.learning_rate_warmup)
else:
with open(os.path.join(training_state_dir, C.SCHEDULER_STATE_NAME), "rb") as fp:
lr_scheduler_instance = pickle.load(fp)
# model configuration
num_embed_source, num_embed_target = args.num_embed
encoder_num_layers, decoder_num_layers = args.num_layers
encoder_embed_dropout, decoder_embed_dropout = args.embed_dropout
encoder_rnn_dropout_inputs, decoder_rnn_dropout_inputs = args.rnn_dropout_inputs
encoder_rnn_dropout_states, decoder_rnn_dropout_states = args.rnn_dropout_states
if encoder_embed_dropout > 0 and encoder_rnn_dropout_inputs > 0:
logger.warning("Setting encoder RNN AND source embedding dropout > 0 leads to "
"two dropout layers on top of each other.")
if decoder_embed_dropout > 0 and decoder_rnn_dropout_inputs > 0:
logger.warning("Setting encoder RNN AND source embedding dropout > 0 leads to "
"two dropout layers on top of each other.")
encoder_rnn_dropout_recurrent, decoder_rnn_dropout_recurrent = args.rnn_dropout_recurrent
if encoder_rnn_dropout_recurrent > 0 or decoder_rnn_dropout_recurrent > 0:
check_condition(args.rnn_cell_type == C.LSTM_TYPE,
"Recurrent dropout without memory loss only supported for LSTMs right now.")
encoder_transformer_preprocess, decoder_transformer_preprocess = args.transformer_preprocess
encoder_transformer_postprocess, decoder_transformer_postprocess = args.transformer_postprocess
config_conv = None
if args.encoder == C.RNN_WITH_CONV_EMBED_NAME:
config_conv = encoder.ConvolutionalEmbeddingConfig(num_embed=num_embed_source,
max_filter_width=args.conv_embed_max_filter_width,
num_filters=args.conv_embed_num_filters,
pool_stride=args.conv_embed_pool_stride,
num_highway_layers=args.conv_embed_num_highway_layers,
dropout=args.conv_embed_dropout)
if args.encoder in (C.TRANSFORMER_TYPE, C.TRANSFORMER_WITH_CONV_EMBED_TYPE):
config_encoder = transformer.TransformerConfig(
model_size=args.transformer_model_size,
attention_heads=args.transformer_attention_heads,
feed_forward_num_hidden=args.transformer_feed_forward_num_hidden,
num_layers=encoder_num_layers,
vocab_size=vocab_source_size,
dropout_attention=args.transformer_dropout_attention,
dropout_relu=args.transformer_dropout_relu,
dropout_prepost=args.transformer_dropout_prepost,
weight_tying=args.weight_tying and C.WEIGHT_TYING_SRC in args.weight_tying_type,
positional_encodings=not args.transformer_no_positional_encodings,
preprocess_sequence=encoder_transformer_preprocess,
postprocess_sequence=encoder_transformer_postprocess,
conv_config=config_conv)
else:
config_encoder = encoder.RecurrentEncoderConfig(
vocab_size=vocab_source_size,
num_embed=num_embed_source,
embed_dropout=encoder_embed_dropout,
rnn_config=rnn.RNNConfig(cell_type=args.rnn_cell_type,
num_hidden=args.rnn_num_hidden,
num_layers=encoder_num_layers,
dropout_inputs=encoder_rnn_dropout_inputs,
dropout_states=encoder_rnn_dropout_states,
dropout_recurrent=encoder_rnn_dropout_recurrent,
residual=args.rnn_residual_connections,
first_residual_layer=args.rnn_first_residual_layer,
forget_bias=args.rnn_forget_bias),
conv_config=config_conv,
reverse_input=args.rnn_encoder_reverse_input)
decoder_weight_tying = args.weight_tying and C.WEIGHT_TYING_TRG in args.weight_tying_type \
and C.WEIGHT_TYING_SOFTMAX in args.weight_tying_type
if args.decoder == C.TRANSFORMER_TYPE:
config_decoder = transformer.TransformerConfig(
model_size=args.transformer_model_size,
attention_heads=args.transformer_attention_heads,
feed_forward_num_hidden=args.transformer_feed_forward_num_hidden,
num_layers=decoder_num_layers,
vocab_size=vocab_target_size,
dropout_attention=args.transformer_dropout_attention,
dropout_relu=args.transformer_dropout_relu,
dropout_prepost=args.transformer_dropout_prepost,
weight_tying=decoder_weight_tying,
positional_encodings=not args.transformer_no_positional_encodings,
preprocess_sequence=decoder_transformer_preprocess,
postprocess_sequence=decoder_transformer_postprocess,
conv_config=None)
else:
attention_num_hidden = args.rnn_num_hidden if not args.attention_num_hidden else args.attention_num_hidden
config_coverage = None
if args.attention_type == C.ATT_COV:
config_coverage = coverage.CoverageConfig(type=args.attention_coverage_type,
num_hidden=args.attention_coverage_num_hidden,
layer_normalization=args.layer_normalization)
config_attention = attention.AttentionConfig(type=args.attention_type,
num_hidden=attention_num_hidden,
