class LanguageModelingTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(LanguageModelingTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'lm'
def _create_task_specific_reader(self):
self._create_vectorizers()
reader_params = self.config_params['reader'] if 'reader' in self.config_params else self.config_params['loader']
reader_params['nctx'] = reader_params.get('nctx', self.config_params.get('nctx', self.config_params.get('nbptt', 35)))
reader_params['clean_fn'] = reader_params.get('clean_fn', self.config_params.get('preproc', {}).get('clean_fn'))
if reader_params['clean_fn'] is not None and self.config_params['dataset'] != 'SST2':
logger.warning('Warning: A reader preprocessing function (%s) is active, it is recommended that all data preprocessing is done outside of baseline to insure data at inference time matches data at training time.', reader_params['clean_fn'])
reader_params['mxlen'] = self.vectorizers[self.primary_key].mxlen
if self.config_params['model'].get('gpus', 1) > 1:
reader_params['truncate'] = True
return baseline.reader.create_reader(self.task_name(), self.vectorizers, self.config_params['preproc'].get('trim', False), **reader_params)
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if backend.name == 'pytorch':
self.config_params.get('preproc', {})['trim'] = True
elif backend.name == 'dy':
self.config_params.get('preproc', {})['trim'] = True
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': batched}
backend.load(self.task_name())
return backend
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
vocab_sources = [self.dataset['train_file'], self.dataset['valid_file']]
# TODO: make this optional
if 'test_file' in self.dataset:
vocab_sources.append(self.dataset['test_file'])
vocabs = self.reader.build_vocab(vocab_sources,
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file'))
self.embeddings, self.feat2index = self._create_embeddings(embeddings_set, vocabs, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _load_dataset(self):
read = self.config_params['reader'] if 'reader' in self.config_params else self.config_params['loader']
tgt_key = read.get('tgt_key', self.primary_key)
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2index,
bsz,
tgt_key=tgt_key
)
self.valid_data = self.reader.load(
self.dataset['valid_file'],
self.feat2index,
vbsz,
tgt_key=tgt_key
)
self.test_data = None
if 'test_file' in self.dataset:
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2index,
1,
tgt_key=tgt_key
)
def _create_model(self):
model = self.config_params['model']
unif = self.config_params.get('unif', 0.1)
model['unif'] = model.get('unif', unif)
model['batchsz'] = self.config_params['batchsz']
model['tgt_key'] = self.config_params.get('reader',
self.config_params.get('loader', {})).get('tgt_key', self.primary_key)
model['src_keys'] = listify(self.config_params.get('reader', list(self.config_params.get('loader', {}).get('src_keys', self.embeddings.keys()))))
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_lang_model(self.embeddings, **model)
def train(self, checkpoint=None):
self._load_dataset()
if self.config_params['train'].get('lr_scheduler_type', None) == 'zaremba':
first_range = int(self.config_params['train']['start_decay_epoch'] * self.train_data.steps)
self.config_params['train']['bounds'] = [first_range] + list(
np.arange(
self.config_params['train']['start_decay_epoch'] + 1,
self.config_params['train']['epochs'] + 1,
dtype=np.int32
) * self.train_data.steps
)
baseline.save_vectorizers(self.get_basedir(), self.vectorizers)
model = self._create_model()
train_params = self.config_params['train']
train_params['checkpoint'] = checkpoint
metrics = baseline.train.fit(model, self.train_data, self.valid_data, self.test_data, **train_params)
baseline.zip_files(self.get_basedir())
self._close_reporting_hooks()
return model, metrics
@staticmethod
def _num_steps_per_epoch(num_examples, nctx, batchsz):
rest = num_examples // batchsz
return rest // nctx
class TaggerTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(TaggerTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'tagger'
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if 'preproc' not in self.config_params:
self.config_params['preproc'] = {}
if backend.name == 'pytorch':
self.config_params['preproc']['trim'] = True
elif backend.name == 'dy':
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
dy_params.set_autobatch(True)
else:
raise Exception('Tagger currently only supports autobatching.'
