def train_pos(args):
######################################################
#
# Data preprocessing
#
######################################################
datasets = {
args.datasetName: #Name of the dataset
{
'columns': {
0: 'tokens',
1: 'POS',
2: 'chunk_BIO'
}, #CoNLL format for the input data. Column 1 contains tokens, column 3 contains POS information
'label': 'POS', #Which column we like to predict
'evaluate': True, #Should we evaluate on this task? Set true always for single task setups
'commentSymbol': None
} #Lines in the input data starting with this string will be skipped. Can be used to skip comments
}
# :: Path on your computer to the word embeddings. Embeddings by Komninos et al. will be downloaded automatically ::
embeddingsPath = args.embeddings
#'komninos_english_embeddings.gz'
# :: Prepares the dataset to be used with the LSTM-network. Creates and stores cPickle files in the pkl/ folder ::
pickleFile = perpareDataset(embeddingsPath, datasets)
######################################################
#
# The training of the network starts here
#
######################################################
#Load the embeddings and the dataset
embeddings, mappings, data = loadDatasetPickle(pickleFile)
# Some network hyperparameters
params = {
'classifier': ['CRF'],
'LSTM-Size': [100],
'dropout': (0.25, 0.25)
}
model = BiLSTM(params)
model.setMappings(mappings, embeddings)
model.setDataset(datasets, data)
model.modelSavePath = args.model_save + '/[ModelName]_[Epoch].h5'
model.fit(epochs=25)
fpath = args.model_save + '/' + args.datasetName + '_1.h5'
save_dir, model_init = os.path.split(fpath)
print(save_dir)
print(model_init)
remove_except_last_model(save_dir, model_init)
prepare_training_data(datasets)
embeddingsPath = 'komninos_english_embeddings.gz' #Word embeddings by Levy et al: https://levyomer.wordpress.com/2014/04/25/dependency-based-word-embeddings/
# :: Prepares the dataset to be used with the LSTM-network. Creates and stores cPickle files in the pkl/ folder ::
pickleFile = perpareDataset(embeddingsPath, datasets)
######################################################
#
# The training of the network starts here
#
######################################################
#Load the embeddings and the dataset
embeddings, mappings, data = loadDatasetPickle(pickleFile)
# Some network hyperparameters
#params = {'classifier': ['CRF'], 'LSTM-Size': [100], 'dropout': (0.25, 0.25),'charEmbeddings': 'CNN',
# 'customClassifier': {'unidep_pos': ['Softmax'], 'conll2000_chunking': [('LSTM', 50), 'CRF']}}
# TODO Replace customClassifier dengan main task + auxiliary task
custom_classifier = {}
custom_classifier[target_task] = [('LSTM', 100), 'CRF']
for task in aux_task:
custom_classifier[task] = ['CRF']
params = {
'classifier': ['CRF'],
'LSTM-Size': [100],
'dropout': (0.25, 0.25),
datasetFiles = [
(datasetName, dataColumns),
]
# :: Prepares the dataset to be used with the LSTM-network. Creates and stores cPickle files in the pkl/ folder ::
pickleFile = perpareDataset(embeddingsPath, datasetFiles)
######################################################
#
# The training of the network starts here
#
######################################################
#Load the embeddings and the dataset
embeddings, word2Idx, datasets = loadDatasetPickle(pickleFile)
data = datasets[datasetName]
print("Dataset:", datasetName)
print(data['mappings'].keys())
print("Label key: ", labelKey)
print("Train Sentences:", len(data['trainMatrix']))
print("Dev Sentences:", len(data['devMatrix']))
print("Test Sentences:", len(data['testMatrix']))
model = BiLSTM(params)
model.setMappings(embeddings, data['mappings'])
model.setTrainDataset(data, labelKey)
model.verboseBuild = True
model.modelSavePath = "models/%s/%s/[DevScore]_[TestScore]_[Epoch].h5" % (
datasetName, labelKey) #Enable this line to save the model to the disk
model_name = model_path.stem
# obtain dataset ID and task from model name
dataset_id, task = model_name.rsplit('_', 1)
task = task.upper()
# obtain dataset language from dataset ID
lang = dataset_id.split('_')[0]
lang = lang.upper()
if lang not in loaded_datasets:
# select fasttext word embeddings
embeddings_path = embeddings_dir / f'{lang.lower()}.fasttext.oov.vec.gz'
# load and cache the embeddings, mappings and datasets
loaded_datasets[lang] = loadDatasetPickle(embeddings_path, lang)
# unpack the embeddings, mappings and datasets
embeddings, mappings, data = loaded_datasets[lang]
# evaluate model in a separate process so that memory is released at the end
proc_args = (model_path, dataset_id, task, evaluator, embeddings, mappings,
data)
proc = Process(target=eval_single_task, args=proc_args)