input_previous_word=args.attention_use_prev_word,
rnn_num_hidden=args.rnn_num_hidden,
layer_normalization=args.layer_normalization,
config_coverage=config_coverage,
num_heads=args.attention_mhdot_heads)
config_decoder = decoder.RecurrentDecoderConfig(
vocab_size=vocab_target_size,
max_seq_len_source=max_seq_len_source,
num_embed=num_embed_target,
rnn_config=rnn.RNNConfig(cell_type=args.rnn_cell_type,
num_hidden=args.rnn_num_hidden,
num_layers=decoder_num_layers,
dropout_inputs=decoder_rnn_dropout_inputs,
dropout_states=decoder_rnn_dropout_states,
dropout_recurrent=decoder_rnn_dropout_recurrent,
residual=args.rnn_residual_connections,
first_residual_layer=args.rnn_first_residual_layer,
forget_bias=args.rnn_forget_bias),
attention_config=config_attention,
embed_dropout=decoder_embed_dropout,
hidden_dropout=args.rnn_decoder_hidden_dropout,
weight_tying=decoder_weight_tying,
state_init=args.rnn_decoder_state_init,
context_gating=args.rnn_context_gating,
layer_normalization=args.layer_normalization,
attention_in_upper_layers=args.attention_in_upper_layers)
config_loss = loss.LossConfig(type=args.loss,
vocab_size=vocab_target_size,
normalize=args.normalize_loss,
smoothed_cross_entropy_alpha=args.smoothed_cross_entropy_alpha)
model_config = model.ModelConfig(config_data=config_data,
max_seq_len_source=max_seq_len_source,
max_seq_len_target=max_seq_len_target,
vocab_source_size=vocab_source_size,
vocab_target_size=vocab_target_size,
config_encoder=config_encoder,
config_decoder=config_decoder,
config_loss=config_loss,
lexical_bias=args.lexical_bias,
learn_lexical_bias=args.learn_lexical_bias,
weight_tying=args.weight_tying,
weight_tying_type=args.weight_tying_type if args.weight_tying else None)
model_config.freeze()
# create training model
training_model = training.TrainingModel(config=model_config,
context=context,
train_iter=train_iter,
fused=args.use_fused_rnn,
bucketing=not args.no_bucketing,
lr_scheduler=lr_scheduler_instance)
# We may consider loading the params in TrainingModule, for consistency
# with the training state saving
if resume_training:
logger.info("Found partial training in directory %s. Resuming from saved state.", training_state_dir)
training_model.load_params_from_file(os.path.join(training_state_dir, C.TRAINING_STATE_PARAMS_NAME))
elif args.params:
logger.info("Training will initialize from parameters loaded from '%s'", args.params)
training_model.load_params_from_file(args.params)
lexicon_array = lexicon.initialize_lexicon(args.lexical_bias,
vocab_source, vocab_target) if args.lexical_bias else None
weight_initializer = initializer.get_initializer(args.weight_init, args.weight_init_scale,
args.rnn_h2h_init, lexicon=lexicon_array)
optimizer = args.optimizer
optimizer_params = {'wd': args.weight_decay,
"learning_rate": args.initial_learning_rate}
if lr_scheduler_instance is not None:
optimizer_params["lr_scheduler"] = lr_scheduler_instance
clip_gradient = none_if_negative(args.clip_gradient)
if clip_gradient is not None:
optimizer_params["clip_gradient"] = clip_gradient
if args.momentum is not None:
optimizer_params["momentum"] = args.momentum
if args.normalize_loss:
# When normalize_loss is turned on we normalize by the number of non-PAD symbols in a batch which implicitly
# already contains the number of sentences and therefore we need to disable rescale_grad.
optimizer_params["rescale_grad"] = 1.0
else:
# Making MXNet module API's default scaling factor explicit
optimizer_params["rescale_grad"] = 1.0 / args.batch_size
logger.info("Optimizer: %s", optimizer)
logger.info("Optimizer Parameters: %s", optimizer_params)
# Handle options that override training settings
max_updates = args.max_updates
max_num_checkpoint_not_improved = args.max_num_checkpoint_not_improved
min_num_epochs = args.min_num_epochs
# Fixed training schedule always runs for a set number of updates
if args.learning_rate_schedule:
max_updates = sum(num_updates for (_, num_updates) in args.learning_rate_schedule)
max_num_checkpoint_not_improved = -1
min_num_epochs = 0
monitor_bleu = args.monitor_bleu
# Turn on BLEU monitoring when the optimized metric is BLEU and it hasn't been enabled yet
if args.optimized_metric == C.BLEU and monitor_bleu == 0:
logger.info("You chose BLEU as the optimized metric, will turn on BLEU monitoring during training. "
"To control how many validation sentences are used for calculating bleu use "
"the --monitor-bleu argument.")
