def testStreaming(self, input_samples):
# prepare non streaming model
stft_layer = stft.STFT(self.frame_size,
self.frame_step,
mode=modes.Modes.TRAINING,
inference_batch_size=1,
padding='causal')
input_tf = tf.keras.layers.Input(shape=(self.input_signal.shape[1], ),
batch_size=1)
net = stft_layer(input_tf)
model_non_stream = tf.keras.models.Model(input_tf, net)
params = test_utils.Params([1])
# shape of input data in the inference streaming mode (excluding batch size)
params.data_shape = (input_samples * stft_layer.frame_step, )
params.step = input_samples
# convert it to streaming model
model_stream = utils.to_streaming_inference(
model_non_stream, params,
modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run streaming inference and compare it with default stft
stream_out = inference.run_stream_inference(params, model_stream,
self.input_signal)
stream_output_length = stream_out.shape[1]
self.assertAllClose(stream_out, self.stft_out[:,
0:stream_output_length])
def test_delay_internal_state(self, delay_also_in_non_streaming):
"""Test delay layer with internal state."""
# model and data parameters
params = test_utils.Params([1], clip_duration_ms=1)
# prepare non stream model
time_delay = 3
model = delay_model(params, time_delay, delay_also_in_non_streaming)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model.summary()
# fill the buffer
for i in range(time_delay):
output = model_stream.predict([i + 1])
self.assertAllEqual(output[0, 0, 0], 0)
# now get the data with delay
for i in range(time_delay):
output = model_stream.predict([0])
self.assertAllEqual(output[0, 0, 0], i + 1)
def test_dynamic_shape(self):
# model and data parameters
params = test_utils.Params([1], clip_duration_ms=0.25)
# prepare input data
x = np.arange(10)
inp_audio = x
inp_audio = np.expand_dims(inp_audio, 0) # add batch dim
# prepare non stream model
params.desired_samples = None
model = conv1d_transpose_model(params,
filters=1,
kernel_size=3,
stride=1)
model.summary()
# run inference on input with dynamic shape
model.predict(inp_audio)
with self.assertRaisesRegex(
ValueError, 'in streaming mode time dimension of input packet '
'should not be dynamic: TFLite limitation'):
# streaming model expected to fail on input data with dynamic shape
params.data_shape = (None, )
utils.to_streaming_inference(model, params,
Modes.STREAM_INTERNAL_STATE_INFERENCE)
def test_streaming_strides(self, stride):
"""Test Conv1DTranspose layer in streaming mode with different strides.
Args:
stride: controls the upscaling factor
"""
# model and data parameters
step = 1 # amount of data fed into streaming model on every iteration
params = test_utils.Params([step], clip_duration_ms=0.25)
# prepare input data
x = np.arange(params.desired_samples)
inp_audio = x
inp_audio = np.expand_dims(inp_audio, 0) # add batch dim
# prepare non stream model
model = conv1d_transpose_model(params,
filters=1,
kernel_size=3,
stride=stride)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
stream_out = test.run_stream_inference(params, model_stream, inp_audio)
self.assertAllClose(stream_out, non_stream_out)
def test_dynamic_shape(self):
# model and data parameters
params = test_utils.Params([1], clip_duration_ms=0.25)
# prepare input data
x = np.random.rand(1, params.desired_samples, 1, self.input_channels)
inp_audio = x
# prepare non stream model
params.desired_samples = None
model = conv2d_transpose_model(
params,
filters=1,
kernel_size=(3, 1),
strides=(1, 1),
channels=self.input_channels)
model.summary()
# run inference on input with dynamic shape
model.predict(inp_audio)
with self.assertRaisesRegex(
ValueError, 'in streaming mode time dimension of input packet '
'should not be dynamic: TFLite limitation'):
# streaming model expected to fail on input data with dynamic shape
params.data_shape = (None, 1, self.input_channels)
utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
def testStreaming(self, input_frames):
params = test_utils.Params([1])
# shape of input data in the inference streaming mode (excluding batch size)
params.data_shape = (1, self.feature_size)
params.step = input_frames
# prepare non streaming model
inverse_stft_layer = inverse_stft.InverseSTFT(
self.frame_size, self.frame_step, use_one_step=(input_frames == 1))
input_tf = tf.keras.layers.Input(shape=self.signal_stft.shape[1:3],
batch_size=1,
dtype=tf.complex64)
net = inverse_stft_layer(input_tf)
model_non_stream = tf.keras.models.Model(input_tf, net)
self.non_stream_out = model_non_stream.predict(self.signal_stft)
# convert it to streaming model
model_stream = utils.to_streaming_inference(
model_non_stream, params,
modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run streaming inference
stream_out = inference.run_stream_inference(params, model_stream,
self.signal_stft)
# several samples in the end will be missing
stream_output_length = stream_out.shape[1]
self.assertAllClose(stream_out,
self.non_stream_out[:, 0:stream_output_length])
def test_transposed_conv(self):
"""Test transposed and standard conv model with 'same' padding."""
