def main():
args = parse_args()
server_address = (args.address, args.port)
httpd = HTTPServer(server_address, NMTHandler)
logger.setLevel(args.logging_level)
parameters = load_parameters()
if args.config is not None:
logger.info("Loading parameters from %s" % str(args.config))
parameters = update_parameters(parameters, pkl2dict(args.config))
if args.online:
online_parameters = load_parameters_online()
parameters = update_parameters(parameters, online_parameters)
try:
for arg in args.changes:
try:
k, v = arg.split('=')
except ValueError:
print(
'Overwritten arguments must have the form key=Value. \n Currently are: %s'
% str(args.changes))
exit(1)
try:
parameters[k] = ast.literal_eval(v)
except ValueError:
parameters[k] = v
except ValueError:
print('Error processing arguments: (', k, ",", v, ")")
exit(2)
dataset = loadDataset(args.dataset)
# For converting predictions into sentences
# Dataset backwards compatibility
bpe_separator = dataset.BPE_separator if hasattr(
dataset,
"BPE_separator") and dataset.BPE_separator is not None else '@@'
# Build BPE tokenizer if necessary
if 'bpe' in parameters['TOKENIZATION_METHOD'].lower():
logger.info('Building BPE')
if not dataset.BPE_built:
dataset.build_bpe(parameters.get(
'BPE_CODES_PATH',
parameters['DATA_ROOT_PATH'] + '/training_codes.joint'),
separator=bpe_separator)
# Build tokenization function
tokenize_f = eval('dataset.' +
parameters.get('TOKENIZATION_METHOD', 'tokenize_bpe'))
detokenize_function = eval(
'dataset.' + parameters.get('DETOKENIZATION_METHOD', 'detokenize_bpe'))
dataset.build_moses_tokenizer(language=parameters['SRC_LAN'])
dataset.build_moses_detokenizer(language=parameters['TRG_LAN'])
tokenize_general = dataset.tokenize_moses
detokenize_general = dataset.detokenize_moses
# Prediction parameters
params_prediction = dict()
params_prediction['max_batch_size'] = parameters.get('BATCH_SIZE', 20)
params_prediction['n_parallel_loaders'] = parameters.get(
'PARALLEL_LOADERS', 1)
params_prediction['beam_size'] = parameters.get('BEAM_SIZE', 6)
params_prediction['maxlen'] = parameters.get('MAX_OUTPUT_TEXT_LEN_TEST',
100)
params_prediction['optimized_search'] = parameters['OPTIMIZED_SEARCH']
params_prediction['model_inputs'] = parameters['INPUTS_IDS_MODEL']
params_prediction['model_outputs'] = parameters['OUTPUTS_IDS_MODEL']
params_prediction['dataset_inputs'] = parameters['INPUTS_IDS_DATASET']
params_prediction['dataset_outputs'] = parameters['OUTPUTS_IDS_DATASET']
params_prediction['search_pruning'] = parameters.get(
'SEARCH_PRUNING', False)
params_prediction['normalize_probs'] = True
params_prediction['alpha_factor'] = parameters.get('ALPHA_FACTOR', 1.0)
params_prediction['coverage_penalty'] = True
params_prediction['length_penalty'] = True
params_prediction['length_norm_factor'] = parameters.get(
'LENGTH_NORM_FACTOR', 0.0)
params_prediction['coverage_norm_factor'] = parameters.get(
'COVERAGE_NORM_FACTOR', 0.0)
params_prediction['pos_unk'] = parameters.get('POS_UNK', False)
params_prediction['heuristic'] = parameters.get('HEURISTIC', 0)
params_prediction['state_below_index'] = -1
params_prediction['output_text_index'] = 0
params_prediction['state_below_maxlen'] = -1 if parameters.get(
'PAD_ON_BATCH', True) else parameters.get('MAX_OUTPUT_TEXT_LEN', 50)
params_prediction['output_max_length_depending_on_x'] = parameters.get(
'MAXLEN_GIVEN_X', True)
params_prediction[
'output_max_length_depending_on_x_factor'] = parameters.get(
'MAXLEN_GIVEN_X_FACTOR', 3)
params_prediction['output_min_length_depending_on_x'] = parameters.get(
'MINLEN_GIVEN_X', True)
params_prediction[
'output_min_length_depending_on_x_factor'] = parameters.get(
'MINLEN_GIVEN_X_FACTOR', 2)
params_prediction['attend_on_output'] = parameters.get(