monitor_bleu = -1
training_model.fit(train_iter, eval_iter,
output_folder=output_folder,
max_params_files_to_keep=args.keep_last_params,
metrics=args.metrics,
initializer=weight_initializer,
max_updates=max_updates,
checkpoint_frequency=args.checkpoint_frequency,
optimizer=optimizer, optimizer_params=optimizer_params,
optimized_metric=args.optimized_metric,
max_num_not_improved=max_num_checkpoint_not_improved,
min_num_epochs=min_num_epochs,
monitor_bleu=monitor_bleu,
use_tensorboard=args.use_tensorboard,
mxmonitor_pattern=args.monitor_pattern,
mxmonitor_stat_func=args.monitor_stat_func)
def main():
params = argparse.ArgumentParser(
description='CLI to train sockeye sequence-to-sequence models.')
arguments.add_io_args(params)
arguments.add_model_parameters(params)
arguments.add_training_args(params)
arguments.add_device_args(params)
args = params.parse_args()
# seed the RNGs
np.random.seed(args.seed)
random.seed(args.seed)
mx.random.seed(args.seed)
if args.use_fused_rnn:
check_condition(not args.use_cpu, "GPU required for FusedRNN cells")
if args.rnn_residual_connections:
check_condition(args.rnn_num_layers > 2,
"Residual connections require at least 3 RNN layers")
check_condition(
args.optimized_metric == C.BLEU
or args.optimized_metric in args.metrics,
"Must optimize either BLEU or one of tracked metrics (--metrics)")
# Checking status of output folder, resumption, etc.
# Create temporary logger to console only
logger = setup_main_logger(__name__,
file_logging=False,
console=not args.quiet)
output_folder = os.path.abspath(args.output)
resume_training = False
training_state_dir = os.path.join(output_folder, C.TRAINING_STATE_DIRNAME)
if os.path.exists(output_folder):
if args.overwrite_output:
logger.info("Removing existing output folder %s.", output_folder)
shutil.rmtree(output_folder)
os.makedirs(output_folder)
elif os.path.exists(training_state_dir):
with open(os.path.join(output_folder, C.ARGS_STATE_NAME),
"r") as fp:
old_args = json.load(fp)
arg_diffs = _dict_difference(
vars(args), old_args) | _dict_difference(old_args, vars(args))
# Remove args that may differ without affecting the training.
arg_diffs -= set(C.ARGS_MAY_DIFFER)
# allow different device-ids provided their total count is the same
if 'device_ids' in arg_diffs and len(
old_args['device_ids']) == len(vars(args)['device_ids']):
arg_diffs.discard('device_ids')
if not arg_diffs:
resume_training = True
else:
# We do not have the logger yet
logger.error(
"Mismatch in arguments for training continuation.")
logger.error("Differing arguments: %s.", ", ".join(arg_diffs))
sys.exit(1)
else:
logger.error("Refusing to overwrite existing output folder %s.",
output_folder)
sys.exit(1)
else:
os.makedirs(output_folder)
logger = setup_main_logger(__name__,
file_logging=True,
console=not args.quiet,
path=os.path.join(output_folder, C.LOG_NAME))
log_sockeye_version(logger)
logger.info("Command: %s", " ".join(sys.argv))
logger.info("Arguments: %s", args)
with open(os.path.join(output_folder, C.ARGS_STATE_NAME), "w") as fp:
json.dump(vars(args), fp)
with ExitStack() as exit_stack:
# context
if args.use_cpu:
logger.info("Device: CPU")
context = [mx.cpu()]
else:
num_gpus = get_num_gpus()
check_condition(
num_gpus >= 1,
"No GPUs found, consider running on the CPU with --use-cpu "
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi "
"binary isn't on the path).")