'Change "batchsz" to 1 and under "train", set "autobatchsz" to your desired batchsz')
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': False}
self.config_params['preproc']['trim'] = True
else:
self.config_params['preproc']['trim'] = False
backend.load(self.task_name())
return backend
def initialize(self, embeddings):
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
vocab_sources = [self.dataset['train_file'], self.dataset['valid_file']]
# TODO: make this optional
if 'test_file' in self.dataset:
vocab_sources.append(self.dataset['test_file'])
vocabs = self.reader.build_vocab(vocab_sources, min_f=Task._get_min_f(self.config_params),
vocab_file
=self.dataset.get('vocab_file'))
self.embeddings, self.feat2index = self._create_embeddings(embeddings_set, vocabs, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _create_model(self):
labels = self.reader.label2index
span_type = self.config_params['train'].get('span_type')
constrain = bool(self.config_params['model'].get('constrain_decode', False))
if span_type is None and constrain:
logger.warning("Constrained Decoding was set but no span type could be found so no Constraints will be applied.")
self.config_params['model']['span_type'] = span_type
if span_type is not None and constrain:
self.config_params['model']['constraint'] = self.backend.transition_mask(
labels, span_type, Offsets.GO, Offsets.EOS, Offsets.PAD
)
model = self.config_params['model']
unif = self.config_params.get('unif', 0.1)
model['unif'] = model.get('unif', unif)
lengths_key = model.get('lengths_key', self.primary_key)
if lengths_key is not None:
if not lengths_key.endswith('_lengths'):
lengths_key = '{}_lengths'.format(lengths_key)
model['lengths_key'] = lengths_key
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_tagger_model(self.embeddings, labels, **self.config_params['model'])
def _load_dataset(self):
# TODO: get rid of sort_key=self.primary_key in favor of something explicit?
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data, _ = self.reader.load(
self.dataset['train_file'],
self.feat2index,
bsz,
shuffle=True,
sort_key='{}_lengths'.format(self.primary_key)
)
self.valid_data, _ = self.reader.load(
self.dataset['valid_file'],
self.feat2index,
vbsz,
sort_key=None
)
self.test_data = None
self.txts = None
if 'test_file' in self.dataset:
self.test_data, self.txts = self.reader.load(
self.dataset['test_file'],
self.feat2index,
tbsz,
shuffle=False,
sort_key=None
)
def train(self, checkpoint=None):
self._load_dataset()
baseline.save_vectorizers(self.get_basedir(), self.vectorizers)
model = self._create_model()
conll_output = self.config_params.get("conll_output", None)
train_params = self.config_params['train']
train_params['checkpoint'] = checkpoint
metrics = baseline.train.fit(model, self.train_data, self.valid_data, self.test_data,
conll_output=conll_output,
txts=self.txts, **train_params)
baseline.zip_files(self.get_basedir())
self._close_reporting_hooks()
return model, metrics
class EncoderDecoderTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(EncoderDecoderTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'seq2seq'
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if 'preproc' not in self.config_params:
self.config_params['preproc'] = {}
self.config_params['preproc']['show_ex'] = show_examples
if backend.name == 'pytorch':
self.config_params['preproc']['trim'] = True
elif backend.name == 'dy':
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
self.config_params['train']['trainer_type'] = 'autobatch'
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': batched}
self.config_params['preproc']['trim'] = True
else:
self.config_params['preproc']['trim'] = True
backend.load(self.task_name())
return backend
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
vocab_sources = [self.dataset['train_file'], self.dataset['valid_file']]
# TODO: make this optional
if 'test_file' in self.dataset:
vocab_sources.append(self.dataset['test_file'])
vocab1, vocab2 = self.reader.build_vocabs(vocab_sources,