test_utils.set_seed(123)
# model and data parameters
cnn_filters = [1, 1]
cnn_kernel_size = [5, 3]
cnn_act = ['linear', 'linear']
cnn_use_bias = [False, False]
cnn_paddings = ['same', 'same']
trans_paddings = ['same', 'causal']
params = test_utils.Params([1], clip_duration_ms=2)
# prepare input data
x = np.arange(params.desired_samples)
inp_audio = x
inp_audio = np.expand_dims(inp_audio, 0)
# prepare non stream model
model = transposed_conv_model(params, cnn_filters, cnn_kernel_size,
cnn_act, cnn_use_bias, cnn_paddings,
trans_paddings)
# set random weights
all_weights = []
for w in model.get_weights():
if isinstance(w, np.ndarray):
shape = w.shape
new_w = np.random.rand(*shape)
all_weights.append(new_w)
else:
all_weights.append(True)
model.set_weights(all_weights)
model.summary()
non_stream_out = model.predict(inp_audio)
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
stream_out = inference.run_stream_inference(params, model_stream,
inp_audio)
# shift defines the index after which data in streaming mode become valid:
# in streaming mode we use ring buffers initialized with zeros and it needs
# several cycles until they are filled with real data.
shift = 2
# the total conv delay is (5//2) * 2 + 3//2 = 5
# (there is no delay from the k=3 s=2 transposed convs, 'same' or 'causal'),
# and the explicit Delay layers add an additional same amount.
total_delay = 10
# normalize output data and compare them
non_stream_out = non_stream_out[0, shift:-(total_delay), ]
stream_out = stream_out[0, total_delay + shift:, ]
self.assertAllClose(stream_out, non_stream_out)
def test_streaming_on_2d_data_strides(self, stride):
"""Tests Conv2DTranspose on 2d in streaming mode with different strides.
Args:
stride: controls the upscaling factor
"""
tf1.reset_default_graph()
config = tf1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf1.Session(config=config)
tf1.keras.backend.set_session(sess)
# model and data parameters
step = 1 # amount of data fed into streaming model on every iteration
params = test_utils.Params([step], clip_duration_ms=0.25)
input_features = 3
# prepare input data: [batch, time, features, channels]
x = np.random.rand(1, params.desired_samples, input_features,
self.input_channels)
inp_audio = x
# prepare non-streaming model
model = conv2d_transpose_model(
params,
filters=1,
kernel_size=(3, 3),
strides=(stride, stride),
features=input_features,
channels=self.input_channels)
model.summary()
# set weights with bias
for layer in model.layers:
if isinstance(layer, tf.keras.layers.Conv2DTranspose):
layer.set_weights([
np.ones(layer.weights[0].shape),
np.zeros(layer.weights[1].shape) + 0.5
])
params.data_shape = (1, input_features, self.input_channels)
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
stream_out = inference.run_stream_inference(params, model_stream, inp_audio)
self.assertAllClose(stream_out, non_stream_out)
def test_residual(self, step, padding, delay_also_in_non_streaming):
"""Test residual connection in streaming mode with conv layer."""
# model and data parameters
cnn_filters = [1, 1]
cnn_kernel_size = [5, 3]
cnn_act = ['elu', 'elu']
cnn_use_bias = [False, False]
cnn_padding = [padding, padding]
params = test_utils.Params([step], clip_duration_ms=2)
# prepare input data
x = np.arange(params.desired_samples)
inp_audio = x
inp_audio = np.expand_dims(inp_audio, 0)
# prepare non stream model
model, sum_delay = residual_model(params, cnn_filters, cnn_kernel_size,
cnn_act, cnn_use_bias, cnn_padding,
delay_also_in_non_streaming)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
stream_out = inference.run_stream_inference(params, model_stream,
inp_audio)
# normalize output data and compare them
channel = 0
non_stream_out = non_stream_out[0, :, channel]
stream_out = stream_out[0, :, channel]
min_len = min(stream_out.shape[0], non_stream_out.shape[0])
stream_out = stream_out[0:min_len]
non_stream_out = non_stream_out[0:min_len]
shift = 1
if delay_also_in_non_streaming:
# Delay was also applied in non-streaming, as well as streaming mode.
non_stream_out = non_stream_out[shift + sum_delay:min_len]
else:
non_stream_out = non_stream_out[shift:min_len - sum_delay]
stream_out = stream_out[sum_delay + shift:]
self.assertAllEqual(non_stream_out.shape, (31 - sum_delay, ))
self.assertAllClose(stream_out, non_stream_out)
def test_average_pooling_stream(self):
# prepare input data
params = test_utils.Params([1])
params.desired_samples = 5
batch_size = 1
time1 = params.desired_samples # it is time dim (will not be averaged out)
time2 = 3 # this dim will be averaged out and become 1
feature = 16 # it is a feature dim
# override data shape for streaming mode testing
params.preprocess = 'custom'
params.data_shape = (1, time2, feature)
inp_audio = np.random.rand(batch_size, time1, time2, feature)
inputs = tf.keras.layers.Input(
shape=(time1, time2, feature), batch_size=batch_size)
net = stream.Stream(
cell=average_pooling2d.AveragePooling2D(
kernel_size=(time1, time2),
padding='valid'),
use_one_step=False,
pad_time_dim='causal')(inputs)
model = tf.keras.Model(inputs, net)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference and compare streaming vs non streaming
non_stream_out = model.predict(inp_audio)
stream_out = test.run_stream_inference(params, model_stream, inp_audio)
self.assertAllClose(stream_out, non_stream_out)
net = tf.keras.layers.GlobalAveragePooling2D()(inputs)
model_global = tf.keras.Model(inputs, net)
model_global.summary()
global_out = model_global.predict(inp_audio)
# last result in streaming output has to be the same with global average
self.assertAllClose(stream_out[0, -1, 0, :], global_out[0, :])
def test_stream_strided_convolution(self, get_model, conv_cell):
# Test streaming convolutional layers with striding, dilation.
cnn_filters = [1, 1, 1, 1]
cnn_kernel_size = [3, 3, 3, 3]
cnn_act = ['linear', 'linear', 'elu', 'elu']
cnn_dilation_rate = [1, 1, 1, 2]
cnn_strides = [2, 1, 3, 1]
cnn_use_bias = [False, False, False, False]
# prepare input data
params = test_utils.Params(cnn_strides)
x = np.arange(params.desired_samples)
frequency = 2.0
inp_audio = np.cos((2.0 * np.pi / params.desired_samples) * frequency *
x) + np.random.rand(1, params.desired_samples) * 0.5
# prepare non stream model
model = get_model(params, conv_cell, cnn_filters, cnn_kernel_size,
cnn_act, cnn_dilation_rate, cnn_strides,
cnn_use_bias)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
stream_out = test.run_stream_inference(params, model_stream, inp_audio)
# normalize output data and compare them
channel = 0
non_stream_out = non_stream_out[0, :, channel]
stream_out = stream_out[0, :, channel]
min_len = min(stream_out.shape[0], non_stream_out.shape[0])
stream_out = stream_out[0:min_len]
non_stream_out = non_stream_out[0:min_len]
self.assertAllEqual(non_stream_out.shape, (42, ))
self.assertAllClose(stream_out, non_stream_out)
def test_residual(self, step):
# model and data parameters
cnn_filters = [1, 1, 1, 1]
cnn_kernel_size = [3, 3, 3, 3]
cnn_act = ['linear', 'linear', 'elu', 'elu']
cnn_use_bias = [False, False, False, False]
cnn_padding = ['causal', 'causal', 'causal', 'causal']
params = test_utils.Params([step], clip_duration_ms=2)
# prepare input data
x = np.arange(params.desired_samples)
frequency = 2.0
inp_audio = np.cos((2.0 * np.pi / params.desired_samples) * frequency *
x) + np.random.rand(1, params.desired_samples) * 0.5
# prepare non stream model
model = residual_model(params, cnn_filters, cnn_kernel_size, cnn_act,
cnn_use_bias, cnn_padding)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
stream_out = test.run_stream_inference(params, model_stream, inp_audio)
# normalize output data and compare them
channel = 0
non_stream_out = non_stream_out[0, :, channel]
stream_out = stream_out[0, :, channel]
min_len = min(stream_out.shape[0], non_stream_out.shape[0])
stream_out = stream_out[0:min_len]
non_stream_out = non_stream_out[0:min_len]
self.assertAllEqual(non_stream_out.shape, (32, ))
self.assertAllClose(stream_out, non_stream_out)
def test_conv(self):
"""Test conv model with 'same' padding."""