if args.disable_device_locking:
context = expand_requested_device_ids(args.device_ids)
else:
context = exit_stack.enter_context(
acquire_gpus(args.device_ids, lock_dir=args.lock_dir))
logger.info("Device(s): GPU %s", context)
context = [mx.gpu(gpu_id) for gpu_id in context]
# load existing or create vocabs
if resume_training:
vocab_source = vocab.vocab_from_json_or_pickle(
os.path.join(output_folder, C.VOCAB_SRC_NAME))
vocab_target = vocab.vocab_from_json_or_pickle(
os.path.join(output_folder, C.VOCAB_TRG_NAME))
else:
num_words_source = args.num_words if args.num_words_source is None else args.num_words_source
vocab_source = _build_or_load_vocab(args.source_vocab, args.source,
num_words_source,
args.word_min_count)
vocab.vocab_to_json(
vocab_source,
os.path.join(output_folder, C.VOCAB_SRC_NAME) + C.JSON_SUFFIX)
num_words_target = args.num_words if args.num_words_target is None else args.num_words_target
vocab_target = _build_or_load_vocab(args.target_vocab, args.target,
num_words_target,
args.word_min_count)
vocab.vocab_to_json(
vocab_target,
os.path.join(output_folder, C.VOCAB_TRG_NAME) + C.JSON_SUFFIX)
vocab_source_size = len(vocab_source)
vocab_target_size = len(vocab_target)
logger.info("Vocabulary sizes: source=%d target=%d", vocab_source_size,
vocab_target_size)
config_data = data_io.DataConfig(
os.path.abspath(args.source), os.path.abspath(args.target),
os.path.abspath(args.validation_source),
os.path.abspath(args.validation_target), args.source_vocab,
args.target_vocab)
# create data iterators
max_seq_len_source = args.max_seq_len if args.max_seq_len_source is None else args.max_seq_len_source
max_seq_len_target = args.max_seq_len if args.max_seq_len_target is None else args.max_seq_len_target
train_iter, eval_iter = data_io.get_training_data_iters(
source=config_data.source,
target=config_data.target,
validation_source=config_data.validation_source,
validation_target=config_data.validation_target,
vocab_source=vocab_source,
vocab_target=vocab_target,
batch_size=args.batch_size,
fill_up=args.fill_up,
max_seq_len_source=max_seq_len_source,
max_seq_len_target=max_seq_len_target,
bucketing=not args.no_bucketing,
bucket_width=args.bucket_width)
# learning rate scheduling
learning_rate_half_life = none_if_negative(
args.learning_rate_half_life)
# TODO: The loading for continuation of the scheduler is done separately from the other parts
if not resume_training:
lr_scheduler_instance = lr_scheduler.get_lr_scheduler(
args.learning_rate_scheduler_type, args.checkpoint_frequency,
learning_rate_half_life, args.learning_rate_reduce_factor,
args.learning_rate_reduce_num_not_improved)
else:
with open(os.path.join(training_state_dir, C.SCHEDULER_STATE_NAME),
"rb") as fp:
lr_scheduler_instance = pickle.load(fp)
# model configuration
num_embed_source = args.num_embed if args.num_embed_source is None else args.num_embed_source
num_embed_target = args.num_embed if args.num_embed_target is None else args.num_embed_target
config_rnn = rnn.RNNConfig(cell_type=args.rnn_cell_type,
num_hidden=args.rnn_num_hidden,
num_layers=args.rnn_num_layers,
dropout=args.dropout,
residual=args.rnn_residual_connections,
forget_bias=args.rnn_forget_bias)
config_conv = None
if args.encoder == C.RNN_WITH_CONV_EMBED_NAME:
config_conv = encoder.ConvolutionalEmbeddingConfig(
num_embed=num_embed_source,
max_filter_width=args.conv_embed_max_filter_width,
num_filters=args.conv_embed_num_filters,
pool_stride=args.conv_embed_pool_stride,
num_highway_layers=args.conv_embed_num_highway_layers,
dropout=args.dropout)
config_encoder = encoder.RecurrentEncoderConfig(
vocab_size=vocab_source_size,
num_embed=num_embed_source,
rnn_config=config_rnn,
conv_config=config_conv)
config_decoder = decoder.RecurrentDecoderConfig(
vocab_size=vocab_target_size,
num_embed=num_embed_target,
rnn_config=config_rnn,
dropout=args.dropout,
weight_tying=args.weight_tying,
context_gating=args.context_gating,
layer_normalization=args.layer_normalization)
attention_num_hidden = args.rnn_num_hidden if not args.attention_num_hidden else args.attention_num_hidden
config_coverage = None
if args.attention_type == "coverage":
config_coverage = coverage.CoverageConfig(
type=args.attention_coverage_type,
num_hidden=args.attention_coverage_num_hidden,
layer_normalization=args.layer_normalization)
config_attention = attention.AttentionConfig(
type=args.attention_type,
num_hidden=attention_num_hidden,
input_previous_word=args.attention_use_prev_word,
rnn_num_hidden=config_rnn.num_hidden,
layer_normalization=args.layer_normalization,
config_coverage=config_coverage)
config_loss = loss.LossConfig(
type=args.loss,
vocab_size=vocab_target_size,
normalize=args.normalize_loss,
smoothed_cross_entropy_alpha=args.smoothed_cross_entropy_alpha)
model_config = model.ModelConfig(
config_data=config_data,
max_seq_len=max_seq_len_source,
vocab_source_size=vocab_source_size,
vocab_target_size=vocab_target_size,
config_encoder=config_encoder,
config_decoder=config_decoder,
config_attention=config_attention,
config_loss=config_loss,
lexical_bias=args.lexical_bias,
learn_lexical_bias=args.learn_lexical_bias)
model_config.freeze()
# create training model
training_model = training.TrainingModel(
config=model_config,
context=context,
train_iter=train_iter,
fused=args.use_fused_rnn,
bucketing=not args.no_bucketing,
lr_scheduler=lr_scheduler_instance)
# We may consider loading the params in TrainingModule, for consistency
# with the training state saving
if resume_training:
logger.info(
"Found partial training in directory %s. Resuming from saved state.",
training_state_dir)
training_model.load_params_from_file(
os.path.join(training_state_dir, C.TRAINING_STATE_PARAMS_NAME))
elif args.params:
logger.info(
"Training will initialize from parameters loaded from '%s'",
args.params)
training_model.load_params_from_file(args.params)
lexicon_array = lexicon.initialize_lexicon(
args.lexical_bias, vocab_source,
vocab_target) if args.lexical_bias else None
weight_initializer = initializer.get_initializer(args.rnn_h2h_init,
lexicon=lexicon_array)
optimizer = args.optimizer
optimizer_params = {
'wd': args.weight_decay,
"learning_rate": args.initial_learning_rate
}
if lr_scheduler_instance is not None:
optimizer_params["lr_scheduler"] = lr_scheduler_instance
clip_gradient = none_if_negative(args.clip_gradient)
if clip_gradient is not None:
optimizer_params["clip_gradient"] = clip_gradient
if args.momentum is not None:
optimizer_params["momentum"] = args.momentum
if args.normalize_loss:
# When normalize_loss is turned on we normalize by the number of non-PAD symbols in a batch which implicitly
# already contains the number of sentences and therefore we need to disable rescale_grad.
optimizer_params["rescale_grad"] = 1.0
else:
# Making MXNet module API's default scaling factor explicit
optimizer_params["rescale_grad"] = 1.0 / args.batch_size
logger.info("Optimizer: %s", optimizer)
logger.info("Optimizer Parameters: %s", optimizer_params)
training_model.fit(
train_iter,
eval_iter,
output_folder=output_folder,
max_params_files_to_keep=args.keep_last_params,
metrics=args.metrics,
initializer=weight_initializer,
max_updates=args.max_updates,
checkpoint_frequency=args.checkpoint_frequency,
optimizer=optimizer,
optimizer_params=optimizer_params,
optimized_metric=args.optimized_metric,
max_num_not_improved=args.max_num_checkpoint_not_improved,
min_num_epochs=args.min_num_epochs,
monitor_bleu=args.monitor_bleu,
use_tensorboard=args.use_tensorboard)
def test_get_num_gpus():
assert utils.get_num_gpus() >= 0
def main():
params = argparse.ArgumentParser(
description='CLI to train sockeye sequence-to-sequence models.')
params = arguments.add_io_args(params)
params = arguments.add_model_parameters(params)
params = arguments.add_training_args(params)
params = arguments.add_device_args(params)
args = params.parse_args()
# seed the RNGs
np.random.seed(args.seed)
random.seed(args.seed)
mx.random.seed(args.seed)
if args.use_fused_rnn:
assert not args.use_cpu, "GPU required for FusedRNN cells"
if args.rnn_residual_connections:
assert args.rnn_num_layers > 2, "Residual connections require at least 3 RNN layers"
assert args.optimized_metric == C.BLEU or args.optimized_metric in args.metrics, \
"Must optimize either BLEU or one of tracked metrics (--metrics)"
# Checking status of output folder, resumption, etc.
# Create temporary logger to console only
logger = setup_main_logger(__name__, console=not args.quiet)
output_folder = os.path.abspath(args.output)
resume_training = False
training_state_dir = os.path.join(output_folder, C.TRAINING_STATE_DIRNAME)
if os.path.exists(output_folder):
if args.overwrite_output:
logger.info("Removing existing output folder %s.", output_folder)
shutil.rmtree(output_folder)
os.makedirs(output_folder)
elif os.path.exists(training_state_dir):
with open(os.path.join(output_folder, C.ARGS_STATE_NAME),
"r") as fp:
old_args = json.load(fp)
arg_diffs = _dict_difference(
vars(args), old_args) | _dict_difference(old_args, vars(args))
# Remove args that may differ without affecting the training.
arg_diffs -= set(C.ARGS_MAY_DIFFER)
if not arg_diffs:
resume_training = True
else:
# We do not have the logger yet
logger.error(
"Mismatch in arguments for training continuation.")
logger.error("Differing arguments: %s.", ", ".join(arg_diffs))
sys.exit(1)
else:
logger.error("Refusing to overwrite existing output folder %s.",
output_folder)
sys.exit(1)
else:
os.makedirs(output_folder)
logger = setup_main_logger(__name__,
console=not args.quiet,
path=os.path.join(output_folder, C.LOG_NAME))
logger.info("Command: %s", " ".join(sys.argv))
logger.info("Arguments: %s", args)
with open(os.path.join(output_folder, C.ARGS_STATE_NAME), "w") as fp:
json.dump(vars(args), fp)
with ExitStack() as exit_stack:
# context
if args.use_cpu:
context = [mx.cpu()]
else:
num_gpus = get_num_gpus()
assert num_gpus > 0, "No GPUs found, consider running on the CPU with --use-cpu " \
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi " \
"binary isn't on the path)."
