def fit(self, X, y=None) -> None:
documents, features = X.shape
ds = CountTensorDataset(X.astype(np.float32))
self.autoencoder = ProdLDA(in_dimension=features,
hidden1_dimension=self.hidden1_dimension,
hidden2_dimension=self.hidden2_dimension,
topics=self.topics)
if self.cuda:
self.autoencoder.cuda()
ae_optimizer = Adam(self.autoencoder.parameters(),
lr=self.lr,
betas=(0.99, 0.999))
train(
ds,
self.autoencoder,
cuda=self.cuda,
validation=None,
epochs=self.epochs,
batch_size=self.batch_size,
optimizer=ae_optimizer,
sampler=WeightedRandomSampler(torch.ones(documents),
min(documents, self.samples)),
silent=True,
num_workers=0 # TODO causes a bug to change this on Mac
)
def test_loss_basic():
vae = ProdLDA(in_dimension=10,
hidden1_dimension=20,
hidden2_dimension=10,
topics=5)
for size in [10, 100, 1000]:
batch = torch.zeros(size, 10)
loss = vae.loss(batch, batch, vae.prior_mean, vae.prior_logvar)
assert loss.shape == (size, )
assert loss.mean().item() == 0
assert torch.all(torch.lt(torch.abs(loss), 0)).item() == 0
def test_parameters():
vae = ProdLDA(in_dimension=10,
hidden1_dimension=20,
hidden2_dimension=10,
topics=5)
# encoder
# two each for the linear units
assert len(tuple(vae.encoder.parameters())) == 4
assert (len(
tuple(param for param in vae.encoder.parameters()
if param.requires_grad)) == 4)
# mean and logvar
# two for the linear, two for the batchnorm
assert len(tuple(vae.mean.parameters())) == 4
assert (len(
tuple(param for param in vae.mean.parameters()
if param.requires_grad)) == 3)
assert len(tuple(vae.logvar.parameters())) == 4
assert (len(
tuple(param for param in vae.logvar.parameters()
if param.requires_grad)) == 3)
# decoder
# one for the linear, two for the batchnorm
assert len(tuple(vae.decoder.parameters())) == 3
# batchnorm has no scale
assert (len(
tuple(param for param in vae.decoder.parameters()
if param.requires_grad)) == 2)
def test_copy_embeddings_model():
lookup = {9: torch.ones(20), 8: torch.tensor(20).fill_(2)}
vae = ProdLDA(in_dimension=10,
hidden1_dimension=20,
hidden2_dimension=10,
topics=5,
word_embeddings=lookup)
assert vae.encoder.linear1.weight[:, 9].eq(1).all()
assert vae.encoder.linear1.weight[:, 8].eq(2).all()
def test_forward_dimensions():
vae = ProdLDA(in_dimension=10,
hidden1_dimension=20,
hidden2_dimension=10,
topics=5)
for size in [10, 100, 1000]:
batch = torch.zeros(size, 10)
recon, mean, logvar = vae(batch)
assert recon.shape == batch.shape
assert mean.shape == (size, 5)
assert logvar.shape == (size, 5)
def test_not_train_embeddings():
vae = ProdLDA(in_dimension=10,
hidden1_dimension=20,
hidden2_dimension=10,
topics=5,
train_word_embeddings=False)
for size in [10, 100, 1000]:
batch = torch.zeros(size, 10)
recon, mean, logvar = vae(batch)
assert recon.shape == batch.shape
assert mean.shape == (size, 5)
assert logvar.shape == (size, 5)
def main(cuda, batch_size, epochs, top_words, testing_mode, verbose_mode):
print("Loading input data")
# TODO fix relative paths
data_train = load_npz("data/train.txt.npz")
data_val = load_npz("data/test.txt.npz")
corpus = Sparse2Corpus(data_train, documents_columns=False)
with open("data/vocab.pkl", "rb") as f:
vocab = pickle.load(f)
reverse_vocab = {vocab[word]: word for word in vocab}
indexed_vocab = [
reverse_vocab[index] for index in range(len(reverse_vocab))
]
writer = SummaryWriter() # create the TensorBoard object
# callback function to call during training, uses writer from the scope
def training_callback(autoencoder, epoch, lr, loss, perplexity):
if verbose_mode:
decoder_weight = autoencoder.decoder.linear.weight.detach().cpu()
topics = [[
reverse_vocab[item.item()] for item in topic
] for topic in decoder_weight.topk(top_words, dim=0)[1].t()]
cm = CoherenceModel(
topics=topics,
corpus=corpus,
dictionary=Dictionary.from_corpus(corpus, reverse_vocab),
coherence="u_mass",
)
coherence = cm.get_coherence()
coherences = cm.get_coherence_per_topic()
for index, topic in enumerate(topics):
print(
str(index) + ":" + str(coherences[index]) + ":" +
",".join(topic))
print(coherence)
else:
coherence = 0
writer.add_scalars(
"data/autoencoder",
{
"lr": lr,
"loss": loss,
"perplexity": perplexity,
"coherence": coherence,
},
global_step=epoch,
)
ds_train = CountTensorDataset(data_train)
ds_val = CountTensorDataset(data_val)
autoencoder = ProdLDA(in_dimension=len(vocab),
hidden1_dimension=100,
hidden2_dimension=100,
topics=50)
if cuda:
autoencoder.cuda()
print("Training stage.")