'ATTEND_ON_OUTPUT', 'transformer' in parameters['MODEL_TYPE'].lower())
# Manage pos_unk strategies
if parameters['POS_UNK']:
mapping = None if dataset.mapping == dict() else dataset.mapping
else:
mapping = None
if 'transformer' in parameters['MODEL_TYPE'].lower():
params_prediction['pos_unk'] = False
params_prediction['coverage_penalty'] = False
# Training parameters
parameters_training = dict()
if args.online:
logger.info('Loading models from %s' % str(args.models))
parameters_training = { # Traning parameters
'n_epochs': parameters['MAX_EPOCH'],
'shuffle': False,
'loss': parameters.get('LOSS', 'categorical_crossentropy'),
'batch_size': parameters.get('BATCH_SIZE', 1),
'homogeneous_batches': False,
'optimizer': parameters.get('OPTIMIZER', 'SGD'),
'lr': parameters.get('LR', 0.1),
'lr_decay': parameters.get('LR_DECAY', None),
'lr_gamma': parameters.get('LR_GAMMA', 1.),
'epochs_for_save': -1,
'verbose': args.verbose,
'eval_on_sets': parameters.get('EVAL_ON_SETS_KERAS', None),
'n_parallel_loaders': parameters['PARALLEL_LOADERS'],
'extra_callbacks': [], # callbacks,
'reload_epoch': parameters['RELOAD'],
'epoch_offset': parameters['RELOAD'],
'data_augmentation': parameters['DATA_AUGMENTATION'],
'patience': parameters.get('PATIENCE', 0),
'metric_check': parameters.get('STOP_METRIC', None),
'eval_on_epochs': parameters.get('EVAL_EACH_EPOCHS', True),
'each_n_epochs': parameters.get('EVAL_EACH', 1),
'start_eval_on_epoch': parameters.get('START_EVAL_ON_EPOCH', 0),
'additional_training_settings': {
'k': parameters.get('K', 1),
'tau': parameters.get('TAU', 1),
'lambda': parameters.get('LAMBDA', 0.5),
'c': parameters.get('C', 0.5),
'd': parameters.get('D', 0.5)
}
}
model_instances = [
TranslationModel(
parameters,
model_type=parameters['MODEL_TYPE'],
verbose=parameters['VERBOSE'],
model_name=parameters['MODEL_NAME'] + '_' + str(i),
vocabularies=dataset.vocabulary,
store_path=parameters['STORE_PATH'],
set_optimizer=False) for i in range(len(args.models))
]
models = [
updateModel(model, path, -1, full_path=True)
for (model, path) in zip(model_instances, args.models)
]
else:
models = [loadModel(m, -1, full_path=True) for m in args.models]
for nmt_model in models:
nmt_model.setParams(parameters)
nmt_model.setOptimizer()
parameters['INPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
parameters['INPUTS_IDS_DATASET'][0]]
parameters['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
parameters['OUTPUTS_IDS_DATASET'][0]]
# Get word2index and index2word dictionaries
index2word_y = dataset.vocabulary[parameters['OUTPUTS_IDS_DATASET']
[0]]['idx2words']
word2index_y = dataset.vocabulary[parameters['OUTPUTS_IDS_DATASET']
[0]]['words2idx']
index2word_x = dataset.vocabulary[parameters['INPUTS_IDS_DATASET']
[0]]['idx2words']
word2index_x = dataset.vocabulary[parameters['INPUTS_IDS_DATASET']
[0]]['words2idx']
excluded_words = None
interactive_beam_searcher = NMTSampler(models,
dataset,
parameters,
params_prediction,
parameters_training,
tokenize_f,
detokenize_function,
tokenize_general,
detokenize_general,
mapping=mapping,
word2index_x=word2index_x,
word2index_y=word2index_y,
index2word_y=index2word_y,
eos_symbol=args.eos_symbol,
excluded_words=excluded_words,
online=args.online,
verbose=args.verbose)
httpd.sampler = interactive_beam_searcher
logger.info('Server starting at %s' % str(server_address))
httpd.serve_forever()
def train_model(params, load_dataset=None):
"""
Training function.
Sets the training parameters from params.
Build or loads the model and launches the training.
:param dict params: Dictionary of network hyperparameters.
:param str load_dataset: Load dataset from file or build it from the parameters.
:return: None
"""
if params['RELOAD'] > 0:
logger.info('Resuming training.')