context = []
for gpu_id in args.device_ids:
if gpu_id < 0:
# get an automatic gpu id:
gpu_id = exit_stack.enter_context(acquire_gpu())
context.append(mx.gpu(gpu_id))
# create vocabs
vocab_source = _build_or_load_vocab(args.source_vocab, args.source,
args.num_words,
args.word_min_count)
sockeye.vocab.vocab_to_json(
vocab_source,
os.path.join(output_folder, C.VOCAB_SRC_NAME) + C.JSON_SUFFIX)
vocab_target = _build_or_load_vocab(args.target_vocab, args.target,
args.num_words,
args.word_min_count)
sockeye.vocab.vocab_to_json(
vocab_target,
os.path.join(output_folder, C.VOCAB_TRG_NAME) + C.JSON_SUFFIX)
vocab_source_size = len(vocab_source)
vocab_target_size = len(vocab_target)
logger.info("Vocabulary sizes: source=%d target=%d", vocab_source_size,
vocab_target_size)
data_info = sockeye.data_io.DataInfo(
os.path.abspath(args.source), os.path.abspath(args.target),
os.path.abspath(args.validation_source),
os.path.abspath(args.validation_target), args.source_vocab,
args.target_vocab)
# create data iterators
train_iter, eval_iter = sockeye.data_io.get_training_data_iters(
source=data_info.source,
target=data_info.target,
validation_source=data_info.validation_source,
validation_target=data_info.validation_target,
vocab_source=vocab_source,
vocab_target=vocab_target,
batch_size=args.batch_size,
fill_up=args.fill_up,
max_seq_len=args.max_seq_len,
bucketing=not args.no_bucketing,
bucket_width=args.bucket_width)
# learning rate scheduling
learning_rate_half_life = none_if_negative(
args.learning_rate_half_life)
# TODO: The loading for continuation of the scheduler is done separately from the other parts
if not resume_training:
lr_scheduler = sockeye.lr_scheduler.get_lr_scheduler(
args.learning_rate_scheduler_type, args.checkpoint_frequency,
learning_rate_half_life, args.learning_rate_reduce_factor,
args.learning_rate_reduce_num_not_improved)
else:
with open(os.path.join(training_state_dir, C.SCHEDULER_STATE_NAME),
"rb") as fp:
lr_scheduler = pickle.load(fp)
# model configuration
num_embed_source = args.num_embed if args.num_embed_source is None else args.num_embed_source
num_embed_target = args.num_embed if args.num_embed_target is None else args.num_embed_target
attention_num_hidden = args.rnn_num_hidden if not args.attention_num_hidden else args.attention_num_hidden
model_config = sockeye.model.ModelConfig(
max_seq_len=args.max_seq_len,
vocab_source_size=vocab_source_size,
vocab_target_size=vocab_target_size,
num_embed_source=num_embed_source,
num_embed_target=num_embed_target,
attention_type=args.attention_type,
attention_num_hidden=attention_num_hidden,
attention_coverage_type=args.attention_coverage_type,
attention_coverage_num_hidden=args.attention_coverage_num_hidden,
attention_use_prev_word=args.attention_use_prev_word,
dropout=args.dropout,
rnn_cell_type=args.rnn_cell_type,
rnn_num_layers=args.rnn_num_layers,
rnn_num_hidden=args.rnn_num_hidden,
rnn_residual_connections=args.rnn_residual_connections,
weight_tying=args.weight_tying,
context_gating=args.context_gating,
lexical_bias=args.lexical_bias,
learn_lexical_bias=args.learn_lexical_bias,
data_info=data_info,
loss=args.loss,
normalize_loss=args.normalize_loss,
smoothed_cross_entropy_alpha=args.smoothed_cross_entropy_alpha)
# create training model
model = sockeye.training.TrainingModel(
model_config=model_config,
context=context,
train_iter=train_iter,
fused=args.use_fused_rnn,
bucketing=not args.no_bucketing,
lr_scheduler=lr_scheduler,
rnn_forget_bias=args.rnn_forget_bias)
# We may consider loading the params in TrainingModule, for consistency
# with the training state saving
if resume_training:
logger.info(
"Found partial training in directory %s. Resuming from saved state.",
training_state_dir)
model.load_params_from_file(
os.path.join(training_state_dir, C.TRAINING_STATE_PARAMS_NAME))
elif args.params:
logger.info(
"Training will initialize from parameters loaded from '%s'",
args.params)
model.load_params_from_file(args.params)
lexicon = sockeye.lexicon.initialize_lexicon(
args.lexical_bias, vocab_source,
vocab_target) if args.lexical_bias else None
initializer = sockeye.initializer.get_initializer(args.rnn_h2h_init,
lexicon=lexicon)
optimizer = args.optimizer
optimizer_params = {
'wd': args.weight_decay,
"learning_rate": args.initial_learning_rate
}
if lr_scheduler is not None:
optimizer_params["lr_scheduler"] = lr_scheduler
clip_gradient = none_if_negative(args.clip_gradient)
if clip_gradient is not None:
optimizer_params["clip_gradient"] = clip_gradient
if args.momentum is not None:
optimizer_params["momentum"] = args.momentum
logger.info("Optimizer: %s", optimizer)
logger.info("Optimizer Parameters: %s", optimizer_params)
model.fit(train_iter,
eval_iter,
output_folder=output_folder,
metrics=args.metrics,
initializer=initializer,
max_updates=args.max_updates,
checkpoint_frequency=args.checkpoint_frequency,
optimizer=optimizer,
optimizer_params=optimizer_params,
optimized_metric=args.optimized_metric,
max_num_not_improved=args.max_num_checkpoint_not_improved,
monitor_bleu=args.monitor_bleu,
use_tensorboard=args.use_tensorboard)
def main():
params = argparse.ArgumentParser(
description='Translate from STDIN to STDOUT')
params = arguments.add_inference_args(params)
params = arguments.add_device_args(params)
args = params.parse_args()
logger = setup_main_logger(__name__, file_logging=False)
assert args.beam_size > 0, "Beam size must be 1 or greater."
