def testCreateModel(self):
model = genomics_cnn.create_model(
batch_size=self.params['batch_size'],
len_seqs=self.params['seq_len'],
num_classes=self.params['num_classes'],
num_motifs=self.params['num_motifs'],
len_motifs=self.params['len_motifs'],
num_denses=self.params['num_denses'])
logits = model(self.rand_data)
loss = tf.reduce_mean(
tf.keras.losses.sparse_categorical_crossentropy(
y_true=self.rand_labels, y_pred=logits, from_logits=True))
self.assertGreater(loss, 0)
def testCreateDifferentModels(self, one_hot, use_mc_dropout, use_spec_norm,
use_gp_layer):
if use_spec_norm:
spec_norm_hparams = {
'spec_norm_bound': 6.0,
'spec_norm_iteration': 1
}
else:
spec_norm_hparams = None
if use_gp_layer:
gp_layer_hparams = {
'gp_input_dim': 128,
'gp_hidden_dim': 1024,
'gp_scale': 2.0,
'gp_bias': 0.0,
'gp_input_normalization': True,
'gp_cov_discount_factor': 0.999,
'gp_cov_ridge_penalty': 1e-3,
}
else:
gp_layer_hparams = None
model = genomics_cnn.create_model(
batch_size=self.params['batch_size'],
len_seqs=self.params['seq_len'],
num_classes=self.params['num_classes'],
num_motifs=self.params['num_motifs'],
len_motifs=self.params['len_motifs'],
num_denses=self.params['num_denses'],
one_hot=one_hot,
use_mc_dropout=use_mc_dropout,
spec_norm_hparams=spec_norm_hparams,
gp_layer_hparams=gp_layer_hparams)
logits = model(self.rand_data)
if isinstance(logits, (tuple, list)):
logits, covmat = logits
self.assertEqual(
covmat.shape,
(self.params['batch_size'], self.params['batch_size']))
self.assertEqual(
logits.shape,
(self.params['batch_size'], self.params['num_classes']))
def get(
model_name: str,
batch_size: int,
**hyperparameters) -> tf.keras.Model:
"""Gets a model builder by name.
Args:
model_name: Name of the model builder class.
batch_size: the training batch size.
**hyperparameters: dict of possible kwargs to be passed to the model
constructor.
Returns:
A model builder class with a method .build(split) which can be called to
get the tf.data.Dataset, which has elements that are a dict with keys
'features' and 'labels'.
Raises:
ValueError: If model_name is unrecognized.
"""
logging.info(
'Building model %s with additional kwargs:\n%s',
model_name,
json.dumps(hyperparameters, indent=2, sort_keys=True))
if model_name not in get_model_names():
raise ValueError('Unrecognized model type: {!r}'.format(model_name))
# load from single_model_uncertainty directory
if model_name == 'genomics_cnn':
return genomics_cnn.create_model(batch_size, **hyperparameters)
if model_name == 'wide_resnet':
return wide_resnet.create_model(batch_size, **hyperparameters)
# load from uncertainty_baselines directory
if model_name == 'criteo_mlp':
return criteo_mlp.create_model(batch_size, **hyperparameters)
if model_name == 'resnet20':
return resnet20.create_model(batch_size, **hyperparameters)
if model_name == 'resnet50':
return resnet50.create_model(batch_size, **hyperparameters)
if model_name == 'textcnn':
return textcnn.create_model(batch_size, **hyperparameters)
if model_name == 'bert':
return bert.create_model(batch_size, **hyperparameters)
def testCreateOptimizer(self, weight_decay):
model = genomics_cnn.create_model(
batch_size=self.params['batch_size'],
len_seqs=self.params['seq_len'],
num_classes=self.params['num_classes'],
num_motifs=self.params['num_motifs'],
len_motifs=self.params['len_motifs'],
num_denses=self.params['num_denses'])
optimizer = ub.optimizers.get(
optimizer_name=self.params['optimizer_name'],
learning_rate=0.001,
weight_decay=weight_decay,
model=model)
with tf.GradientTape() as tape:
logits = model(self.rand_data)
loss = tf.reduce_mean(
tf.keras.losses.sparse_categorical_crossentropy(
y_true=self.rand_labels, y_pred=logits, from_logits=True))
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(list(zip(grads, model.trainable_variables)))