def get_triviaqa_train_params(n_epochs, n_dev, n_train):
return TrainParams(SerializableOptimizer("Adadelta",
dict(learning_rate=1)),
num_epochs=n_epochs,
ema=0.9999,
max_checkpoints_to_keep=2,
async_encoding=10,
log_period=30,
eval_period=1800,
save_period=1800,
eval_samples=dict(dev=n_dev, train=n_train))
def train_params(n_epochs):
return TrainParams(SerializableOptimizer("Adadelta",
dict(learning_rate=1.0)),
ema=0.999,
max_checkpoints_to_keep=3,
async_encoding=10,
num_epochs=n_epochs,
log_period=30,
eval_period=1200,
save_period=1200,
eval_samples=dict(dev=None, train=8000))
def train_params(n_epochs):
return TrainParams(SerializableOptimizer("Adadelta",
dict(learning_rate=1.0)),
ema=0.999,
max_checkpoints_to_keep=3,
async_encoding=10,
num_epochs=n_epochs,
log_period=30,
eval_period=900,
save_period=300,
best_weights=("dev", "race data accuracy"),
eval_samples=dict(dev=None, train=5000))
def get_training_params(train_config):
return TrainParams(
SerializableOptimizer(
train_config.optimizer,
dict(learning_rate=train_config.learning_rate)
),
num_epochs=train_config.n_epochs,
ema=train_config.ema,
max_checkpoints_to_keep=train_config.max_checkpoints_to_keep,
async_encoding=train_config.async_encoding,
log_period=train_config.log_period,
eval_period=train_config.eval_period,
save_period=train_config.save_period,
best_weights=("dev", "b8/question-text-f1"),
eval_samples=dict(dev=None, train=6000),
eval_at_zero=False
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("corpus", choices=["en", "en_trans_de", "en_trans_zh"])
parser.add_argument(
'mode',
choices=["confidence", "merge", "shared-norm", "sigmoid", "paragraph"])
# Note I haven't tested modes other than `shared-norm` on this corpus, so
# some things might need adjusting
parser.add_argument("name", help="Where to store the model")
parser.add_argument("-t",
"--n_tokens",
default=400,
type=int,
help="Paragraph size")
parser.add_argument(
'-n',
'--n_processes',
type=int,
default=2,
help="Number of processes (i.e., select which paragraphs to train on) "
"the data with")
args = parser.parse_args()
mode = args.mode
corpus = args.corpus
out = args.name + "-" + datetime.now().strftime("%m%d-%H%M%S")
model = get_model(100, 140, mode, WithIndicators())
extract = ExtractMultiParagraphsPerQuestion(MergeParagraphs(args.n_tokens),
ShallowOpenWebRanker(16),
model.preprocessor,
intern=True)
eval = [
LossEvaluator(),
MultiParagraphSpanEvaluator(8,
"triviaqa",
mode != "merge",
per_doc=False)
]
oversample = [
1
] * 2 # Sample the top two answer-containing paragraphs twice
if mode == "paragraph":
n_epochs = 120
test = RandomParagraphSetDatasetBuilder(120, "flatten", True,
oversample)
train = StratifyParagraphsBuilder(ClusteredBatcher(
60, ContextLenBucketedKey(3), True),
oversample,
only_answers=True)
elif mode == "confidence" or mode == "sigmoid":
if mode == "sigmoid":
n_epochs = 640
else:
n_epochs = 160
test = RandomParagraphSetDatasetBuilder(120, "flatten", True,
oversample)
train = StratifyParagraphsBuilder(
ClusteredBatcher(60, ContextLenBucketedKey(3), True), oversample)
else:
n_epochs = 80
test = RandomParagraphSetDatasetBuilder(
120, "merge" if mode == "merge" else "group", True, oversample)
train = StratifyParagraphSetsBuilder(30, mode == "merge", True,
oversample)
data = XQADataset(corpus)
params = TrainParams(SerializableOptimizer("Adadelta",
dict(learning_rate=1)),
num_epochs=n_epochs,
ema=0.999,
max_checkpoints_to_keep=2,
async_encoding=10,
log_period=30,
eval_period=1800,
save_period=1800,
best_weights=("dev", "b8/question-text-f1"),
eval_samples=dict(dev=None, train=6000))
data = PreprocessedData(data, extract, train, test, eval_on_verified=False)
data.preprocess(args.n_processes, 1000)
with open(__file__, "r") as f:
notes = f.read()
notes = "Mode: " + args.mode + "\n" + notes
trainer.start_training(data, model, params, eval, model_dir.ModelDir(out),
notes)
def main():
"""
A close-as-possible impelemntation of BiDaF, its based on the `dev` tensorflow 1.1 branch of Ming's repo
which, in particular, uses Adam not Adadelta. I was not able to replicate the results in paper using Adadelta,
but with Adam i was able to get to 78.0 F1 on the dev set with this scripts. I believe this approach is
an exact reproduction up the code in the repo, up to initializations.
Notes: Exponential Moving Average is very important, as is early stopping. This is also in particualr best run
on a GPU due to the large number of parameters and batch size involved.
"""
out = get_output_name_from_cli()
train_params = TrainParams(SerializableOptimizer(
"Adam", dict(learning_rate=0.001)),
num_epochs=12,
ema=0.999,
async_encoding=10,
log_period=30,
eval_period=1000,
save_period=1000,
eval_samples=dict(dev=None, train=8000))
# recurrent_layer = BiRecurrentMapper(LstmCellSpec(100, keep_probs=0.8))
# recurrent_layer = FusedLstm()
recurrent_layer = SequenceMapperSeq(DropoutLayer(0.8), CudnnLstm(100))
model = Attention(
encoder=DocumentAndQuestionEncoder(SingleSpanAnswerEncoder()),
word_embed=FixedWordEmbedder(vec_name="glove.6B.100d",
word_vec_init_scale=0,
learn_unk=False),
char_embed=CharWordEmbedder(embedder=LearnedCharEmbedder(16, 49, 8),
layer=ReduceLayer("max",
Conv1d(100, 5, 0.8),
mask=False),
shared_parameters=True),
word_embed_layer=None,
embed_mapper=SequenceMapperSeq(HighwayLayer(activation="relu"),
HighwayLayer(activation="relu"),
recurrent_layer),
preprocess=None,
question_mapper=None,
context_mapper=None,
memory_builder=NullBiMapper(),
attention=BiAttention(TriLinear(bias=True), True),
match_encoder=NullMapper(),
predictor=BoundsPredictor(
ChainConcat(start_layer=SequenceMapperSeq(recurrent_layer,
recurrent_layer),
end_layer=recurrent_layer)),
)
with open(__file__, "r") as f:
notes = f.read()
eval = [LossEvaluator(), SpanEvaluator(bound=[17], text_eval="squad")]
corpus = SquadCorpus()
train_batching = ClusteredBatcher(60, ContextLenBucketedKey(3), True,
False)
eval_batching = ClusteredBatcher(60, ContextLenKey(), False, False)
data = DocumentQaTrainingData(corpus, None, train_batching, eval_batching)
trainer.start_training(data, model, train_params, eval,
model_dir.ModelDir(out), notes)
def main():
parser = argparse.ArgumentParser(
description='Train a model on TriviaQA unfiltered')
parser.add_argument(
'mode',
choices=["confidence", "merge", "shared-norm", "sigmoid", "paragraph"])
parser.add_argument("name", help="Where to store the model")
parser.add_argument("-t",
"--n_tokens",
default=400,
type=int,
help="Paragraph size")
parser.add_argument(
'-n',
'--n_processes',
type=int,
default=2,
help="Number of processes (i.e., select which paragraphs to train on) "
"the data with")
parser.add_argument("-s",
"--source_dir",
type=str,
default=None,
help="where to take input files")
parser.add_argument("--n_epochs",
type=int,
default=None,
help="Max number of epoches to train on ")
parser.add_argument("--char_th",
type=int,
default=None,
help="char level embeddings")
parser.add_argument("--hl_dim",
type=int,
default=None,
help="hidden layer dim size")
parser.add_argument("--regularization",
type=int,
default=None,
help="hidden layer dim size")
parser.add_argument("--LR",
type=float,
default=1.0,
help="hidden layer dim size")
parser.add_argument("--save_every",
type=int,
default=1800,
help="save period")
parser.add_argument("--init_from",
type=str,
default=None,
help="model to init from")
args = parser.parse_args()
mode = args.mode
#out = args.name + "-" + datetime.now().strftime("%m%d-%H%M%S")
out = join('models', args.name)
char_th = 100
hl_dim = 140
if args.char_th is not None:
print(args.char_th)
char_th = int(args.char_th)
out += '--th' + str(char_th)
if args.hl_dim is not None:
print(args.hl_dim)
hl_dim = int(args.hl_dim)
out += '--hl' + str(hl_dim)
if args.init_from is None:
model = get_model(char_th, hl_dim, mode, WithIndicators())
else:
md = model_dir.ModelDir(args.init_from)
model = md.get_model()
extract = ExtractMultiParagraphsPerQuestion(MergeParagraphs(args.n_tokens),
ShallowOpenWebRanker(16),
model.preprocessor,
intern=True)
eval = [
LossEvaluator(),
MultiParagraphSpanEvaluator(8,
"triviaqa",
mode != "merge",
per_doc=False)
]
oversample = [1] * 4
if mode == "paragraph":
n_epochs = 120
test = RandomParagraphSetDatasetBuilder(120, "flatten", True,
oversample)
train = StratifyParagraphsBuilder(ClusteredBatcher(
60, ContextLenBucketedKey(3), True),
oversample,
only_answers=True)
elif mode == "confidence" or mode == "sigmoid":
if mode == "sigmoid":
n_epochs = 640
else:
n_epochs = 160
test = RandomParagraphSetDatasetBuilder(120, "flatten", True,
oversample)
train = StratifyParagraphsBuilder(
ClusteredBatcher(60, ContextLenBucketedKey(3), True), oversample)
else:
n_epochs = 80
test = RandomParagraphSetDatasetBuilder(
120, "merge" if mode == "merge" else "group", True, oversample)
train = StratifyParagraphSetsBuilder(30, mode == "merge", True,
oversample)
if args.n_epochs is not None:
n_epochs = args.n_epochs
out += '--' + str(n_epochs)
if args.LR != 1.0:
out += '--' + str(args.LR)
data = TriviaQaOpenDataset(args.source_dir)
async_encoding = 10
#async_encoding = 0
params = TrainParams(SerializableOptimizer("Adadelta",
dict(learning_rate=args.LR)),
num_epochs=n_epochs,
num_of_steps=250000,
ema=0.999,
max_checkpoints_to_keep=2,
async_encoding=async_encoding,
log_period=30,
eval_period=1800,
save_period=args.save_every,
eval_samples=dict(dev=None, train=6000),
regularization_weight=None)
data = PreprocessedData(data, extract, train, test, eval_on_verified=False)
data.preprocess(args.n_processes, 1000)
with open(__file__, "r") as f:
notes = f.read()
notes = "Mode: " + args.mode + "\n" + notes
if args.init_from is not None:
init_from = model_dir.ModelDir(args.init_from).get_best_weights()
if init_from is None:
init_from = model_dir.ModelDir(
args.init_from).get_latest_checkpoint()
else:
init_from = None
trainer.start_training(data,
model,
params,
eval,
model_dir.ModelDir(out),
notes,
initialize_from=init_from)