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file'))
# To keep the config file simple, share a list between source and destination (tgt)
features_src = []
features_tgt = None
for feature in self.config_params['features']:
if feature['name'] == 'tgt':
features_tgt = feature
else:
features_src += [feature]
self.src_embeddings, self.feat2src = self._create_embeddings(embeddings_set, vocab1, features_src)
# For now, dont allow multiple vocabs of output
baseline.save_vocabs(self.get_basedir(), self.feat2src)
self.tgt_embeddings, self.feat2tgt = self._create_embeddings(embeddings_set, {'tgt': vocab2}, [features_tgt])
baseline.save_vocabs(self.get_basedir(), self.feat2tgt)
self.tgt_embeddings = self.tgt_embeddings['tgt']
self.feat2tgt = self.feat2tgt['tgt']
def _load_dataset(self):
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2src, self.feat2tgt,
bsz,
shuffle=True,
sort_key='{}_lengths'.format(self.primary_key)
)
self.valid_data = self.reader.load(
self.dataset['valid_file'],
self.feat2src, self.feat2tgt,
vbsz,
shuffle=True
)
self.test_data = None
if 'test_file' in self.dataset:
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2src, self.feat2tgt,
tbsz,
)
def _create_model(self):
self.config_params['model']["unif"] = self.config_params["unif"]
model = self.config_params['model']
unif = self.config_params.get('unif', 0.1)
model['unif'] = model.get('unif', unif)
lengths_key = model.get('src_lengths_key', self.primary_key)
if lengths_key is not None:
if not lengths_key.endswith('_lengths'):
lengths_key = '{}_lengths'.format(lengths_key)
model['src_lengths_key'] = lengths_key
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_seq2seq_model(self.src_embeddings, self.tgt_embeddings, **self.config_params['model'])
def train(self, checkpoint=None):
num_ex = self.config_params['num_valid_to_show']
rlut1 = revlut(self.feat2src[self.primary_key])
rlut2 = revlut(self.feat2tgt)
if num_ex > 0:
logger.info('Showing examples')
preproc = self.config_params.get('preproc', {})
show_ex_fn = preproc['show_ex']
self.config_params['train']['after_train_fn'] = lambda model: show_ex_fn(model,
self.valid_data, rlut1, rlut2,
self.feat2tgt,
preproc['mxlen'], False, 0,
num_ex, reverse=False)
self.config_params['train']['tgt_rlut'] = rlut2
return super(EncoderDecoderTask, self).train(checkpoint)
class EncoderDecoderTask(Task):
def __init__(self, logging_file, mead_config, **kwargs):
super(EncoderDecoderTask, self).__init__(logging_file, mead_config,
**kwargs)
self.task = None
def _create_task_specific_reader(self):
preproc = self.config_params['preproc']
reader = baseline.create_parallel_corpus_reader(
preproc['mxlen'], preproc['vec_alloc'], preproc['trim'],
preproc['word_trans_fn'], **self.config_params['loader'])
return reader
def _setup_task(self):
# If its not vanilla seq2seq, dont bother reversing
do_reverse = self.config_params['model']['model_type'] == 'default'
backend = self.config_params.get('backend', 'tensorflow')
if backend == 'pytorch':
print('PyTorch backend')
from baseline.pytorch import long_0_tensor_alloc as vec_alloc
from baseline.pytorch import tensor_shape as vec_shape
from baseline.pytorch import tensor_reverse_2nd as rev2nd
import baseline.pytorch.seq2seq as seq2seq
self.config_params['preproc']['vec_alloc'] = vec_alloc
self.config_params['preproc']['vec_shape'] = vec_shape
src_vec_trans = rev2nd if do_reverse else None
self.config_params['preproc']['word_trans_fn'] = src_vec_trans
self.config_params['preproc'][
'show_ex'] = baseline.pytorch.show_examples_pytorch
self.config_params['preproc']['trim'] = True
else:
import baseline.tf.seq2seq as seq2seq
import mead.tf
self.ExporterType = mead.tf.Seq2SeqTensorFlowExporter
self.config_params['preproc']['vec_alloc'] = np.zeros
self.config_params['preproc']['vec_shape'] = np.shape
self.config_params['preproc']['trim'] = False
src_vec_trans = baseline.reverse_2nd if do_reverse else None
self.config_params['preproc']['word_trans_fn'] = src_vec_trans
self.config_params['preproc'][
'show_ex'] = baseline.tf.show_examples_tf
self.task = seq2seq