if args.checkpoints is not None:
assert len(args.checkpoints) == len(
args.models), "must provide checkpoints for each model"
logger.info("Command: %s", " ".join(sys.argv))
logger.info("Arguments: %s", args)
output_stream = sys.stdout
output_handler = sockeye.output_handler.get_output_handler(
args.output_type, output_stream, args.align_plot_prefix,
args.sure_align_threshold)
with ExitStack() as exit_stack:
if args.use_cpu:
context = mx.cpu()
else:
num_gpus = get_num_gpus()
assert num_gpus > 0, "No GPUs found, consider running on the CPU with --use-cpu " \
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi " \
"binary isn't on the path)."
assert len(
args.device_ids) == 1, "cannot run on multiple devices for now"
gpu_id = args.device_ids[0]
if gpu_id < 0:
# get a gpu id automatically:
gpu_id = exit_stack.enter_context(acquire_gpu())
context = mx.gpu(gpu_id)
translator = sockeye.inference.Translator(
context, args.ensemble_mode,
*sockeye.inference.load_models(context, args.max_input_len,
args.beam_size, args.models,
args.checkpoints,
args.softmax_temperature))
total_time = 0
i = 0
for i, line in enumerate(sys.stdin, 1):
trans_input = translator.make_input(i, line)
logger.debug(" IN: %s", trans_input)
tic = time.time()
trans_output = translator.translate(trans_input)
trans_wall_time = time.time() - tic
total_time += trans_wall_time
logger.debug("OUT: %s", trans_output)
logger.debug("OUT: time=%.2f", trans_wall_time)
output_handler.handle(trans_input, trans_output)
if i != 0:
logger.info(
"Processed %d lines. Total time: %.4f sec/sent: %.4f sent/sec: %.4f",
i, total_time, total_time / i, i / total_time)
else:
logger.info("Processed 0 lines.")
def main():
params = argparse.ArgumentParser(
description='CLI to train sockeye sequence-to-sequence models.')
params = arguments.add_io_args(params)
params = arguments.add_model_parameters(params)
params = arguments.add_training_args(params)
params = arguments.add_device_args(params)
args = params.parse_args()
# seed the RNGs
np.random.seed(args.seed)
random.seed(args.seed)
mx.random.seed(args.seed)
if args.use_fused_rnn:
assert not args.use_cpu, "GPU required for FusedRNN cells"
if args.rnn_residual_connections:
assert args.rnn_num_layers > 2, "Residual connections require at least 3 RNN layers"
assert args.optimized_metric == C.BLEU or args.optimized_metric in args.metrics, \
"Must optimize either BLEU or one of tracked metrics (--metrics)"
output_folder = os.path.abspath(args.output)
if not os.path.exists(output_folder):
os.makedirs(output_folder)
logger = setup_main_logger(__name__,
console=not args.quiet,
path=os.path.join(output_folder, C.LOG_NAME))
logger.info("Command: %s", " ".join(sys.argv))
logger.info("Arguments: %s", args)
with ExitStack() as exit_stack:
# context
if args.use_cpu:
context = [mx.cpu()]
else:
num_gpus = get_num_gpus()
assert num_gpus > 0, "No GPUs found, consider running on the CPU with --use-cpu " \
"(note: check depends on nvidia-smi and this could also mean that the nvidia-smi " \
"binary isn't on the path)."