# Load data
if load_dataset is None:
if params['REBUILD_DATASET']:
logger.info('Rebuilding dataset.')
dataset = build_dataset(params)
else:
logger.info('Updating dataset.')
dataset = loadDataset(params['DATASET_STORE_PATH'] +
'/Dataset_' + params['DATASET_NAME'] +
'_' + params['SRC_LAN'] +
params['TRG_LAN'] + '.pkl')
epoch_offset = 0 if dataset.len_train == 0 else int(
params['RELOAD'] * params['BATCH_SIZE'] /
dataset.len_train)
params['EPOCH_OFFSET'] = params['RELOAD'] if params[
'RELOAD_EPOCH'] else epoch_offset
for split, filename in iteritems(params['TEXT_FILES']):
dataset = update_dataset_from_file(
dataset,
params['DATA_ROOT_PATH'] + '/' + filename +
params['SRC_LAN'],
params,
splits=list([split]),
output_text_filename=params['DATA_ROOT_PATH'] + '/' +
filename + params['TRG_LAN'],
remove_outputs=False,
compute_state_below=True,
recompute_references=True)
dataset.name = params['DATASET_NAME'] + '_' + params[
'SRC_LAN'] + params['TRG_LAN']
saveDataset(dataset, params['DATASET_STORE_PATH'])
else:
logger.info('Reloading and using dataset.')
dataset = loadDataset(load_dataset)
else:
# Load data
if load_dataset is None:
dataset = build_dataset(params)
else:
dataset = loadDataset(load_dataset)
params['INPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['INPUTS_IDS_DATASET'][0]]
params['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['OUTPUTS_IDS_DATASET'][0]]
# Build model
set_optimizer = True if params['RELOAD'] == 0 else False
clear_dirs = True if params['RELOAD'] == 0 else False
# build new model
nmt_model = TranslationModel(params,
model_type=params['MODEL_TYPE'],
verbose=params['VERBOSE'],
model_name=params['MODEL_NAME'],
vocabularies=dataset.vocabulary,
store_path=params['STORE_PATH'],
set_optimizer=set_optimizer,
clear_dirs=clear_dirs)
# Define the inputs and outputs mapping from our Dataset instance to our model
inputMapping = dict()
for i, id_in in enumerate(params['INPUTS_IDS_DATASET']):
pos_source = dataset.ids_inputs.index(id_in)
id_dest = nmt_model.ids_inputs[i]
inputMapping[id_dest] = pos_source
nmt_model.setInputsMapping(inputMapping)
outputMapping = dict()
for i, id_out in enumerate(params['OUTPUTS_IDS_DATASET']):
pos_target = dataset.ids_outputs.index(id_out)
id_dest = nmt_model.ids_outputs[i]
outputMapping[id_dest] = pos_target
nmt_model.setOutputsMapping(outputMapping)
if params['RELOAD'] > 0:
nmt_model = updateModel(nmt_model,
params['STORE_PATH'],
params['RELOAD'],
reload_epoch=params['RELOAD_EPOCH'])
nmt_model.setParams(params)
nmt_model.setOptimizer()
if params.get('EPOCH_OFFSET') is None:
params['EPOCH_OFFSET'] = params['RELOAD'] if params['RELOAD_EPOCH'] else \
int(params['RELOAD'] * params['BATCH_SIZE'] / dataset.len_train)
# Store configuration as pkl
dict2pkl(params, params['STORE_PATH'] + '/config')
# Callbacks
callbacks = buildCallbacks(params, nmt_model, dataset)
# Training
total_start_time = timer()
logger.debug('Starting training!')