def initialize(self, embeddings):
embeddings_set = mead.utils.index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache,
True).download()
print(
"[train file]: {}\n[valid file]: {}\n[test file]: {}\n[vocab file]: {}"
.format(self.dataset['train_file'], self.dataset['valid_file'],
self.dataset['test_file'],
self.dataset.get('vocab_file', "None")))
vocab_file = self.dataset.get('vocab_file', None)
if vocab_file is not None:
vocab1, vocab2 = self.reader.build_vocabs([vocab_file])
else:
vocab1, vocab2 = self.reader.build_vocabs([
self.dataset['train_file'], self.dataset['valid_file'],
self.dataset['test_file']
])
self.embeddings1, self.feat2index1 = self._create_embeddings(
embeddings_set, {'word': vocab1})
self.embeddings2, self.feat2index2 = self._create_embeddings(
embeddings_set, {'word': vocab2})
def _load_dataset(self):
self.train_data = self.reader.load(self.dataset['train_file'],
self.feat2index1['word'],
self.feat2index2['word'],
self.config_params['batchsz'],
shuffle=True)
self.valid_data = self.reader.load(self.dataset['valid_file'],
self.feat2index1['word'],
self.feat2index2['word'],
self.config_params['batchsz'],
shuffle=True)
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2index1['word'], self.feat2index2['word'],
self.config_params.get('test_batchsz', 1))
def _create_model(self):
return self.task.create_model(self.embeddings1['word'],
self.embeddings2['word'],
**self.config_params['model'])
def train(self):
num_ex = self.config_params['num_valid_to_show']
if num_ex > 0:
print('Showing examples')
preproc = self.config_params['preproc']
show_ex_fn = preproc['show_ex']
rlut1 = baseline.revlut(self.feat2index1['word'])
rlut2 = baseline.revlut(self.feat2index2['word'])
self.config_params['train'][
'after_train_fn'] = lambda model: show_ex_fn(model,
self.valid_data,
rlut1,
rlut2,
self.embeddings2[
'word'],
preproc['mxlen'],
False,
0,
num_ex,
reverse=False)
super(EncoderDecoderTask, self).train()
class ClassifierTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(ClassifierTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'classify'
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if backend.name == 'dy':
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
self.config_params['train']['trainer_type'] = 'autobatch'
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': batched}
backend.load(self.task_name())
return backend
def _setup_task(self, **kwargs):
super(ClassifierTask, self)._setup_task(**kwargs)
if self.config_params.get('preproc', {}).get('clean', False) is True:
self.config_params.get('preproc', {})['clean_fn'] = baseline.TSVSeqLabelReader.do_clean
logger.info('Clean')
else:
self.config_params.setdefault('preproc', {})
self.config_params['preproc']['clean_fn'] = None
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
vocab_sources = [self.dataset['train_file'], self.dataset['valid_file']]
# TODO: make this optional
if 'test_file' in self.dataset:
vocab_sources.append(self.dataset['test_file'])
vocab, self.labels = self.reader.build_vocab(vocab_sources,
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file'),
label_file=self.dataset.get('label_file'))
self.embeddings, self.feat2index = self._create_embeddings(embeddings_set, vocab, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _create_model(self):
unif = self.config_params.get('unif', 0.1)
model = self.config_params['model']
model['unif'] = model.get('unif', unif)
lengths_key = model.get('lengths_key', self.primary_key)
if lengths_key is not None:
if not lengths_key.endswith('_lengths'):
lengths_key = '{}_lengths'.format(lengths_key)
model['lengths_key'] = lengths_key
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_model(self.embeddings, self.labels, **model)
def _load_dataset(self):
read = self.config_params['reader'] if 'reader' in self.config_params else self.config_params['loader']
sort_key = read.get('sort_key')
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2index,
bsz,
shuffle=True,
sort_key=sort_key,
)
self.valid_data = self.reader.load(
