def test_receptive_field_size(self, kernel3):
batch_size = 2
num_classes = 10
images = _construct_images(batch_size)
config = bagnet_config.get_config()
config.num_classes = num_classes
config.kernel3 = kernel3
config.batch_norm.enable = False # Batch norm false to use random model.
config.strides = [2, 2, 2, 1]
config.blocks = [3, 4, 6, 3]
model = bagnet_model.BagNet(config)
_, endpoints = model(images, is_training=True)
logits2d = endpoints["logits2d"]
init = tf.global_variables_initializer()
self.evaluate(init)
# Compute gradient with respect to input of one of the outputs.
loss = tf.losses.sparse_softmax_cross_entropy(
labels=tf.random.uniform(shape=(batch_size, ),
minval=0,
maxval=num_classes,
dtype=tf.int32),
logits=logits2d[:, 0, 0, :])
grad = self.evaluate(tf.gradients(loss, images)[0])
# Infer the receptive field as the number of non-zero gradients.
receptive_field_size = 0
for j in range(batch_size):
for i in range(images.shape.as_list()[-1]):
is_nonzero_grad = np.amax(grad[j, :, :, i], 1) > 1e-12
receptive_field_size = max(np.sum(is_nonzero_grad),
receptive_field_size)
# Ensure receptive field is correct.
self.assertEqual(model.receptive_field[0], receptive_field_size)
def test_reuse_model(self, avg_pool, is_training):
batch_size = 2
num_classes = 10
images = _construct_images(batch_size)
config = bagnet_config.get_config()
config.num_classes = num_classes
config.avg_pool = avg_pool
model = bagnet_model.BagNet(config)
# Build once.
logits1, _ = model(images, is_training)
num_params = len(tf.all_variables())
l2_loss1 = tf.losses.get_regularization_loss()
# Build twice.
logits2, _ = model(
images, is_training)
l2_loss2 = tf.losses.get_regularization_loss()
# Ensure variables are reused.
self.assertLen(tf.all_variables(), num_params)
init = tf.global_variables_initializer()
self.evaluate(init)
logits1, logits2 = self.evaluate((logits1, logits2))
l2_loss1, l2_loss2 = self.evaluate((l2_loss1, l2_loss2))
np.testing.assert_almost_equal(logits1, logits2, decimal=9)
np.testing.assert_almost_equal(l2_loss1, l2_loss2, decimal=9)
def test_build_model(self, avg_pool, is_training):
batch_size = 2
num_classes = 10
images = _construct_images(batch_size)
config = bagnet_config.get_config()
config.num_classes = num_classes
config.avg_pool = avg_pool
model = bagnet_model.BagNet(config)
logits, _ = model(images, is_training)
final_shape = (batch_size, num_classes)
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
self.assertEqual(final_shape, sess.run(logits).shape)
def __init__(self, config, variable_scope="saccader"):
self.config = copy.deepcopy(config)
representation_config = self.config.representation_config
representation_config.num_classes = self.config.num_classes
self.glimpse_shape = (-1, -1)
self.variable_scope = variable_scope
self.saccader_cell = SaccaderCell(soft_attention=self.config.soft_attention)
self.representation_network = bagnet_model.BagNet(representation_config)
self.var_list_saccader_cell = []
self.var_list_representation_network = []
self.var_list_attention_network = []
self.var_list_classification = []
self.var_list_location = []
self.var_list = []
self.init_op = None