def extract(cls, node):
pb = node.parameters
collect_until_token(pb, b'<MaxChange>')
max_change = read_binary_float_token(pb)
collect_until_token(pb, b'<L2Regularize>')
collect_until_token(pb, b'<LearningRate>')
collect_until_token(pb, b'<TimeOffsets>')
time_offsets = read_binary_vector(pb, False, np.int32)
collect_until_token(pb, b'<LinearParams>')
weights, weights_shape = read_binary_matrix(pb)
collect_until_token(pb, b'<BiasParams>')
bias_params = read_binary_vector(pb)
collect_until_token(pb, b'<OrthonormalConstraint>')
orthonormal_constraint = read_binary_float_token(
pb) # used only on training
collect_until_token(pb, b'<UseNaturalGradient>')
use_natural_grad = read_binary_bool_token(pb) # used only on training
collect_until_token(pb, b'<NumSamplesHistory>')
num_samples_hist = read_binary_float_token(pb)
collect_until_token(pb, b'<AlphaInOut>')
alpha_in_out = read_binary_float_token(pb), read_binary_float_token(
pb) # for training, usually (4, 4)
# according to Kaldi documentation http://kaldi-asr.org/doc/classkaldi_1_1nnet3_1_1TdnnComponent.html#details
# it looks like it's used only during training (but not 100% sure)
collect_until_token(pb, b'<RankInOut>')
rank_in_out = read_binary_integer32_token(
pb), read_binary_integer32_token(pb)
biases = np.array(bias_params) if len(bias_params) != 0 else None
attrs = {
'weights': np.reshape(weights, weights_shape),
'biases': biases,
'time_offsets': time_offsets,
}
TdnnComponent.update_node_stat(node, attrs)
return cls.enabled
def extract(cls, node):
pb = node.parameters
try:
collect_until_token(pb, b'<Dim>')
except Error:
try:
pb.seek(0)
collect_until_token(pb, b'<InputDim>')
except Error:
raise Error("Neither <Dim> nor <InputDim> were found")
in_dim = read_binary_integer32_token(pb)
try:
collect_until_token(pb, b'<TargetRms>')
target_rms = read_binary_float_token(pb)
except Error:
# model does not contain TargetRms
target_rms = 1.0
try:
collect_until_token(pb, b'<AddLogStddev>')
add_log = read_binary_bool_token(pb)
except Error:
# model does not contain AddLogStddev
add_log = False
if add_log is not False:
raise Error(
"AddLogStddev True in Normalize component is not supported")
scale = target_rms * np.sqrt(in_dim)
attrs = {
'eps': 0.00000001,
'across_spatial': 0,
'channel_shared': 1,
'in_dim': in_dim,
}
embed_input(attrs, 1, 'weights', [scale])
NormalizeOp.update_node_stat(node, attrs)
return cls.enabled
def extract(node):
pb = node.parameters
collect_until_token(pb, b'<Dim>')
dim = read_binary_integer32_token(pb)
collect_until_token(pb, b'<BlockDim>')
block_dim = read_binary_integer32_token(pb)
if block_dim != dim:
raise Error(
"Dim is not equal BlockDim for BatchNorm is not supported")
collect_until_token(pb, b'<Epsilon>')
eps = read_binary_float_token(pb)
collect_until_token(pb, b'<TargetRms>')
target_rms = read_binary_float_token(pb)
collect_until_token(pb, b'<TestMode>')
test_mode = read_binary_bool_token(pb)
if test_mode is not False:
raise Error("Test mode True for BatchNorm is not supported")
collect_until_token(pb, b'<StatsMean>')
mean = read_binary_vector(pb)
collect_until_token(pb, b'<StatsVar>')
var = read_binary_vector(pb)
scale = target_rms / np.sqrt(var + eps)
shift = -target_rms * mean / np.sqrt(var + eps)
attrs = {}
embed_input(attrs, 1, 'weights', scale)
embed_input(attrs, 2, 'biases', shift)
ScaleShiftOp.update_node_stat(node, attrs)
return __class__.enabled
def test_read_binary_bool_token(self):
true_value = self.bytesio_from(self.pack_value(True, self.bool_fmt))
false_value = self.bytesio_from(self.pack_value(False, self.bool_fmt))
self.assertTrue(read_binary_bool_token(true_value))
self.assertFalse(read_binary_bool_token(false_value))