self.dataset['valid_file'],
self.feat2index,
vbsz,
)
self.test_data = None
if 'test_file' in self.dataset:
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2index,
tbsz,
)
class LanguageModelingTask(Task):
def __init__(self, logging_config, mead_settings_config, **kwargs):
super(LanguageModelingTask,
self).__init__(logging_config, mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'lm'
def _create_task_specific_reader(self):
self._create_vectorizers()
reader_params = self.config_params['loader']
reader_params['nctx'] = reader_params.get(
'nctx',
self.config_params.get('nctx', self.config_params.get('nbptt',
35)))
reader_params['clean_fn'] = reader_params.get(
'clean_fn',
self.config_params.get('preproc', {}).get('clean_fn'))
reader_params['mxlen'] = self.vectorizers[self.primary_key].mxlen
if self.config_params['model'].get('gpus', 1) > 1:
reader_params['truncate'] = True
return baseline.reader.create_reader(
self.task_name(), self.vectorizers,
self.config_params['preproc'].get('trim', False), **reader_params)
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if backend.name == 'pytorch':
self.config_params.get('preproc', {})['trim'] = True
elif backend.name == 'dy':
self.config_params.get('preproc', {})['trim'] = True
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {
'pc': _dynet.ParameterCollection(),
'batched': batched
}
backend.load(self.task_name())
return backend
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset,
self.data_download_cache).download()
print_dataset_info(self.dataset)
vocabs = self.reader.build_vocab(
[
self.dataset['train_file'], self.dataset['valid_file'],
self.dataset['test_file']
],
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file'))
self.embeddings, self.feat2index = self._create_embeddings(
embeddings_set, vocabs, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _load_dataset(self):
tgt_key = self.config_params['loader'].get('tgt_key', self.primary_key)
self.train_data = self.reader.load(self.dataset['train_file'],
self.feat2index,
self.config_params['batchsz'],
tgt_key=tgt_key)
self.valid_data = self.reader.load(self.dataset['valid_file'],
self.feat2index,
self.config_params.get(
'valid_batchsz',
self.config_params['batchsz']),
tgt_key=tgt_key)
self.test_data = self.reader.load(self.dataset['test_file'],
self.feat2index,
1,
tgt_key=tgt_key)
def _create_model(self):
model = self.config_params['model']
unif = self.config_params.get('unif', 0.1)
model['unif'] = model.get('unif', unif)
model['batchsz'] = self.config_params['batchsz']
model['tgt_key'] = self.config_params['loader'].get(
'tgt_key', self.primary_key)
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_lang_model(self.embeddings, **model)
def train(self):
self._load_dataset()
if self.config_params['train'].get('lr_scheduler_type',
None) == 'zaremba':
first_range = int(
self.config_params['train']['start_decay_epoch'] *
self.train_data.steps)
self.config_params['train']['bounds'] = [first_range] + list(
np.arange(self.config_params['train']['start_decay_epoch'] + 1,
self.config_params['train']['epochs'] + 1,
dtype=np.int32) * self.train_data.steps)
baseline.save_vectorizers(self.get_basedir(), self.vectorizers)
model = self._create_model()
baseline.train.fit(model, self.train_data, self.valid_data,
self.test_data, **self.config_params['train'])
baseline.zip_files(self.get_basedir())
self._close_reporting_hooks()
@staticmethod
def _num_steps_per_epoch(num_examples, nctx, batchsz):
rest = num_examples // batchsz
return rest // nctx
class ClassifierTask(Task):
def __init__(self, logging_config, mead_settings_config, **kwargs):
super(ClassifierTask, self).__init__(logging_config,
mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'classify'
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if backend.name == 'dy':
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
self.config_params['train']['trainer_type'] = 'autobatch'
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {
'pc': _dynet.ParameterCollection(),
'batched': batched
}
elif backend.name == 'tf':