# model and data parameters
cnn_filters = [1, 1, 1]
cnn_kernel_size = [5, 3, 5]
cnn_act = ['elu', 'elu', 'elu']
cnn_use_bias = [False, False, False]
cnn_padding = ['same', 'causal', 'same']
params = test_utils.Params([1], clip_duration_ms=2)
# prepare input data
x = np.arange(params.desired_samples)
inp_audio = x
inp_audio = np.expand_dims(inp_audio, 0)
# prepare non stream model
model, sum_delay, sum_shift = conv_model(params, cnn_filters,
cnn_kernel_size, cnn_act,
cnn_use_bias, cnn_padding)
model.summary()
non_stream_out = model.predict(inp_audio)
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
stream_out = inference.run_stream_inference(params, model_stream,
inp_audio)
shift = sum_shift + 1
# normalize output data and compare them
non_stream_out = non_stream_out[0, shift:-(sum_delay), ]
stream_out = stream_out[0, sum_delay + shift:, ]
self.assertAllClose(stream_out, non_stream_out)
def test_stream_strided_convolution(self, get_model, conv_cell):
# Test streaming convolutional layers with striding, dilation.
cnn_filters = [1, 1, 1, 1]
cnn_kernel_size = [3, 3, 3, 3]
cnn_act = ['linear', 'linear', 'elu', 'elu']
cnn_dilation_rate = [1, 1, 1, 2]
cnn_strides = [2, 1, 3, 1]
cnn_use_bias = [False, False, False, False]
# prepare input data
params = test_utils.Params(cnn_strides)
x = np.arange(params.desired_samples)
frequency = 2.0
inp_audio = np.cos((2.0 * np.pi / params.desired_samples) * frequency *
x) + np.random.rand(1, params.desired_samples) * 0.5
if conv_cell == tf.keras.layers.SeparableConv1D:
kwargs = dict(
depthwise_initializer=tf.keras.initializers.GlorotUniform(
seed=123),
pointwise_initializer=tf.keras.initializers.GlorotUniform(
seed=456))
else:
kwargs = dict(
kernel_initializer=tf.keras.initializers.GlorotUniform(
seed=123))
# Prepare Keras native model.
model_native = conv_model_keras_native(params, conv_cell, cnn_filters,
cnn_kernel_size, cnn_act,
cnn_dilation_rate, cnn_strides,
cnn_use_bias, **kwargs)
model_native.summary()
# prepare non stream model
model = get_model(params, conv_cell, cnn_filters, cnn_kernel_size,
cnn_act, cnn_dilation_rate, cnn_strides,
cnn_use_bias, **kwargs)
model.summary()
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
native_out = model_native.predict(inp_audio)
stream_out = test.run_stream_inference(params, model_stream, inp_audio)
# normalize output data and compare them
channel = 0
non_stream_out = non_stream_out[0, :, channel]
native_out = native_out[0, :, channel]
stream_out = stream_out[0, :, channel]
min_len = min(stream_out.shape[0], non_stream_out.shape[0])
stream_out = stream_out[0:min_len]
native_out = native_out[0:min_len]
non_stream_out = non_stream_out[0:min_len]
self.assertAllEqual(non_stream_out.shape,
(params.desired_samples / np.prod(cnn_strides), ))
with self.subTest(name='stream_vs_non_stream'):
self.assertAllClose(stream_out, non_stream_out)
with self.subTest(name='non_stream_vs_native'):
self.assertAllClose(non_stream_out, native_out)
def test_cnn_model_end_to_end(self):
config = tf1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf1.Session(config=config)
tf1.keras.backend.set_session(sess)
test_utils.set_seed(123)