ae_optimizer = Adam(autoencoder.parameters(), 0.0001, betas=(0.99, 0.999))
train(
ds_train,
autoencoder,
cuda=cuda,
validation=ds_val,
epochs=epochs,
batch_size=batch_size,
optimizer=ae_optimizer,
update_callback=training_callback,
sampler=WeightedRandomSampler(torch.ones(data_train.shape[0]), 20000),
num_workers=4,
)
autoencoder.eval()
decoder_weight = autoencoder.decoder.linear.weight.detach().cpu()
topics = [[reverse_vocab[item.item()] for item in topic]
for topic in decoder_weight.topk(top_words, dim=0)[1].t()]
cm = CoherenceModel(
topics=topics,
corpus=corpus,
dictionary=Dictionary.from_corpus(corpus, reverse_vocab),
coherence="u_mass",
)
coherence = cm.get_coherence()
coherences = cm.get_coherence_per_topic()
for index, topic in enumerate(topics):
print(
str(index) + ":" + str(coherences[index]) + ":" + ",".join(topic))
print(coherence)
if not testing_mode:
writer.add_embedding(
autoencoder.encoder.linear1.weight.detach().cpu().t(),
metadata=indexed_vocab,
tag="feature_embeddings",
)
writer.close()
def main(
cuda,
batch_size,
epochs,
top_words,
testing_mode,
):
print("Loading input data")
# TODO fix relative paths
data_train = np.load("data/train.txt.npy", encoding="bytes")
data_val = np.load("data/test.txt.npy", encoding="bytes")
with open("data/vocab.pkl", "rb") as f:
vocab = pickle.load(f)
reverse_vocab = {vocab[word]: word for word in vocab}
indexed_vocab = [
reverse_vocab[index] for index in range(len(reverse_vocab))
]
writer = SummaryWriter() # create the TensorBoard object
# callback function to call during training, uses writer from the scope
def training_callback(autoencoder, epoch, lr, loss, perplexity):
writer.add_scalars(
"data/autoencoder",
{
"lr": lr,
"loss": loss,
"perplexity": perplexity,
},
global_step=epoch,
)
decoder_weight = autoencoder.decoder.linear.weight.detach().cpu()
topics = [[reverse_vocab[item.item()] for item in topic]
for topic in decoder_weight.topk(top_words, dim=0)[1].t()]
for index, topic in enumerate(topics):
print(str(index) + ":" + ",".join(topic))
ds_train = TensorDataset(torch.from_numpy(data_train).float())
ds_val = TensorDataset(torch.from_numpy(data_val).float())
autoencoder = ProdLDA(in_dimension=len(vocab),
hidden1_dimension=100,
hidden2_dimension=100,
topics=50)
if cuda:
autoencoder.cuda()
print("Training stage.")
ae_optimizer = Adam(autoencoder.parameters(), 0.001, betas=(0.99, 0.999))
train(
ds_train,
autoencoder,
cuda=cuda,
validation=ds_val,
epochs=epochs,
batch_size=batch_size,
optimizer=ae_optimizer,
update_callback=training_callback,
)
autoencoder.eval()
decoder_weight = autoencoder.decoder.linear.weight.detach().cpu()
topics = [[reverse_vocab[item.item()] for item in topic]
for topic in decoder_weight.topk(top_words, dim=0)[1].t()]
for topic in topics:
print(",".join(topic))
if not testing_mode:
writer.add_embedding(
autoencoder.encoder.linear1.weight.detach().cpu().t(),
metadata=indexed_vocab,
tag="feature_embeddings",
)
writer.close()
def main(
cuda,
batch_size,
epochs,
top_words,
testing_mode,
):
print('Loading input data')
# TODO fix relative paths
input_train = np.load('data/train.txt.npy', encoding='bytes')
input_val = np.load('data/test.txt.npy', encoding='bytes')
with open('data/vocab.pkl', 'rb') as f:
vocab = pickle.load(f)
reverse_vocab = {vocab[word]: word for word in vocab}
indexed_vocab = [
reverse_vocab[index] for index in range(len(reverse_vocab))
]
data_train = np.array([
np.bincount(doc.astype('int'), minlength=len(vocab))
for doc in input_train if doc.sum() > 0
])
data_val = np.array([
np.bincount(doc.astype('int'), minlength=len(vocab))
for doc in input_val if doc.sum() > 0
])
writer = SummaryWriter() # create the TensorBoard object
# callback function to call during training, uses writer from the scope
def training_callback(autoencoder, epoch, lr, loss, perplexity):
writer.add_scalars('data/autoencoder', {
'lr': lr,
'loss': loss,
'perplexity': perplexity,
},
global_step=epoch)
decoder_weight = autoencoder.decoder.linear.weight.detach().cpu()
topics = [[reverse_vocab[item.item()] for item in topic]
for topic in decoder_weight.topk(top_words, dim=0)[1].t()]
for index, topic in enumerate(topics):
print(str(index) + ':' + ','.join(topic))
ds_train = TensorDataset(torch.from_numpy(data_train).float())
ds_val = TensorDataset(torch.from_numpy(data_val).float())
autoencoder = ProdLDA(in_dimension=len(vocab),
hidden1_dimension=100,
hidden2_dimension=100,
topics=50)
if cuda:
autoencoder.cuda()
print('Training stage.')