# FIXME this should be registered as well!
exporter_type = kwargs.get('exporter_type', 'default')
if exporter_type == 'default':
from mead.tf.exporters import ClassifyTensorFlowExporter
backend.exporter = ClassifyTensorFlowExporter
elif exporter_type == 'preproc':
from mead.tf.preproc_exporters import ClassifyTensorFlowPreProcExporter
import mead.tf.preprocessors
backend.exporter = ClassifyTensorFlowPreProcExporter
backend.load(self.task_name())
return backend
def _setup_task(self, **kwargs):
super(ClassifierTask, self)._setup_task(**kwargs)
if self.config_params.get('preproc', {}).get('clean', False) is True:
self.config_params.get(
'preproc',
{})['clean_fn'] = baseline.TSVSeqLabelReader.do_clean
print('Clean')
else:
self.config_params['preproc'] = {}
self.config_params['preproc']['clean_fn'] = None
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset,
self.data_download_cache).download()
print_dataset_info(self.dataset)
vocab, self.labels = self.reader.build_vocab(
[
self.dataset['train_file'], self.dataset['valid_file'],
self.dataset['test_file']
],
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file'),
label_file=self.dataset.get('label_file'))
self.embeddings, self.feat2index = self._create_embeddings(
embeddings_set, vocab, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _create_model(self):
unif = self.config_params.get('unif', 0.1)
model = self.config_params['model']
model['unif'] = model.get('unif', unif)
lengths_key = model.get('lengths_key', self.primary_key)
if lengths_key is not None:
if not lengths_key.endswith('_lengths'):
lengths_key = '{}_lengths'.format(lengths_key)
model['lengths_key'] = lengths_key
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_model(self.embeddings, self.labels,
**model)
def _load_dataset(self):
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2index,
self.config_params['batchsz'],
shuffle=True,
sort_key=self.config_params['loader'].get('sort_key'))
self.valid_data = self.reader.load(
self.dataset['valid_file'], self.feat2index,
self.config_params.get('valid_batchsz',
self.config_params['batchsz']))
self.test_data = self.reader.load(
self.dataset['test_file'], self.feat2index,
self.config_params.get('test_batchsz', 1))
class LanguageModelingTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(LanguageModelingTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'lm'
def _create_task_specific_reader(self):
self._create_vectorizers()
reader_params = self.config_params['reader'] if 'reader' in self.config_params else self.config_params['loader']
reader_params['nctx'] = reader_params.get('nctx', self.config_params.get('nctx', self.config_params.get('nbptt', 35)))
reader_params['clean_fn'] = reader_params.get('clean_fn', self.config_params.get('preproc', {}).get('clean_fn'))
if reader_params['clean_fn'] is not None and self.config_params['dataset'] != 'SST2':
logger.warning('Warning: A reader preprocessing function (%s) is active, it is recommended that all data preprocessing is done outside of baseline to insure data at inference time matches data at training time.', reader_params['clean_fn'])
reader_params['mxlen'] = self.vectorizers[self.primary_key].mxlen
if self.config_params['model'].get('gpus', 1) > 1:
reader_params['truncate'] = True
return baseline.reader.create_reader(self.task_name(), self.vectorizers, self.config_params['preproc'].get('trim', False), **reader_params)
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if backend.name == 'pytorch':
self.config_params.get('preproc', {})['trim'] = True
elif backend.name == 'dy':
self.config_params.get('preproc', {})['trim'] = True
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': batched}
backend.load(self.task_name())
return backend
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
vocabs = self.reader.build_vocab(
[self.dataset['train_file'], self.dataset['valid_file'], self.dataset['test_file']],
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file')
)