# data parameters
num_time_bins = 12
feature_size = 12
# model params.
total_stride = 2
params = test_utils.Params([total_stride], 0)
params.model_name = 'cnn'
params.cnn_filters = '2'
params.cnn_kernel_size = '(3,3)'
params.cnn_act = "'relu'"
params.cnn_dilation_rate = '(1,1)'
params.cnn_strides = '(2,2)'
params.dropout1 = 0.5
params.units2 = ''
params.act2 = ''
params.label_count = 2
params.return_softmax = True
params.quantize = 1 # apply quantization aware training
params.data_shape = (num_time_bins, feature_size)
params.preprocess = 'custom'
model = cnn.model(params)
model.summary()
# prepare training and testing data
train_images, train_labels = test_utils.generate_data(
img_size_y=num_time_bins, img_size_x=feature_size, n_samples=32)
test_images = train_images
test_labels = train_labels
# create and train quantization aware model in non streaming mode
model.compile(
optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(),
metrics=['accuracy'])
model.fit(
train_images,
train_labels,
epochs=1,
validation_data=(test_images, test_labels))
model.summary()
# one test image
train_image = train_images[:1,]
# run tf non streaming inference
non_stream_output_tf = model.predict(train_image)
# specify input data shape for streaming mode
params.data_shape = (total_stride, feature_size)
# TODO(rybakov) add params structure for model with no feature extractor
# prepare tf streaming model and use it to generate representative_dataset
with quantize.quantize_scope():
stream_quantized_model = utils.to_streaming_inference(
model, params, Modes.STREAM_EXTERNAL_STATE_INFERENCE)
calibration_data = prepare_calibration_data(stream_quantized_model,
total_stride, train_image)
def representative_dataset(dtype):
def _representative_dataset_gen():
for i in range(len(calibration_data)):
yield [
calibration_data[i][0].astype(dtype), # input audio packet
calibration_data[i][1].astype(dtype), # conv state
calibration_data[i][2].astype(dtype) # flatten state
]
return _representative_dataset_gen
# convert streaming quantization aware model to tflite
# and apply post training quantization
with quantize.quantize_scope():
tflite_streaming_model = utils.model_to_tflite(
sess, model, params,
Modes.STREAM_EXTERNAL_STATE_INFERENCE,
optimizations=[tf.lite.Optimize.DEFAULT],
inference_type=tf.int8,
experimental_new_quantizer=True,
representative_dataset=representative_dataset(np.float32))
# run tflite in streaming mode and compare output logits with tf
interpreter = tf.lite.Interpreter(model_content=tflite_streaming_model)
interpreter.allocate_tensors()
input_states = []
for detail in interpreter.get_input_details():
input_states.append(np.zeros(detail['shape'], dtype=np.float32))
stream_out_tflite = inference.run_stream_inference_classification_tflite(
params, interpreter, train_image, input_states)
self.assertAllClose(stream_out_tflite, non_stream_output_tf, atol=0.001)
def test_streaming_on_1d_data_strides(self, stride):
"""Tests Conv2DTranspose on 1d in streaming mode with different strides.
Args:
stride: controls the upscaling factor
"""
tf1.reset_default_graph()
config = tf1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf1.Session(config=config)
tf1.keras.backend.set_session(sess)
# model and data parameters
step = 1 # amount of data fed into streaming model on every iteration
params = test_utils.Params([step], clip_duration_ms=0.25)
# prepare input data: [batch, time, 1, channels]
x = np.random.rand(1, params.desired_samples, 1, self.input_channels)
inp_audio = x
# prepare non-streaming model
model = conv2d_transpose_model(
params,
filters=1,
kernel_size=(3, 1),
strides=(stride, 1),
channels=self.input_channels)
model.summary()
# set weights with bias
for layer in model.layers:
if isinstance(layer, tf.keras.layers.Conv2DTranspose):
layer.set_weights([
np.ones(layer.weights[0].shape),
np.zeros(layer.weights[1].shape) + 0.5
])
params.data_shape = (1, 1, self.input_channels)
# prepare streaming model
model_stream = utils.to_streaming_inference(
model, params, modes.Modes.STREAM_INTERNAL_STATE_INFERENCE)
model_stream.summary()
# run inference
non_stream_out = model.predict(inp_audio)
stream_out = inference.run_stream_inference(params, model_stream, inp_audio)
self.assertAllClose(stream_out, non_stream_out)
# Convert TF non-streaming model to TFLite external-state streaming model.
tflite_streaming_model = utils.model_to_tflite(
sess, model, params, modes.Modes.STREAM_EXTERNAL_STATE_INFERENCE)
self.assertTrue(tflite_streaming_model)
# Run TFLite external-state streaming inference.
interpreter = tf.lite.Interpreter(model_content=tflite_streaming_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
input_states = []
# before processing test sequence we create model state
for s in range(len(input_details)):
input_states.append(np.zeros(input_details[s]['shape'], dtype=np.float32))
stream_out_tflite_external_st = inference.run_stream_inference_tflite(
params, interpreter, inp_audio, input_states, concat=True)
# compare streaming TFLite with external-state vs TF non-streaming
self.assertAllClose(stream_out_tflite_external_st, non_stream_out)