def train(self, training_data: TrainingData) -> None:
log.info('Preprocessing training data...')
x_train, y_train, x_test, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data, create_runs=self.train_on_q_runs, full_question=self.train_on_full_q)
if self.wiki_data_frac > 0:
log.info('Using wikipedia with fraction: {}'.format(self.wiki_data_frac))
wiki_data = FilteredWikipediaDataset().training_data()
results = preprocess_dataset(
wiki_data,
train_size=1,
vocab=vocab,
class_to_i=class_to_i,
i_to_class=i_to_class)
x_train.extend(results[0])
y_train.extend(results[1])
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
log.info('Creating embeddings...')
embeddings, embedding_lookup = load_embeddings(vocab=vocab, expand_glove=self.expand_we, mask_zero=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
log.info('Converting dataset to embeddings...')
x_train = [nn.convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_train]
x_test = [nn.convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_test]
self.n_classes = nn.compute_n_classes(training_data[1])
self.max_len = nn.compute_max_len(training_data)
x_train = np.array(nn.tf_format(x_train, self.max_len, 0))
x_test = np.array(nn.tf_format(x_test, self.max_len, 0))
log.info('Building keras model...')
self.model = self.build_model()
log.info('Training model...')
callbacks = [
TensorBoard(),
EarlyStopping(patience=self.max_patience, monitor='val_sparse_categorical_accuracy'),
ModelCheckpoint(
safe_path(CNN_MODEL_TMP_TARGET),
save_best_only=True,
monitor='val_sparse_categorical_accuracy'
)
]
if self.decay_lr_on_plateau:
callbacks.append(ReduceLROnPlateau(monitor='val_sparse_categorical_accuracy', factor=.5, patience=5))
history = self.model.fit(
x_train, y_train,
validation_data=(x_test, y_test),
batch_size=self.batch_size, epochs=self.max_n_epochs,
callbacks=callbacks, verbose=2
)
self.history = history.history
log.info('Done training')
def train(self,
training_data: TrainingData) -> None:
log.info('Preprocessing training data...')
x_train, y_train, _, x_test, y_test, _, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
if self.use_wiki:
wiki_training_data = WikipediaDataset(self.min_answers).training_data()
x_train_wiki, y_train_wiki, _, _, _, _, _, _, _ = preprocess_dataset(
wiki_training_data, train_size=1, vocab=vocab, class_to_i=class_to_i,
i_to_class=i_to_class)
log.info('Creating embeddings...')
embeddings, embedding_lookup = _load_embeddings(vocab=vocab)
self.embeddings = embeddings
self.embedding_lookup = self.embedding_lookup
log.info('Converting dataset to embeddings...')
x_train = [_convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_train]
x_train_lengths = _compute_lengths(x_train)
x_test = [_convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_test]
x_test_lengths = _compute_lengths(x_test)
if self.use_wiki:
x_train_wiki = [_convert_text_to_embeddings_indices(q, embedding_lookup)
for q in x_train_wiki]
x_train_lengths_wiki = _compute_lengths(x_train_wiki)
x_train.extend(x_train_wiki)
y_train.extend(y_train_wiki)
x_train_lengths = np.concatenate([x_train_lengths, x_train_lengths_wiki])
log.info('Computing number of classes and max paragraph length in words')
self.n_classes = _compute_n_classes(training_data[1])
self.max_len = _compute_max_len(x_train)
x_train = _tf_format(x_train, self.max_len, embeddings.shape[0])
x_test = _tf_format(x_test, self.max_len, embeddings.shape[0])
log.info('Training deep model...')
self.model = TFDanModel(self.dan_params, self.max_len, self.n_classes)
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = np.array(x_test)
y_test = np.array(y_test)
train_losses, train_accuracies, holdout_losses, holdout_accuracies = self.model.train(
x_train, y_train, x_train_lengths, x_test, y_test, x_test_lengths)
def train(self, training_data):
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data
)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
embeddings, embedding_lookup = load_embeddings(vocab=vocab, expand_glove=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
x_train = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_train_text]
for row in x_train:
if len(row) == 0:
row.append(embedding_lookup['UNK'])
x_train = np.array(x_train)
y_train = np.array(y_train)
mems_train = load_memories([' '.join(x) for x in x_train_text], self.n_memories)
mems_indices_train = memories_to_indices(mems_train, embedding_lookup)
x_test = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_test_text]
for row in x_test:
if len(row) == 0:
row.append(embedding_lookup['UNK'])
x_test = np.array(x_test)
y_test = np.array(y_test)
mems_test = load_memories([' '.join(x) for x in x_test_text], self.n_memories)
mem_indices_test = memories_to_indices(mems_test, embedding_lookup)
self.n_classes = compute_n_classes(training_data[1])
def get_quizbowl():
qb_dataset = QuizBowlDataset(guesser_train=True, buzzer_train=False)
training_data = qb_dataset.training_data()
train_x, train_y, dev_x, dev_y, i_to_word, class_to_i, i_to_class = preprocess_dataset(
training_data)
i_to_word = ['<unk>', '<eos>'] + sorted(i_to_word)
word_to_i = {x: i for i, x in enumerate(i_to_word)}
train = transform_to_array(zip(train_x, train_y), word_to_i)
dev = transform_to_array(zip(dev_x, dev_y), word_to_i)
return train, dev, word_to_i, i_to_class
def train(self, training_data: TrainingData) -> None:
x_train, y_train, x_val, y_val, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
log.info('Batchifying data')
train_batches = batchify(x_train, y_train, shuffle=True)
val_batches = batchify(x_val, y_val, shuffle=False)
self.model = ElmoModel(len(i_to_class), dropout=self.dropout)
if CUDA:
self.model = self.model.cuda()
log.info(f'Parameters:\n{self.parameters()}')
log.info(f'Model:\n{self.model}')
parameters = list(self.model.classifier.parameters())
for mix in self.model.elmo._scalar_mixes:
parameters.extend(list(mix.parameters()))
self.optimizer = Adam(parameters)
self.criterion = nn.CrossEntropyLoss()
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer,
patience=5,
verbose=True,
mode='max')
temp_prefix = get_tmp_filename()
self.model_file = f'{temp_prefix}.pt'
manager = TrainingManager([
BaseLogger(log_func=log.info),
TerminateOnNaN(),
EarlyStopping(monitor='test_acc', patience=10, verbose=1),
MaxEpochStopping(100),
ModelCheckpoint(create_save_model(self.model),
self.model_file,
monitor='test_acc')
])
log.info('Starting training')
epoch = 0
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(train_batches)
random.shuffle(train_batches)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(val_batches,
train=False)
stop_training, reasons = manager.instruct(train_time, train_loss,
train_acc, test_time,
test_loss, test_acc)
if stop_training:
log.info(' '.join(reasons))
break
else:
self.scheduler.step(test_acc)
epoch += 1
def train(self, training_data: TrainingData) -> None:
x_train, y_train, x_val, y_val, vocab, class_to_i, i_to_class = preprocess_dataset(training_data)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
log.info('Batchifying data')
train_batches = batchify(x_train, y_train, shuffle=True)
val_batches = batchify(x_val, y_val, shuffle=False)
self.model = ElmoModel(len(i_to_class), dropout=self.dropout)
if CUDA:
self.model = self.model.cuda()
log.info(f'Parameters:\n{self.parameters()}')
log.info(f'Model:\n{self.model}')
parameters = list(self.model.classifier.parameters())
for mix in self.model.elmo._scalar_mixes:
parameters.extend(list(mix.parameters()))
self.optimizer = Adam(parameters)
self.criterion = nn.CrossEntropyLoss()
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer, patience=5, verbose=True, mode='max')
temp_prefix = get_tmp_filename()
self.model_file = f'{temp_prefix}.pt'
manager = TrainingManager([
BaseLogger(log_func=log.info), TerminateOnNaN(), EarlyStopping(monitor='test_acc', patience=10, verbose=1),
MaxEpochStopping(100), ModelCheckpoint(create_save_model(self.model), self.model_file, monitor='test_acc')
])
log.info('Starting training')
epoch = 0
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(train_batches)
random.shuffle(train_batches)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(val_batches, train=False)
stop_training, reasons = manager.instruct(
train_time, train_loss, train_acc,
test_time, test_loss, test_acc
)
if stop_training:
log.info(' '.join(reasons))
break
else:
self.scheduler.step(test_acc)
epoch += 1
def train(self, training_data: TrainingData) -> None:
log.info('Preprocessing training data...')
x_train, y_train, properties_train, x_test, y_test, properties_test, vocab,\
class_to_i, i_to_class = preprocess_dataset(training_data)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
self.ans_type_i_to_class, self.ans_type_class_to_i = compute_ans_type_classes(
properties_train)
self.n_ans_type_classes = len(self.ans_type_class_to_i)
compute_ans_type_classes(properties_test)
self.category_i_to_class, self.category_class_to_i = compute_category_classes(
properties_train)
self.n_category_classes = len(self.category_class_to_i)
compute_category_classes(properties_test)
self.gender_i_to_class, self.gender_class_to_i = compute_gender_classes(
properties_train)
self.n_gender_classes = len(self.gender_class_to_i)
compute_gender_classes(properties_test)
ans_type_labels_train = np.array([
self.ans_type_class_to_i[prop['ans_type']]
for prop in properties_train
])
category_labels_train = np.array([
self.category_class_to_i[prop['category']]
for prop in properties_train
])
gender_labels_train = np.array([
self.gender_class_to_i[prop['gender']] for prop in properties_train
])
ans_type_labels_test = np.array([
self.ans_type_class_to_i[prop['ans_type']]
for prop in properties_test
])
category_labels_test = np.array([
self.category_class_to_i[prop['category']]
for prop in properties_test
])
gender_labels_test = np.array([
self.gender_class_to_i[prop['gender']] for prop in properties_test
])
if self.use_wiki:
wiki_training_data = WikipediaDataset(
self.min_answers).training_data()
x_train_wiki, y_train_wiki, _, _, _, _, _, _, _ = preprocess_dataset(
wiki_training_data,
train_size=1,
vocab=vocab,
class_to_i=class_to_i,
i_to_class=i_to_class)
log.info('Creating embeddings...')
embeddings, embedding_lookup = _load_embeddings(vocab=vocab)
self.embeddings = embeddings
self.embedding_lookup = self.embedding_lookup
log.info('Converting dataset to embeddings...')
x_train = [
convert_text_to_embeddings_indices(q, embedding_lookup)
for q in x_train
]
x_train_lengths = compute_lengths(x_train)
x_test = [
convert_text_to_embeddings_indices(q, embedding_lookup)
for q in x_test
]
x_test_lengths = compute_lengths(x_test)
log.info(
'Computing number of classes and max paragraph length in words')
self.n_classes = compute_n_classes(training_data[1])
self.max_len = compute_max_len(x_train)
x_train = tf_format(x_train, self.max_len, embeddings.shape[0])
x_test = tf_format(x_test, self.max_len, embeddings.shape[0])
log.info('Training deep model...')
self.model = AuxDanModel(self.dan_params, self.max_len, self.n_classes,
self.n_ans_type_classes,
self.n_gender_classes,
self.n_category_classes)
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = np.array(x_test)
y_test = np.array(y_test)
property_labels_train = {
'ans_type': ans_type_labels_train,
'category': category_labels_train,
'gender': gender_labels_train
}
property_labels_test = {
'ans_type': ans_type_labels_test,
'category': category_labels_test,
'gender': gender_labels_test
}
train_losses, train_accuracies, holdout_losses, holdout_accuracies = self.model.train(
x_train, y_train, x_train_lengths, property_labels_train, x_test,
y_test, x_test_lengths, property_labels_test)
def train(self, training_data: TrainingData) -> None:
if self.use_qb:
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data)
if self.use_wiki:
wiki_dataset = WikipediaDataset(set(training_data[1]))
wiki_train_data = wiki_dataset.training_data()
w_x_train_text, w_train_y, _, _, _, _, _ = preprocess_dataset(
wiki_train_data,
train_size=1,
vocab=vocab,
class_to_i=class_to_i,
i_to_class=i_to_class)
x_train_text.extend(w_x_train_text)
y_train.extend(w_train_y)
else:
if self.use_wiki:
wiki_dataset = WikipediaDataset(set(training_data[1]))
wiki_train_data = wiki_dataset.training_data()
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
wiki_train_data)
else:
raise ValueError(
'use_wiki and use_qb cannot both be false, otherwise there is no training data'
)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
embeddings, embedding_lookup = load_embeddings(vocab=vocab,
expand_glove=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
x_train = [
convert_text_to_embeddings_indices(q, embedding_lookup)
for q in x_train_text
]
for r in x_train:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = [
convert_text_to_embeddings_indices(q, embedding_lookup)
for q in x_test_text
]
for r in x_test:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_test = np.array(x_test)
y_test = np.array(y_test)
self.n_classes = compute_n_classes(training_data[1])
n_batches_train, t_x_train, t_offset_train, t_y_train = batchify(
self.batch_size, x_train, y_train, truncate=True)
n_batches_test, t_x_test, t_offset_test, t_y_test = batchify(
self.batch_size, x_test, y_test, truncate=False)
self.vocab_size = embeddings.shape[0]
self.model = DanModel(self.vocab_size,
self.n_classes,
dropout_prob=self.dropout_prob,
k_softmaxes=self.k_softmaxes,
n_hidden_units=self.n_hidden_units,
non_linearity=self.non_linearity)
log.info(f'Parameters:\n{pformat(self.parameters())}')
log.info(f'Torch Model:\n{self.model}')
self.model.init_weights(initial_embeddings=embeddings)
if CUDA:
self.model = self.model.cuda()
self.optimizer = Adam(self.model.parameters(), lr=self.learning_rate)
self.criterion = nn.NLLLoss()
# self.criterion = nn.CrossEntropyLoss()
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer,
patience=5,
verbose=True)
tb_experiment = ' '.join(
f'{param}={value}' for param, value in
[('model',
'dan'), ('n_hidden_units',
self.n_hidden_units), ('dropout_prob',
self.dropout_prob),
('k_softmaxes',
self.k_softmaxes), ('non_linearity', self.non_linearity),
('learning_rate',
self.learning_rate), ('batch_size', self.batch_size)])
manager = TrainingManager([
BaseLogger(log_func=log.info),
TerminateOnNaN(),
EarlyStopping(monitor='test_loss', patience=10, verbose=1),
MaxEpochStopping(100),
ModelCheckpoint(create_save_model(self.model),
'/tmp/dan.pt',
monitor='test_loss'),
Tensorboard(tb_experiment)
])
log.info('Starting training...')
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(n_batches_train,
t_x_train,
t_offset_train,
t_y_train,
evaluate=False)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(n_batches_test,
t_x_test,
t_offset_test,
t_y_test,
evaluate=True)
stop_training, reasons = manager.instruct(train_time, train_loss,
train_acc, test_time,
test_loss, test_acc)
if stop_training:
log.info(' '.join(reasons))
break
else:
self.scheduler.step(test_loss)
log.info('Done training')
def train(self, training_data: TrainingData) -> None:
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data
)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
embeddings, embedding_lookup = load_embeddings(vocab=vocab, expand_glove=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
x_train = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_train_text]
for r in x_train:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_test_text]
for r in x_test:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_test = np.array(x_test)
y_test = np.array(y_test)
self.n_classes = compute_n_classes(training_data[1])
n_batches_train, t_x_train, t_offset_train, t_y_train = batchify(
self.batch_size, x_train, y_train, truncate=True)
n_batches_test, t_x_test, t_offset_test, t_y_test = batchify(
self.batch_size, x_test, y_test, truncate=False)
self.model = DanModel(embeddings.shape[0], self.n_classes)
self.model.init_weights(initial_embeddings=embeddings)
self.model.cuda()
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.learning_rate)
self.criterion = nn.CrossEntropyLoss()
manager = TrainingManager([
BaseLogger(log_func=log.info), TerminateOnNaN(),
EarlyStopping(monitor='test_acc', patience=10, verbose=1), MaxEpochStopping(100),
ModelCheckpoint(create_save_model(self.model), '/tmp/dan.pt', monitor='test_acc')
# Tensorboard('dan', log_dir='tb-logs')
])
log.info('Starting training...')
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(
n_batches_train,
t_x_train, t_offset_train, t_y_train, evaluate=False
)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(
n_batches_test,
t_x_test, t_offset_test, t_y_test, evaluate=True
)
stop_training, reasons = manager.instruct(
train_time, train_loss, train_acc,
test_time, test_loss, test_acc
)
if stop_training:
log.info(' '.join(reasons))
break
log.info('Done training')
def train(self, training_data: TrainingData):
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data)
if self.use_wiki:
wiki_dataset = FilteredWikipediaDataset()
wiki_train_data = wiki_dataset.training_data()
w_x_train_text, w_train_y, _, _, _, _, _ = preprocess_dataset(
wiki_train_data,
train_size=1,
vocab=vocab,
class_to_i=class_to_i,
i_to_class=i_to_class)
x_train_text.extend(w_x_train_text)
y_train.extend(w_train_y)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
embeddings, embedding_lookup = load_embeddings(vocab=vocab,
expand_glove=True,
mask_zero=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
x_train = [
convert_text_to_embeddings_indices(q,
embedding_lookup,
random_unk_prob=.05)
for q in x_train_text
]
for r in x_train:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = [
convert_text_to_embeddings_indices(q,
embedding_lookup,
random_unk_prob=.05)
for q in x_test_text
]
for r in x_test:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_test = np.array(x_test)
y_test = np.array(y_test)
self.n_classes = compute_n_classes(training_data[1])
n_batches_train, t_x_train, lengths_train, t_y_train, _ = batchify(
self.batch_size, x_train, y_train, truncate=True)
n_batches_test, t_x_test, lengths_test, t_y_test, _ = batchify(
self.batch_size, x_test, y_test, truncate=False)
self.model = RnnModel(embeddings.shape[0], self.n_classes)
self.model.init_weights(embeddings=embeddings)
self.model.cuda()
self.optimizer = Adam(self.model.parameters(), lr=self.learning_rate)
self.criterion = nn.CrossEntropyLoss()
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer,
'max',
patience=5,
verbose=True)
manager = TrainingManager([
BaseLogger(log_func=log.info),
TerminateOnNaN(),
EarlyStopping(monitor='test_acc', patience=10, verbose=1),
MaxEpochStopping(100),
ModelCheckpoint(create_save_model(self.model),
'/tmp/rnn.pt',
monitor='test_acc')
#Tensorboard('rnn', log_dir='tb-logs')
])
log.info('Starting training...')
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(n_batches_train,
t_x_train,
lengths_train,
t_y_train,
evaluate=False)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(n_batches_test,
t_x_test,
lengths_test,
t_y_test,
evaluate=True)
stop_training, reasons = manager.instruct(train_time, train_loss,
train_acc, test_time,
test_loss, test_acc)
if stop_training:
log.info(' '.join(reasons))
break
else:
self.scheduler.step(test_acc)
log.info('Done training')
def train(self, training_data: TrainingData) -> None:
if self.use_qb:
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data
)
if self.use_wiki:
wiki_dataset = WikipediaDataset(set(training_data[1]))
wiki_train_data = wiki_dataset.training_data()
w_x_train_text, w_train_y, _, _, _, _, _ = preprocess_dataset(
wiki_train_data, train_size=1, vocab=vocab, class_to_i=class_to_i, i_to_class=i_to_class
)
x_train_text.extend(w_x_train_text)
y_train.extend(w_train_y)
else:
if self.use_wiki:
wiki_dataset = WikipediaDataset(set(training_data[1]))
wiki_train_data = wiki_dataset.training_data()
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
wiki_train_data
)
else:
raise ValueError('use_wiki and use_qb cannot both be false, otherwise there is no training data')
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
embeddings, embedding_lookup = load_embeddings(vocab=vocab, expand_glove=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
x_train = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_train_text]
for r in x_train:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_test_text]
for r in x_test:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_test = np.array(x_test)
y_test = np.array(y_test)
self.n_classes = compute_n_classes(training_data[1])
log.info(f'Batching: {len(x_train)} train questions and {len(x_test)} test questions')
n_batches_train, t_x_train, t_offset_train, t_y_train = batchify(
self.batch_size, x_train, y_train, truncate=True)
n_batches_test, t_x_test, t_offset_test, t_y_test = batchify(
self.batch_size, x_test, y_test, truncate=False)
self.vocab_size = embeddings.shape[0]
self.model = DanModel(self.vocab_size, self.n_classes, embeddings=embeddings)
if CUDA:
self.model = self.model.cuda()
log.info(f'Model:\n{self.model}')
self.optimizer = Adam(self.model.parameters(), lr=self.learning_rate)
self.criterion = nn.CrossEntropyLoss()
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optimizer, patience=5, verbose=True, mode='max')
manager = TrainingManager([
BaseLogger(log_func=log.info), TerminateOnNaN(),
EarlyStopping(monitor='test_acc', patience=10, verbose=1), MaxEpochStopping(100),
ModelCheckpoint(create_save_model(self.model), '/tmp/dan.pt', monitor='test_acc')
])
log.info('Starting training...')
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(
n_batches_train,
t_x_train, t_offset_train, t_y_train, evaluate=False
)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(
n_batches_test,
t_x_test, t_offset_test, t_y_test, evaluate=True
)
stop_training, reasons = manager.instruct(
train_time, train_loss, train_acc,
test_time, test_loss, test_acc
)
if stop_training:
log.info(' '.join(reasons))
break
else:
self.scheduler.step(test_acc)
log.info('Done training')
def train(self, training_data) -> None:
log.info('Preprocessing data')
x_train_text, y_train, x_test_text, y_test, vocab, class_to_i, i_to_class = preprocess_dataset(
training_data
)
self.class_to_i = class_to_i
self.i_to_class = i_to_class
self.vocab = vocab
embeddings, embedding_lookup = load_embeddings(vocab=vocab, expand_glove=True, mask_zero=True)
self.embeddings = embeddings
self.embedding_lookup = embedding_lookup
x_train = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_train_text]
for r in x_train:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test = [convert_text_to_embeddings_indices(q, embedding_lookup) for q in x_test_text]
for r in x_test:
if len(r) == 0:
r.append(embedding_lookup['UNK'])
x_test = np.array(x_test)
y_test = np.array(y_test)
log.info('Batching data')
n_batches_train, t_x_train, lengths_train, masks_train, t_y_train = batchify(
self.batch_size, x_train, y_train, truncate=True
)
n_batches_test, t_x_test, lengths_test, masks_test, t_y_test = batchify(
self.batch_size, x_test, y_test, truncate=False, shuffle=False
)
self.n_classes = compute_n_classes(training_data[1])
log.info('Creating model')
self.model = BCN(
300, 500, embeddings.shape[0], self.n_classes,
We=torch.from_numpy(embeddings)
).cuda()
self.optimizer = Adam(self.model.parameters())
self.criterion = nn.NLLLoss()
log.info(f'Model:\n{self.model}')
manager = TrainingManager([
BaseLogger(log_func=log.info), TerminateOnNaN(),
EarlyStopping(monitor='test_acc', patience=10, verbose=1), MaxEpochStopping(100),
ModelCheckpoint(create_save_model(self.model), '/tmp/bcn.pt', monitor='test_acc'),
Tensorboard('bcn')
])
log.info('Starting training...')
while True:
self.model.train()
train_acc, train_loss, train_time = self.run_epoch(
n_batches_train,
t_x_train, lengths_train, masks_train, t_y_train, evaluate=False
)
self.model.eval()
test_acc, test_loss, test_time = self.run_epoch(
n_batches_test,
t_x_test, lengths_test, masks_test, t_y_test, evaluate=True
)
stop_training, reasons = manager.instruct(
train_time, train_loss, train_acc,
test_time, test_loss, test_acc
)
if stop_training:
log.info(' '.join(reasons))
break