self.embeddings, self.feat2index = self._create_embeddings(embeddings_set, vocabs, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _load_dataset(self):
read = self.config_params['reader'] if 'reader' in self.config_params else self.config_params['loader']
tgt_key = read.get('tgt_key', self.primary_key)
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2index,
bsz,
tgt_key=tgt_key
)
self.valid_data = self.reader.load(
self.dataset['valid_file'],
self.feat2index,
vbsz,
tgt_key=tgt_key
)
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2index,
1,
tgt_key=tgt_key
)
def _create_model(self):
model = self.config_params['model']
unif = self.config_params.get('unif', 0.1)
model['unif'] = model.get('unif', unif)
model['batchsz'] = self.config_params['batchsz']
model['tgt_key'] = self.config_params.get('reader', self.config_params.get('loader', {})).get('tgt_key', self.primary_key)
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_lang_model(self.embeddings, **model)
def train(self, checkpoint=None):
self._load_dataset()
if self.config_params['train'].get('lr_scheduler_type', None) == 'zaremba':
first_range = int(self.config_params['train']['start_decay_epoch'] * self.train_data.steps)
self.config_params['train']['bounds'] = [first_range] + list(
np.arange(
self.config_params['train']['start_decay_epoch'] + 1,
self.config_params['train']['epochs'] + 1,
dtype=np.int32
) * self.train_data.steps
)
baseline.save_vectorizers(self.get_basedir(), self.vectorizers)
model = self._create_model()
train_params = self.config_params['train']
train_params['checkpoint'] = checkpoint
metrics = baseline.train.fit(model, self.train_data, self.valid_data, self.test_data, **train_params)
baseline.zip_files(self.get_basedir())
self._close_reporting_hooks()
return model, metrics
@staticmethod
def _num_steps_per_epoch(num_examples, nctx, batchsz):
rest = num_examples // batchsz
return rest // nctx
class EncoderDecoderTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(EncoderDecoderTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'seq2seq'
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if 'preproc' not in self.config_params:
self.config_params['preproc'] = {}
self.config_params['preproc']['show_ex'] = show_examples
if backend.name == 'pytorch':
self.config_params['preproc']['trim'] = True
elif backend.name == 'dy':
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
self.config_params['train']['trainer_type'] = 'autobatch'
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': batched}
self.config_params['preproc']['trim'] = True
else:
self.config_params['preproc']['trim'] = True
backend.load(self.task_name())
return backend
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
vocab1, vocab2 = self.reader.build_vocabs(
[self.dataset['train_file'], self.dataset['valid_file'], self.dataset['test_file']],
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file')
)
# To keep the config file simple, share a list between source and destination (tgt)
features_src = []
features_tgt = None
for feature in self.config_params['features']:
if feature['name'] == 'tgt':
features_tgt = feature
else:
features_src += [feature]
self.src_embeddings, self.feat2src = self._create_embeddings(embeddings_set, vocab1, features_src)
# For now, dont allow multiple vocabs of output
baseline.save_vocabs(self.get_basedir(), self.feat2src)
self.tgt_embeddings, self.feat2tgt = self._create_embeddings(embeddings_set, {'tgt': vocab2}, [features_tgt])
baseline.save_vocabs(self.get_basedir(), self.feat2tgt)
self.tgt_embeddings = self.tgt_embeddings['tgt']
self.feat2tgt = self.feat2tgt['tgt']
def _load_dataset(self):
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2src, self.feat2tgt,
bsz,
shuffle=True,
sort_key='{}_lengths'.format(self.primary_key)
)
self.valid_data = self.reader.load(
self.dataset['valid_file'],
self.feat2src, self.feat2tgt,
vbsz,
shuffle=True
)
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2src, self.feat2tgt,