training_params = {
'n_epochs':
params['MAX_EPOCH'],
'batch_size':
params['BATCH_SIZE'],
'homogeneous_batches':
params['HOMOGENEOUS_BATCHES'],
'maxlen':
params['MAX_OUTPUT_TEXT_LEN'],
'joint_batches':
params['JOINT_BATCHES'],
'lr_decay':
params.get('LR_DECAY', None), # LR decay parameters
'initial_lr':
params.get('LR', 1.0),
'reduce_each_epochs':
params.get('LR_REDUCE_EACH_EPOCHS', True),
'start_reduction_on_epoch':
params.get('LR_START_REDUCTION_ON_EPOCH', 0),
'lr_gamma':
params.get('LR_GAMMA', 0.9),
'lr_reducer_type':
params.get('LR_REDUCER_TYPE', 'linear'),
'lr_reducer_exp_base':
params.get('LR_REDUCER_EXP_BASE', 0),
'lr_half_life':
params.get('LR_HALF_LIFE', 50000),
'lr_warmup_exp':
params.get('WARMUP_EXP', -1.5),
'min_lr':
params.get('MIN_LR', 1e-9),
'epochs_for_save':
params['EPOCHS_FOR_SAVE'],
'verbose':
params['VERBOSE'],
'eval_on_sets':
params['EVAL_ON_SETS_KERAS'],
'n_parallel_loaders':
params['PARALLEL_LOADERS'],
'extra_callbacks':
callbacks,
'reload_epoch':
params['RELOAD'],
'epoch_offset':
params.get('EPOCH_OFFSET', 0),
'data_augmentation':
params['DATA_AUGMENTATION'],
'patience':
params.get('PATIENCE', 0), # early stopping parameters
'metric_check':
params.get('STOP_METRIC', None)
if params.get('EARLY_STOP', False) else None,
'eval_on_epochs':
params.get('EVAL_EACH_EPOCHS', True),
'each_n_epochs':
params.get('EVAL_EACH', 1),
'start_eval_on_epoch':
params.get('START_EVAL_ON_EPOCH', 0),
'tensorboard':
params.get('TENSORBOARD', False),
'n_gpus':
params.get('N_GPUS', 1),
'tensorboard_params': {
'log_dir':
params.get('LOG_DIR', 'tensorboard_logs'),
'histogram_freq':
params.get('HISTOGRAM_FREQ', 0),
'batch_size':
params.get('TENSORBOARD_BATCH_SIZE', params['BATCH_SIZE']),
'write_graph':
params.get('WRITE_GRAPH', True),
'write_grads':
params.get('WRITE_GRADS', False),
'write_images':
params.get('WRITE_IMAGES', False),
'embeddings_freq':
params.get('EMBEDDINGS_FREQ', 0),
'embeddings_layer_names':
params.get('EMBEDDINGS_LAYER_NAMES', None),
'embeddings_metadata':
params.get('EMBEDDINGS_METADATA', None),
'label_word_embeddings_with_vocab':
params.get('LABEL_WORD_EMBEDDINGS_WITH_VOCAB', False),
'word_embeddings_labels':
params.get('WORD_EMBEDDINGS_LABELS', None),
}
}
nmt_model.trainNet(dataset, training_params)
total_end_time = timer()
time_difference = total_end_time - total_start_time
logger.info('In total is {0:.2f}s = {1:.2f}m'.format(
time_difference, time_difference / 60.0))
'd': params.get('D', 0.5)
}
}
params['INPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['INPUTS_IDS_DATASET'][0]]
params['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['OUTPUTS_IDS_DATASET'][0]]
logger.info("<<< Using an ensemble of %d models >>>" % len(args.models))
# Load trainable model(s)
logging.info('Loading models from %s' % str(args.models))
model_instances = [
TranslationModel(params,
model_type=params['MODEL_TYPE'],
verbose=params['VERBOSE'],
model_name=params['MODEL_NAME'] + '_' + str(i),
vocabularies=dataset.vocabulary,
store_path=params['STORE_PATH'],
clear_dirs=False,
set_optimizer=False) for i in range(len(args.models))
]
models = [
updateModel(model, path, -1, full_path=True)
for (model, path) in zip(model_instances, args.models)
]
# Set additional inputs to models if using a custom loss function
params['USE_CUSTOM_LOSS'] = True if 'PAS' in params['OPTIMIZER'] else False
if params.get('N_BEST_OPTIMIZER', False):
logging.info('Using N-best optimizer')
models = build_online_models(models, params)
src_model = loadModel(SRC_MODEL_PATH, epoch_choice)
params = src_model.params
params['USE_CUDNN'] = False
# params['BIDIRECTIONAL_ENCODER'] = False # Set to False to get the RNN type displayed.
params['MODEL_NAME'] = 'CPU'
params['STORE_PATH'] = DST_MODEL_PATH
params['MODE'] = 'sampling'
params['RELOAD'] = epoch_choice
# params['INPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[params['INPUTS_IDS_DATASET'][0]]
# params['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[params['OUTPUTS_IDS_DATASET'][0]]
cpu_model = TranslationModel(params,
model_type=params['MODEL_TYPE'],
verbose=True,
model_name=params['MODEL_NAME'],
vocabularies=dataset.vocabulary,
store_path=params['STORE_PATH'],
set_optimizer=True,
clear_dirs=True)
exit()
cpu_model = updateModel(cpu_model,
SRC_MODEL_PATH,
params['RELOAD'],
reload_epoch=True)
saveModel(cpu_model,
update_num=epoch_choice,
path=DST_MODEL_PATH,
full_path=True)