def _RunModel(self, input_1_val, input_2_val, dropout, custom_grad):
self._ClearCachedSession()
tf.reset_default_graph()
with self.session() as sess:
tf.random.set_seed(321)
input_1 = tf.placeholder(tf.float32)
input_2 = tf.placeholder(tf.float32)
revnet_params = self._SimpleRevNetParams('revnet', dropout,
custom_grad)
revnet_params.params_init = py_utils.WeightInit.Xavier(scale=1.0,
seed=0)
revnet = revnet_params.Instantiate()
h = revnet.FPropDefaultTheta(
py_utils.NestedMap(split1=input_1, split2=input_2))
dfx = tf.gradients(h.Flatten(), [input_1, input_2])
dfw = tf.gradients(
h.Flatten(),
revnet.theta.Flatten(),
unconnected_gradients=tf.UnconnectedGradients.ZERO)
tf.global_variables_initializer().run()
dfx_val, dfw_val, h_val = sess.run([dfx, dfw, h],
feed_dict={
input_1: input_1_val,
input_2: input_2_val,
})
return h_val, dfx_val, dfw_val
def setUp(self):
"""Generates test minibatches, the expected output and the layer config."""
super().setUp()
tf.reset_default_graph()
# Generate the common list of minibatches
self.input_batch_list = self._GenerateListOfMinibatches()
# Get the expected test results to compare with
self.expected_outputs = self._GetExpectedTestOutputs()
# Set up the basic cumulative statistics layer config
input_dim = self.input_batch_list[0].features.shape[2]
p = cumulative_statistics_layer.CumulativeStatisticsLayer.Params()
p.stats_type = 'PASS_THRU'
p.use_weighted_frames = False
p.input_dim = input_dim
p.features_name = 'features'
p.paddings_name = 'paddings'
p.frame_weight_ffn.activation = ['SIGMOID']
p.frame_weight_ffn.has_bias = [True]
p.frame_weight_ffn.hidden_layer_dims = [1]
p.frame_weight_ffn.input_dim = input_dim
self.params = p
def testRespectsInfeedBatchSize(self):
p = ToyInputGenerator.Params()
p.batch_size = 3
self._tmpdir, p.input_files = _CreateFakeTFRecordFiles()
p.dataset_type = tf.data.TFRecordDataset
ig = p.Instantiate()
batch = ig.GetPreprocessedInputBatch()
self.assertEqual(batch.audio.shape[0], p.batch_size)
self.assertEqual(p.batch_size, ig.InfeedBatchSize())
tf.reset_default_graph()
ig = p.Instantiate()
with mock.patch.object(ig, 'InfeedBatchSize',
return_value=42) as mock_method:
batch = ig.GetPreprocessedInputBatch()
self.assertEqual(batch.audio.shape[0], 42)
mock_method.assert_called()
def setUp(self):
"""Generates test examples and the attentive scoring configuration."""
super().setUp()
tf.reset_default_graph()
# Generate the example to test the attentive scoring
(self.test_data, self.enroll_data,
self.data_info) = self._GenerateExample()
# Set up the basic attentive scoring layer config
p = attentive_scoring_layer.AttentiveScoringLayer.Params()
p.num_keys = self.data_info['num_keys']
p.key_dim = self.data_info['key_dim']
p.value_dim = self.data_info['value_dim']
p.scale_factor = 1.0
p.use_trainable_scale_factor = False
p.apply_l2_norm_to_keys = False
p.apply_l2_norm_to_values = True
p.apply_global_l2_norm_to_concat_form = False
self.params = p
def setUp(self):
super(BaseExampleInputGeneratorTest, self).setUp()
tf.reset_default_graph()
def setUp(self): tf.reset_default_graph()
def setUp(self):
super().setUp()
tf.reset_default_graph()