ae_optimizer = Adam(autoencoder.parameters(), 0.001, betas=(0.99, 0.999))
train(ds_train,
autoencoder,
cuda=cuda,
validation=ds_val,
epochs=epochs,
batch_size=batch_size,
optimizer=ae_optimizer,
update_callback=training_callback)
autoencoder.eval()
decoder_weight = autoencoder.decoder.linear.weight.detach().cpu()
topics = [[reverse_vocab[item.item()] for item in topic]
for topic in decoder_weight.topk(top_words, dim=0)[1].t()]
for topic in topics:
print(','.join(topic))
if not testing_mode:
writer.add_embedding(
autoencoder.encoder.linear1.weight.detach().cpu().t(),
metadata=indexed_vocab,
tag='feature_embeddings',
)
writer.close()
class ProdLDATransformer(TransformerMixin, BaseEstimator):
def __init__(self,
cuda=None,
batch_size=200,
epochs=80,
hidden1_dimension=100,
hidden2_dimension=100,
topics=50,
lr=0.001,
samples=20000,
score_num=7,
score_type='coherence') -> None:
self.cuda = torch.cuda.is_available() if cuda is None else cuda
self.batch_size = batch_size
self.epochs = epochs
self.hidden1_dimension = hidden1_dimension
self.hidden2_dimension = hidden2_dimension
self.topics = topics
self.lr = lr
self.samples = samples
self.autoencoder = None
self.score_type = score_type
self.score_num = score_num
if self.score_type not in ['coherence']:
raise ValueError('score_type must be "coherence"')
def fit(self, X, y=None) -> None:
documents, features = X.shape
ds = CountTensorDataset(X.astype(np.float32))
self.autoencoder = ProdLDA(in_dimension=features,
hidden1_dimension=self.hidden1_dimension,
hidden2_dimension=self.hidden2_dimension,
topics=self.topics)
if self.cuda:
self.autoencoder.cuda()
ae_optimizer = Adam(self.autoencoder.parameters(),
lr=self.lr,
betas=(0.99, 0.999))
train(
ds,
self.autoencoder,
cuda=self.cuda,
validation=None,
epochs=self.epochs,
batch_size=self.batch_size,
optimizer=ae_optimizer,
sampler=WeightedRandomSampler(torch.ones(documents),
min(documents, self.samples)),
silent=True,
num_workers=0 # TODO causes a bug to change this on Mac
)
def transform(self, X):
if self.autoencoder is None:
raise NotFittedError
self.autoencoder.eval()
ds = CountTensorDataset(X.astype(np.float32))
output = predict(
ds,
self.autoencoder,
encode=True,
silent=True,
batch_size=self.batch_size,
num_workers=0 # TODO causes a bug to change this on Mac
)
return output.numpy()
def score(self, X, y=None, sample_weight=None) -> float:
# TODO this needs further testing for correctness, WIP
if self.autoencoder is None:
raise NotFittedError
self.autoencoder.eval()
corpus = Sparse2Corpus(X, documents_columns=False)
decoder_weight = self.autoencoder.decoder.linear.weight.detach().cpu()
id2word = {index: str(index) for index in range(X.shape[1])}
topics = [[str(item.item()) for item in topic]
for topic in decoder_weight.topk(
min(self.score_num, X.shape[1]), dim=0)[1].t()]
cm = CoherenceModel(topics=topics,
corpus=corpus,
dictionary=Dictionary.from_corpus(corpus, id2word),
coherence='u_mass')
return cm.get_coherence()
