def _run_non_stream_model(self):
# below model expects that input_data are already initialized in tu.TestBase
# in setUp, by default input_data should have 3 dimensions
# size of each dimesnion is constant and is defiend by self.weights
mode = modes.Modes.TRAINING
input_tf = tf.keras.layers.Input(shape=(
None,
self.input_data.shape[2],
))
dense1 = tf.keras.layers.Dense(units=self.weights[0].shape[1],
use_bias=False)
depthwisecnn1 = depthwise_conv1d.DepthwiseConv1D(
memory_size=self.memory_size,
inference_batch_size=self.batch_size,
mode=mode)
output_dense1_tf = dense1(input_tf)
output_tf = depthwisecnn1(inputs=output_dense1_tf)
dense1.set_weights([self.weights[0]])
depthwisecnn1.set_weights([self.weights[1], self.weights[2]])
model_tf = tf.keras.models.Model(input_tf, output_tf)
output_np = model_tf.predict(self.input_data)
return output_np, model_tf
def test_streaming_inference_internal_state(self):
output_non_stream_np, _ = self._run_non_stream_model()
mode = modes.Modes.STREAM_INTERNAL_STATE_INFERENCE
input_tf = tf.keras.layers.Input(shape=(
1,
self.input_data.shape[2],
),
batch_size=self.batch_size)
dense1 = tf.keras.layers.Dense(units=self.weights[0].shape[1],
use_bias=False)
depthwisecnn1 = depthwise_conv1d.DepthwiseConv1D(
memory_size=self.memory_size,
inference_batch_size=self.batch_size,
mode=mode)
output_dense1_tf = dense1(input_tf)
output_tf = depthwisecnn1(inputs=output_dense1_tf)
model = tf.keras.Model(input_tf, output_tf)
input_states_np = np.zeros(
[self.batch_size, self.memory_size, self.weights[1].shape[-1]])
dense1.set_weights([self.weights[0]])
depthwisecnn1.set_weights(
[self.weights[1], self.weights[2], input_states_np])
# streaming emulation: loop over every element in time
for i in range(self.input_data.shape[1]):
input_batch_np = self.input_data[:, i, :]
input_batch_np = np.expand_dims(input_batch_np, 1)
output_np = model.predict(input_batch_np)
for b in range(self.input_data.shape[0]): # loop over batch
self.assertAllClose(output_np[b][0],
output_non_stream_np[b][i])
def build(self, input_shape):
super(Svdf, self).build(input_shape)
if self.mode == modes.Modes.TRAINING:
self.dropout1 = non_scaling_dropout.NonScalingDropout(
self.dropout, training=True)
else:
self.dropout1 = tf.keras.layers.Lambda(lambda x: x)
self.dense1 = tf.keras.layers.Dense(units=self.units1,
use_bias=self.use_bias1)
self.depth_cnn1 = depthwise_conv1d.DepthwiseConv1D(
memory_size=self.memory_size,
inference_batch_size=self.inference_batch_size,
use_bias=self.use_bias,
mode=self.mode,
pad=self.pad)
if self.units2 > 0:
self.dense2 = tf.keras.layers.Dense(units=self.units2,
use_bias=True)
else:
self.dense2 = tf.keras.layers.Lambda(lambda x: x)
if self.use_batch_norm:
self.batch_norm = tf.keras.layers.BatchNormalization(
scale=self.bn_scale)
else:
self.batch_norm = tf.keras.layers.Lambda(lambda x: x)
