def _get_vae_estimator(self, config, reporter, ctx):
"""Take a model configuration - specified by a config file or as determined by model selection and
return a VAE topic model ready for training.
Parameters:
config (dict): an autogluon configuration/argument object, instantiated to particular parameters
reporter (`autogluon.core.scheduler.reporter.Reporter`): object for reporting model evaluations to scheduler
ctx (`mxnet.context.Context`): Mxnet compute context
Returns:
Estimator (:class:`tmnt.estimator.BaseEstimator`): Either BowEstimator or MetaBowEstimator
"""
lr = config.lr
latent_distrib = config.latent_distribution
optimizer = config.optimizer
n_latent = int(config.n_latent)
enc_hidden_dim = int(config.enc_hidden_dim)
coherence_reg_penalty = float(config.coherence_loss_wt)
redundancy_reg_penalty = float(config.redundancy_loss_wt)
batch_size = int(config.batch_size)
embedding_source = config.embedding.source
fixed_embedding = config.embedding.get('fixed') == True
covar_net_layers = config.covar_net_layers
n_encoding_layers = config.num_enc_layers
enc_dr = config.enc_dr
epochs = int(config.epochs)
ldist_def = config.latent_distribution
kappa = 0.0
alpha = 1.0
latent_distrib = ldist_def.dist_type
if latent_distrib == 'vmf':
kappa = ldist_def.kappa
elif latent_distrib == 'logistic_gaussian':
alpha = ldist_def.alpha
vocab, emb_size = self._initialize_vocabulary(embedding_source)
if emb_size < 0 and 'size' in config.embedding:
emb_size = config.embedding.size
if self.c_args.use_labels_as_covars:
#n_covars = len(self.label_map) if self.label_map else 1
n_covars = -1
model = \
MetaBowEstimator(vocab, coherence_coefficient=8.0, reporter=reporter, num_val_words=self.total_tst_words,
wd_freqs=self.wd_freqs,
label_map=self.label_map,
covar_net_layers=1, ctx=ctx, lr=lr, latent_distribution=latent_distrib, optimizer=optimizer,
n_latent=n_latent, kappa=kappa, alpha=alpha, enc_hidden_dim=enc_hidden_dim,
coherence_reg_penalty=coherence_reg_penalty,
redundancy_reg_penalty=redundancy_reg_penalty, batch_size=batch_size,
embedding_source=embedding_source, embedding_size=emb_size, fixed_embedding=fixed_embedding,
num_enc_layers=n_encoding_layers, enc_dr=enc_dr, seed_matrix=self.seed_matrix, hybridize=False,
epochs=epochs, log_method='log')
else:
print("Encoder coherence = {}".format(
self.c_args.encoder_coherence))
model = \
BowEstimator(vocab, coherence_coefficient=8.0, reporter=reporter, num_val_words=self.total_tst_words,
wd_freqs=self.wd_freqs,
ctx=ctx, lr=lr, latent_distribution=latent_distrib, optimizer=optimizer,
n_latent=n_latent, kappa=kappa, alpha=alpha, enc_hidden_dim=enc_hidden_dim,
coherence_reg_penalty=coherence_reg_penalty,
redundancy_reg_penalty=redundancy_reg_penalty, batch_size=batch_size,
embedding_source=embedding_source, embedding_size=emb_size, fixed_embedding=fixed_embedding,
num_enc_layers=n_encoding_layers, enc_dr=enc_dr, seed_matrix=self.seed_matrix, hybridize=False,
epochs=epochs, log_method='log', coherence_via_encoder=self.c_args.encoder_coherence,
pretrained_param_file = self.pretrained_param_file)
model.validate_each_epoch = self.validate_each_epoch
return model
def test_train_and_topics_categorical():
model = MetaBowEstimator(vocabulary, batch_size=32)
model.fit(X_scipy, y_numpy)
model.get_topic_vectors()
assert (True)
def test_train_and_npmi_categorical():
model = MetaBowEstimator(vocabulary, batch_size=32)
model.fit(X_scipy, y_numpy)
assert (model._npmi_per_covariate(X_scipy, y_numpy, 10) == 0)
def test_train_and_transform_categorical():
model = MetaBowEstimator(vocabulary, batch_size=32)
model.fit(X_scipy, y_numpy)
trans = model.transform(X_scipy, y_numpy)
assert (np.all(trans == trans[0]))
def test_train_and_perplexity_categorical():
model = MetaBowEstimator(vocabulary, batch_size=32)
model.fit(X_scipy, y_numpy)
model.perplexity(X_scipy, y_numpy)
assert (True)