context = []
for gpu_id in args.device_ids:
if gpu_id < 0:
# get an automatic gpu id:
gpu_id = exit_stack.enter_context(acquire_gpu())
context.append(mx.gpu(gpu_id))
# create vocabs
vocab_source = _build_or_load_vocab(args.source_vocab, args.source,
args.num_words,
args.word_min_count)
sockeye.vocab.vocab_to_json(
vocab_source,
os.path.join(output_folder, C.VOCAB_SRC_NAME) + C.JSON_SUFFIX)
vocab_target = _build_or_load_vocab(args.target_vocab, args.target,
args.num_words,
args.word_min_count)
sockeye.vocab.vocab_to_json(
vocab_target,
os.path.join(output_folder, C.VOCAB_TRG_NAME) + C.JSON_SUFFIX)
vocab_source_size = len(vocab_source)
vocab_target_size = len(vocab_target)
logger.info("Vocabulary sizes: source=%d target=%d", vocab_source_size,
vocab_target_size)
data_info = sockeye.data_io.DataInfo(
os.path.abspath(args.source), os.path.abspath(args.target),
os.path.abspath(args.validation_source),
os.path.abspath(args.validation_target), args.source_vocab,
args.target_vocab)
# create data iterators
train_iter, eval_iter = sockeye.data_io.get_training_data_iters(
source=data_info.source,
target=data_info.target,
validation_source=data_info.validation_source,
validation_target=data_info.validation_target,
vocab_source=vocab_source,
vocab_target=vocab_target,
batch_size=args.batch_size,
fill_up=args.fill_up,
max_seq_len=args.max_seq_len,
bucketing=not args.no_bucketing,
bucket_width=args.bucket_width)
# learning rate scheduling
learning_rate_half_life = none_if_negative(
args.learning_rate_half_life)
lr_scheduler = sockeye.lr_scheduler.get_lr_scheduler(
args.learning_rate_scheduler_type, args.checkpoint_frequency,
learning_rate_half_life, args.learning_rate_reduce_factor,
args.learning_rate_reduce_num_not_improved)
# model configuration
num_embed_source = args.num_embed if args.num_embed_source is None else args.num_embed_source
num_embed_target = args.num_embed if args.num_embed_target is None else args.num_embed_target
attention_num_hidden = args.rnn_num_hidden if not args.attention_num_hidden else args.attention_num_hidden
model_config = sockeye.model.ModelConfig(
max_seq_len=args.max_seq_len,
vocab_source_size=vocab_source_size,
vocab_target_size=vocab_target_size,
num_embed_source=num_embed_source,
num_embed_target=num_embed_target,
attention_type=args.attention_type,
attention_num_hidden=attention_num_hidden,
attention_coverage_type=args.attention_coverage_type,
attention_coverage_num_hidden=args.attention_coverage_num_hidden,
attention_use_prev_word=args.attention_use_prev_word,
dropout=args.dropout,
rnn_cell_type=args.rnn_cell_type,
rnn_num_layers=args.rnn_num_layers,
rnn_num_hidden=args.rnn_num_hidden,
rnn_residual_connections=args.rnn_residual_connections,
weight_tying=args.weight_tying,
context_gating=args.context_gating,
lexical_bias=args.lexical_bias,
learn_lexical_bias=args.learn_lexical_bias,
data_info=data_info,
loss=args.loss,
normalize_loss=args.normalize_loss,
smoothed_cross_entropy_alpha=args.smoothed_cross_entropy_alpha)
# create training model
model = sockeye.training.TrainingModel(
model_config=model_config,
context=context,
train_iter=train_iter,
fused=args.use_fused_rnn,
bucketing=not args.no_bucketing,
lr_scheduler=lr_scheduler,
rnn_forget_bias=args.rnn_forget_bias)
if args.params:
model.load_params_from_file(args.params)
logger.info(
"Training will continue from parameters loaded from '%s'",
args.params)
lexicon = sockeye.lexicon.initialize_lexicon(
args.lexical_bias, vocab_source,
vocab_target) if args.lexical_bias else None
initializer = sockeye.initializer.get_initializer(args.rnn_h2h_init,
lexicon=lexicon)
optimizer = args.optimizer
optimizer_params = {
'wd': args.weight_decay,
"learning_rate": args.initial_learning_rate
}
if lr_scheduler is not None:
optimizer_params["lr_scheduler"] = lr_scheduler
clip_gradient = none_if_negative(args.clip_gradient)
if clip_gradient is not None:
optimizer_params["clip_gradient"] = clip_gradient
if args.momentum is not None:
optimizer_params["momentum"] = args.momentum
logger.info("Optimizer: %s", optimizer)
logger.info("Optimizer Parameters: %s", optimizer_params)
model.fit(train_iter,
eval_iter,
output_folder=output_folder,
metrics=args.metrics,
initializer=initializer,
max_updates=args.max_updates,
checkpoint_frequency=args.checkpoint_frequency,
optimizer=optimizer,
optimizer_params=optimizer_params,
optimized_metric=args.optimized_metric,
max_num_not_improved=args.max_num_checkpoint_not_improved,
monitor_bleu=args.monitor_bleu,
use_tensorboard=args.use_tensorboard)