tbsz,
)
def _create_model(self):
self.config_params['model']["unif"] = self.config_params["unif"]
model = self.config_params['model']
unif = self.config_params.get('unif', 0.1)
model['unif'] = model.get('unif', unif)
lengths_key = model.get('src_lengths_key', self.primary_key)
if lengths_key is not None:
if not lengths_key.endswith('_lengths'):
lengths_key = '{}_lengths'.format(lengths_key)
model['src_lengths_key'] = lengths_key
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_seq2seq_model(self.src_embeddings, self.tgt_embeddings, **self.config_params['model'])
def train(self, checkpoint=None):
num_ex = self.config_params['num_valid_to_show']
rlut1 = revlut(self.feat2src[self.primary_key])
rlut2 = revlut(self.feat2tgt)
if num_ex > 0:
logger.info('Showing examples')
preproc = self.config_params.get('preproc', {})
show_ex_fn = preproc['show_ex']
self.config_params['train']['after_train_fn'] = lambda model: show_ex_fn(model,
self.valid_data, rlut1, rlut2,
self.feat2tgt,
preproc['mxlen'], False, 0,
num_ex, reverse=False)
self.config_params['train']['tgt_rlut'] = rlut2
return super(EncoderDecoderTask, self).train(checkpoint)
class ClassifierTask(Task):
def __init__(self, mead_settings_config, **kwargs):
super(ClassifierTask, self).__init__(mead_settings_config, **kwargs)
@classmethod
def task_name(cls):
return 'classify'
def _create_backend(self, **kwargs):
backend = Backend(self.config_params.get('backend', 'tf'))
if backend.name == 'dy':
import _dynet
dy_params = _dynet.DynetParams()
dy_params.from_args()
dy_params.set_requested_gpus(1)
if 'autobatchsz' in self.config_params['train']:
self.config_params['train']['trainer_type'] = 'autobatch'
dy_params.set_autobatch(True)
batched = False
else:
batched = True
dy_params.init()
backend.params = {'pc': _dynet.ParameterCollection(), 'batched': batched}
backend.load(self.task_name())
return backend
def _setup_task(self, **kwargs):
super(ClassifierTask, self)._setup_task(**kwargs)
if self.config_params.get('preproc', {}).get('clean', False) is True:
self.config_params.get('preproc', {})['clean_fn'] = baseline.TSVSeqLabelReader.do_clean
logger.info('Clean')
else:
self.config_params['preproc'] = {}
self.config_params['preproc']['clean_fn'] = None
def initialize(self, embeddings):
embeddings = read_config_file_or_json(embeddings, 'embeddings')
embeddings_set = index_by_label(embeddings)
self.dataset = DataDownloader(self.dataset, self.data_download_cache).download()
print_dataset_info(self.dataset)
vocab, self.labels = self.reader.build_vocab(
[self.dataset['train_file'], self.dataset['valid_file'], self.dataset['test_file']],
min_f=Task._get_min_f(self.config_params),
vocab_file=self.dataset.get('vocab_file'),
label_file=self.dataset.get('label_file')
)
self.embeddings, self.feat2index = self._create_embeddings(embeddings_set, vocab, self.config_params['features'])
baseline.save_vocabs(self.get_basedir(), self.feat2index)
def _create_model(self):
unif = self.config_params.get('unif', 0.1)
model = self.config_params['model']
model['unif'] = model.get('unif', unif)
lengths_key = model.get('lengths_key', self.primary_key)
if lengths_key is not None:
if not lengths_key.endswith('_lengths'):
lengths_key = '{}_lengths'.format(lengths_key)
model['lengths_key'] = lengths_key
if self.backend.params is not None:
for k, v in self.backend.params.items():
model[k] = v
return baseline.model.create_model(self.embeddings, self.labels, **model)
def _load_dataset(self):
read = self.config_params['reader'] if 'reader' in self.config_params else self.config_params['loader']
sort_key = read.get('sort_key')
bsz, vbsz, tbsz = Task._get_batchsz(self.config_params)
self.train_data = self.reader.load(
self.dataset['train_file'],
self.feat2index,
bsz,
shuffle=True,
sort_key=sort_key,
)
self.valid_data = self.reader.load(
self.dataset['valid_file'],
self.feat2index,
vbsz,
)
self.test_data = self.reader.load(
self.dataset['test_file'],
self.feat2index,
tbsz,
)