proc.start()
proc.join()
# write the evaluation tables
evaluator.write_tables(tables_dir / 'single_task')
# build and train the model
model.buildModel()
model.fit(
epochs=500) # do not limit training by epochs - use early stopping
for lang in ['PT', 'ES', None]:
# select fasttext word embeddings
lang_prefix = lang.lower() if lang is not None else 'es2pt'
embeddings_path = embeddings_dir / f'{lang_prefix}.fasttext.oov.vec.gz'
# prepare the datasets to be used with the LSTM network
prepareDatasets(embeddings_path, lang)
# load the embeddings and the datasets
embeddings, mappings, data = loadDatasetPickle(embeddings_path, lang)
# iterate through the multiple dataset combinations of language and task
for task in ['POS', 'NER', None]:
if lang is None and task is not None: continue
# obtain datasets for the experiment
datasets = Datasets(exclude=['pt_colonia'], lang=lang, task=task)
datasets_dict = datasets.to_dict()
# run experiment in a separate process so that memory is released at the end
proc_args = (datasets_dict, lang, task, embeddings, mappings, data)
proc = Process(target=run_experiment, args=proc_args)
proc.start()
proc.join()
logger.info(f'Completed experiment: lang {"all" if lang is None else lang} - ' \
def train_pos(args):
######################################################
#
# Data preprocessing
#
######################################################
datasets = {
args.datasetName: #Name of the dataset
{
'columns': {
0: 'tokens',
1: 'POS',
2: 'chunk_BIO'
}, #CoNLL format for the input data. Column 0 contains tokens, column 1 contains POS and column 2 contains chunk information using BIO encoding
'label': 'POS', #Which column we like to predict
'evaluate': True, #Should we evaluate on this task? Set true always for single task setups
'commentSymbol': None
} #Lines in the input data starting with this string will be skipped. Can be used to skip comments
}
# :: Prepares the dataset to be used with the LSTM-network. Creates and stores cPickle files in the pkl/ folder ::
embeddings_file = None
elmo_options_file = args.elmo_options
elmo_weight_file = args.elmo_weights
elmo_mode = 'weighted_average'
#elmo_options_file= 'pretrained/elmo_2x4096_512_2048cnn_2xhighway_5.5B_options.json'
#elmo_weight_file = 'pretrained/elmo_2x4096_512_2048cnn_2xhighway_5.5B_weights.hdf5'
elmo_cuda_device = args.cuda_device #Which GPU to use. -1 for CPU
embLookup = ELMoWordEmbeddings(embeddings_file, elmo_options_file,
elmo_weight_file, elmo_mode,
elmo_cuda_device)
# You can use a cache to precompute the ELMo embeddings once. See Create_ELMo_Cache.py for an example.
embLookup.loadCache(args.pkl_path)
pickleFile = perpareDataset(datasets, embLookup)
######################################################
#
# The training of the network starts here
#
######################################################
#Load the embeddings and the dataset
mappings, data = loadDatasetPickle(pickleFile)
# Some network hyperparameters
params = {
'classifier': ['CRF'],
'LSTM-Size': [100, 100],
'dropout': (0.5, 0.5)
}
model = ELMoBiLSTM(embLookup, params)
model.setMappings(mappings)
model.setDataset(datasets, data)
model.modelSavePath = args.model_save + "/[ModelName]_[Epoch].h5"
model.fit(epochs=25)
fpath = args.model_save + '/' + args.datasetName + '_1.h5'
save_dir, model_init = os.path.split(fpath)
print(save_dir)
print(model_init)
# remove trained files except from the last file
remove_except_last_model(save_dir, model_init)
# :: Path on your computer to the word embeddings. Embeddings by Reimers et al. will be downloaded automatically ::
embeddingsPath = '/datastore/liu121/nosqldb2/emnlp_ukplab/skipgram'
# :: Prepares the dataset to be used with the LSTM-network. Creates and stores cPickle files in the pkl/ folder ::
pickleFile_train, pickleFile_dev, pickleFile_test = perpareDataset(
embeddingsPath, datasets, args.k_shot)
print('data prepare successful: %s, %s, and %s' %
(pickleFile_train, pickleFile_dev, pickleFile_test))
######################################################
#
# The training of the network starts here
#
######################################################
#Load the embeddings and the dataset
embeddings, mappings, data_train = loadDatasetPickle(pickleFile_train)
embeddings, mappings, data_dev = loadDatasetPickle(pickleFile_dev)
embeddings, mappings, data_test = loadDatasetPickle(pickleFile_test)
# print('mappings type: ',type(mappings))
# for key in mappings:
# print(key)
# print(mappings[key])
# print('===============')
# print('embeddings type:',type(data))
# for key in data:
# print(key)
# for subkey in data[key]:
# print('--',subkey)
# for subsubkey in data[key][subkey]:
