def prepare_data(self):
test, train, val = utils.load_test_train_val(self.data_num) # df
train_texts = list(train.posts)
glove = Glove()
glove.create_custom_embedding([word for text in train_texts for word in text.split()])
self.train_tuple = utils.process_data(train, glove, self.max_words, self.max_posts)
self.test_tuple = utils.process_data(test, glove, self.max_words, self.max_posts)
self.val_tuple = utils.process_data(val, glove, self.max_words, self.max_posts)
def train_and_eval(thread_ids,
posts,
labels,
max_posts=20,
max_words=400,
frac=[0.8, 0.1, 0.1],
seed=0,
batch_size=9,
embedding='glove',
max_epoch=500,
validate=False,
result_dir=None):
# preliminary check
if len(thread_ids) != len(posts) or \
len(thread_ids) != len(labels) or \
len(posts) != len(labels):
raise Exception('Invalid length of data.')
if len(frac) != 3 or frac[0] + frac[1] + frac[2] != 1:
raise Exception('Invalid value of frac.')
if frac[0] <= 0 or frac[1] <= 0 or frac[2] <= 0:
raise Exception('Invalid value(s) for one or more frac element(s).')
if embedding not in ['glove']:
raise Exception('Invalid embedding.')
train_texts, train_labels, test_texts, test_labels, val_texts, val_labels = utils.filter_and_shuffle_data(
thread_ids, posts, labels, max_words, max_posts, seed, frac)
print(f'''----------
Data Split Result:
Train data = {len(train_texts)}
Test data = {len(test_texts)}
Val data = {len(val_texts)}
----------''')
# from here on is glove specific implementation (may need to extract to a function)
print('Init embedding')
glove = Glove()
glove.create_custom_embedding(
[item for sublist in train_texts for item in sublist])
glove.add_to_embedding(['.', '!', '?'])
print('Padding and packing data into data loader')
for i, thread in enumerate(train_texts):
for j, post_text in enumerate(thread):
train_texts[i][j] = glove.sentence_to_indices(post_text,
seq_len=max_words)
for i, thread in enumerate(test_texts):
for j, post_text in enumerate(thread):
test_texts[i][j] = glove.sentence_to_indices(post_text,
seq_len=max_words)
for i, thread in enumerate(val_texts):
for j, post_text in enumerate(thread):
val_texts[i][j] = glove.sentence_to_indices(post_text,
seq_len=max_words)
# padding at the post level
post_padding = [glove.word2idx[glove.pad_token]] * max_words
for posts in [train_texts, test_texts, val_texts]:
for sublist in posts:
if len(sublist) < max_posts:
sublist.extend([post_padding] * (max_posts - len(sublist)))
for labels in [train_labels, test_labels, val_labels]:
for sublist in labels:
if len(sublist) < max_posts:
sublist.extend([0] * (max_posts - len(sublist)))
train_loader = utils.to_data_loader(batch_size, train_texts, train_labels)
test_loader = utils.to_data_loader(batch_size, test_texts, test_labels)
val_loader = utils.to_data_loader(batch_size, val_texts, val_labels)
print('Creating model')
embedding = create_emb_layer(
torch.from_numpy(glove.weights_matrix).float().to(utils.get_device()))
model = hLSTM(input_size=glove.emb_dim,
hidden_size=glove.emb_dim,
output_size=glove.emb_dim,
batch_size=batch_size,
num_layers=1,
bidirectional=False,
embedding=embedding,
drop_prob=0.5,
max_output=max_posts,
device=utils.get_device())
labels = [label for sublist in train_labels for label in sublist]
intervention_ratio = len([label
for label in labels if label == 1]) / len(labels)
loss_fn = WeightedBCELoss(zero_weight=intervention_ratio,
one_weight=1 - intervention_ratio)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
writer = None
if result_dir is not None:
writer = SummaryWriter(f'runs/{result_dir}')
if not os.path.exists(f'models/{result_dir}'):
os.makedirs(f'models/{result_dir}')
if not validate:
val_loader = None
print('Start training model')
train_model(model, train_loader, max_epoch, loss_fn, optimizer, val_loader,
writer)
print('Evaluating model')
f1, precision, recall = eval_model(model, test_loader, False)
print(f'''
Test results:
F1 = {f1}
Precision = {precision}
Recall = {recall}
''')
if result_dir is not None:
print('Saving final model')
torch.save(model.state_dict(), f'models/{result_dir}/final_model.pth')
print('DONE :)))')
def train_and_eval_crf(thread_ids, posts, labels, max_posts=20,
max_words=400, frac=[0.8, 0.1, 0.1], seed=0,
batch_size=9, embedding='glove', max_epoch=500,
validate=False, result_dir=None):
# preliminary check
if len(thread_ids) != len(posts) or \
len(thread_ids) != len(labels) or \
len(posts) != len(labels):
raise Exception('Invalid length of data.')
if len(frac) != 3 or frac[0]+frac[1]+frac[2] != 1:
raise Exception('Invalid value of frac.')
if frac[0] <= 0 or frac[1] <= 0 or frac[2] <= 0:
raise Exception('Invalid value(s) for one or more frac element(s).')
if embedding not in ['glove']:
raise Exception('Invalid embedding.')
train_texts, train_labels, test_texts, test_labels, val_texts, val_labels = utils.filter_and_shuffle_data(thread_ids, posts, labels, max_words, max_posts, seed, frac)
# from here on is glove specific implementation (may need to extract to a function)
print('Init embedding')
glove = Glove()
glove.create_custom_embedding([item for sublist in train_texts for item in sublist])
glove.add_to_embedding(['.', '!', '?'])
print('Padding and packing data into data loader')
for i, thread in enumerate(train_texts):
for j, post_text in enumerate(thread):
train_texts[i][j] = glove.sentence_to_indices(post_text, seq_len=max_words)
for i, thread in enumerate(test_texts):
for j, post_text in enumerate(thread):
test_texts[i][j] = glove.sentence_to_indices(post_text, seq_len=max_words)
for i, thread in enumerate(val_texts):
for j, post_text in enumerate(thread):
val_texts[i][j] = glove.sentence_to_indices(post_text, seq_len=max_words)
# padding at the post level
post_padding = [glove.word2idx[glove.pad_token]] * max_words
for posts in [train_texts, test_texts, val_texts]:
for sublist in posts:
if len(sublist) < max_posts:
sublist.extend([post_padding] * (max_posts - len(sublist)))
train_masks, test_masks, val_masks = [], [], []
def get_to_append(ones):
to_append = [1] * len(labels)
if len(to_append) < max_posts:
to_append.extend([0] * (max_posts-len(labels)))
return to_append
for labels in train_labels:
to_append = get_to_append(len(labels))
train_masks.append(to_append)
for labels in test_labels:
to_append = get_to_append(len(labels))
test_masks.append(to_append)
for labels in val_labels:
to_append = get_to_append(len(labels))
val_masks.append(to_append)
for labels in [train_labels, test_labels, val_labels]:
for sublist in labels:
if len(sublist) < max_posts:
sublist.extend([0] * (max_posts-len(sublist)))
train_loader = utils.to_data_loader(batch_size, train_texts, train_labels, train_masks)
test_loader = utils.to_data_loader(batch_size, test_texts, test_labels, test_masks)
val_loader = utils.to_data_loader(batch_size, val_texts, val_labels, val_masks)
print('Creating model')
embedding = create_emb_layer(torch.from_numpy(glove.weights_matrix).float().to(utils.get_device()))
model = hLSTM_CRF(num_tags=2,
input_size=glove.emb_dim,
hidden_size=glove.emb_dim,
output_size=glove.emb_dim,
batch_size=batch_size,
num_layers=1,
bidirectional=False,
embedding=embedding,
drop_prob=0.5,
max_output=max_posts,
device=utils.get_device())
labels = [label for sublist in train_labels for label in sublist]
intervention_ratio = len([label for label in labels if label == 1]) / len(labels)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
writer = None
if result_dir is not None:
writer = SummaryWriter(f'runs/{result_dir}')
if not os.path.exists(f'models/{result_dir}'):
os.makedirs(f'models/{result_dir}')
if not validate:
val_loader = None
print('Start training model')
model.zero_grad()
model.train()
running_loss = 0.0
for epoch in range(max_epoch):
if (epoch + 1) % 20 == 0:
print(f'Training model ({epoch + 1} / {max_epoch})')
for i, (inputs, labels, masks) in enumerate(train_loader):
inputs, labels, masks = inputs.to(utils.get_device()), labels.to(utils.get_device()), masks.to(utils.get_device())
loss = model.loss(inputs, labels, masks)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5)
optimizer.step()
running_loss += loss.item()
if i % 1000 == 999: # every 100 mini-batches
if writer is not None:
writer.add_scalar('training loss',
running_loss / 1000,
epoch * len(train_loader) + i)
running_loss = 0.0
if val_loader is not None:
f1, _, _ = eval_model(model, val_loader)
writer.add_scalar('validation f1', f1,
epoch * len(train_loader) + i)
print('Evaluating model')
f1, precision, recall = eval_model(model, test_loader, False)
print(f'''
Test results:
F1 = {f1}
Precision = {precision}
Recall = {recall}
''')
if result_dir is not None:
print('Saving final model')
torch.save(model.state_dict(), f'models/{result_dir}/final_model.pth')
print('DONE :)))')
INTERVENED_RATIO = 0.25
EPOCHS = 1
CLIP = 5
VAL_EVERY = 200
TB_FOLDER = 'hlstm_data1_ep1'
test, train, val = utils.load_test_train_val(DATA_NUM) # df
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
writer_v = SummaryWriter(f'runs/{TB_FOLDER}_val')
writer_t = SummaryWriter(f'runs/{TB_FOLDER}_train')
train_texts = list(train.posts)
print('Init GloVe embedding')
glove = Glove()
glove.create_custom_embedding([word for text in train_texts for word in text.split()])
print(len(glove.word2idx))
print('Padding and packing data into data loader')
train_indices, train_labels, train_wl, train_pl = utils.process_data(train, glove, MAX_WORDS, MAX_POSTS)
test_indices, test_labels, test_wl, test_pl = utils.process_data(test, glove, MAX_WORDS, MAX_POSTS)
val_indices, val_labels, val_wl, val_pl = utils.process_data(val, glove, MAX_WORDS, MAX_POSTS)
train_loader = utils.to_data_loader(train_indices, train_labels, train_wl, train_pl, BATCH_SIZE)
test_loader = utils.to_data_loader(test_indices, test_labels, test_wl, test_pl, BATCH_SIZE)
val_loader = utils.to_data_loader(val_indices, val_labels, val_wl, val_pl, BATCH_SIZE)
print('Creating model')
embedding = create_emb_layer(torch.from_numpy(glove.weights_matrix).float().to(device))
criterion = WeightedBCELoss(zero_weight=INTERVENED_RATIO, one_weight=1-INTERVENED_RATIO)
