def call(self, inputs, params=None):
if params[self.name + '/depthwise_kernel:0'] is None:
return super(layers.SeparableConv2D, self).call(inputs)
else:
depthwise_kernel = params.get(self.name + '/depthwise_kernel:0')
pointwise_kernel = params.get(self.name + '/pointwise_kernel:0')
bias = params.get(self.name + '/bias:0')
# Apply the actual ops.
if self.data_format == 'channels_last':
strides = (1, ) + self.strides + (1, )
else:
strides = (1, 1) + self.strides
outputs = nn.separable_conv2d(
inputs,
depthwise_kernel,
pointwise_kernel,
strides=strides,
padding=self.padding.upper(),
rate=self.dilation_rate,
data_format=conv_utils.convert_data_format(self.data_format,
ndim=4))
if self.use_bias:
outputs = nn.bias_add(outputs,
bias,
data_format=conv_utils.convert_data_format(
self.data_format, ndim=4))
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, inputs):
inputs_shape = array_ops.shape(inputs)
batch_size = inputs_shape[0]
if self.data_format == 'channels_first':
w_axis = 2
else:
w_axis = 1
width = inputs_shape[w_axis]
kernel_w, = self.kernel_size
stride_w, = self.strides
if self.output_padding is None:
out_pad_w = None
else:
out_pad_w, = self.output_padding
# Infer the dynamic output shape:
out_width = conv_utils.deconv_output_length(
width,
kernel_w,
padding=self.padding,
output_padding=out_pad_w,
stride=stride_w,
dilation=self.dilation_rate[0],
)
if self.data_format == 'channels_first':
output_shape = (batch_size, self.filters, out_width)
else:
output_shape = (batch_size, out_width, self.filters)
output_shape_tensor = array_ops.stack(output_shape)
outputs = conv1d_transpose(
inputs,
self.kernel,
output_shape_tensor,
stride=stride_w,
padding=self.padding.upper(),
data_format=conv_utils.convert_data_format(self.data_format, ndim=3),
)
if not context.executing_eagerly():
# Infer the static output shape:
out_shape = self.compute_output_shape(inputs.shape)
outputs.set_shape(out_shape)
if self.use_bias:
outputs = nn.bias_add(
outputs,
self.bias,
data_format=conv_utils.convert_data_format(self.data_format, ndim=4),
)
if self.activation is not None:
return self.activation(outputs)
return outputs
def build(self, input_shape):
with tf.variable_scope(self.name):
input_shape = TensorShape(input_shape)
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if input_shape[channel_axis].value is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
input_dim = int(input_shape[channel_axis])
kernel_shape = self.kernel_size + (input_dim, self.filters)
print("input_shape: {}".format(input_shape))
self.kernel = sn_kernel(shape=kernel_shape, scope='kernel')
self.bias = tf.get_variable(
name='bias',
shape=[self.filters],
initializer=tf.initializers.zeros(dtype=self.dtype))
self._convolution_op = Convolution(
input_shape,
filter_shape=self.kernel.get_shape(),
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self.padding.upper(),
data_format=conv_utils.convert_data_format(
self.data_format, self.rank + 2))
self.built = True
def build_dilated_conv(self, input_shape, input_dim):
"""
Define trainable kernel and bias variables for dilated conv layers
Prepare conv operations
"""
kernel_shape = self.kernel_size + (input_dim, self.filters)
self.kernel = self.add_weight(name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
for dilation in self.dilations:
convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.get_shape(),
dilation_rate=dilation,
strides=self.strides,
padding="SAME",
data_format=conv_utils.convert_data_format(
self.data_format, self.rank + 2))
self.conv_ops.append(convolution_op)
def call(self, inputs):
data_format = conv_utils.convert_data_format(self.data_format,
self.rank + 2)
inputs, tf_data_format = K._preprocess_conv2d_input(
inputs, self.data_format)
inputs = tf.extract_image_patches(
inputs,
ksizes=(1, ) + K.int_shape(self.kernel)[:2] + (1, ),
strides=(1, ) + self.strides + (1, ),
rates=(1, ) + self.dilation_rate + (1, ),
padding=self.padding.upper(),
)
kernel = K.reshape(self.kernel, (-1, self.filters))
outputs = self.kernel_function([inputs, kernel])
if self.data_format == 'channels_first':
outputs = K.permute_dimensions(outputs, (0, 3, 1, 2))
if self.use_bias:
outputs = nn.bias_add(outputs, self.bias, data_format=data_format)
if self.activation is not None:
outputs = self.activation(outputs)
return outputs
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
kernel_shape = self.kernel_size + (input_channel, self.filters)
self.kernel = self.add_weight(name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
try:
# Disable variable partitioning when creating the variable
if hasattr(self, '_scope') and self._scope:
partitioner = self._scope.partitioner
self._scope.set_partitioner(None)
else:
partitioner = None
self.u = self.add_weight(
name='sn_u',
shape=(1, tf.reduce_prod(kernel_shape[:-1])),
dtype=self.dtype,
initializer=tf.keras.initializers.ones,
synchronization=tf.VariableSynchronization.ON_READ,
trainable=False,
aggregation=tf.VariableAggregation.MEAN)
finally:
if partitioner:
self._scope.set_partitioner(partitioner)
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
channel_axis = self._get_channel_axis()
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_channel})
self._build_conv_op_input_shape = input_shape
self._build_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._padding_op,
data_format=self._conv_op_data_format)
self.built = True
def __init__(
self,
wavelet,
level=1,
strides=2,
rank=1,
mode='symmetric',
data_format=None,
**kwargs
):
L.Layer.__init__(self, **kwargs)
Wavelet.__init__(self, wavelet, level)
self.mode = mode
self.rank = rank
self.strides = conv_utils.normalize_tuple(strides, rank, "strides")
self.data_format = conv_utils.normalize_data_format(data_format)
self._channels_first = self.data_format == "channels_first"
# All the convolutions will be permuted to channels last
self.conv_format = conv_utils.convert_data_format(
"channels_last", self.rank+2)
self.filters = self.build_filters(rank, self.dec_lo, self.dec_hi)
offset = 1 if self.p % 2 or self.p <= 2 else 2
padding = self.rank * ([self.p-1, self.p-offset],)
if self._channels_first:
self.padding = tf.constant([[0, 0], [0, 0], *padding])
else:
self.padding = tf.constant([[0, 0], *padding, [0, 0]])
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
kernel_shape = self.kernel_size + (input_channel, self.filters)
self.kernel = self.add_weight(name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
channel_axis = self._get_channel_axis()
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_channel})
self._build_conv_op_input_shape = input_shape
self._build_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._padding_op,
data_format=self._conv_op_data_format)
self.mask = self.create_mask(self.mask_type, self.color_conditioning)
self.built = True
def Conv2D(self, error, layer):
"""
Todo: Propagate error through Conv2D layer
:param error: Propagated Error Tensor from previous layer --> l-1
:param layer: Current Layer for doing error Propagation --> l
:return: Propagated Error for next layer --> l
"""
if error is None:
return None
logging.info(f'Propagate Error {error.shape}')
error = tf.nn.conv2d(input=error,
filters=layer.weights[0]**2,
strides=layer.strides,
# Hier muss man das groß schreiben im model klein (Conv(Layer) fkt. _get_padding_op)!
padding=layer.padding.upper(),
# conv_utils.py convert_data_format convertiert zu NH...
data_format=convert_data_format(layer.data_format, error.shape.ndims),
dilations=layer.dilation_rate
)
if self._debug[self.DEBUG.plot]:
if not isinstance(error, (np.ndarray)):
error = error.numpy()
sns.kdeplot(error.flatten(), cumulative=False, bw=0.001, shade=True,
label=f"Error-{layer.__class__.__name__}")
plt.legend()
''' Check if an activation function is placed within the layer'''
return self.Activation(error, layer)
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if input_shape.dims[channel_axis].value is None:
raise ValueError(
'The channel dimension of the inputs should be defined. Found `None`.'
)
input_dim = int(input_shape[channel_axis])
wshape = [
self.kernel_size[0] * self.kernel_size[1] - 1, input_dim,
self.filters
]
W = self.add_weight(name='kernel',
shape=wshape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
self.kernel = array_ops.concat(
(W[:wshape[0] // 2], array_ops.zeros(
(1, wshape[1], wshape[2])), W[wshape[0] // 2:]),
axis=0)
self.kernel = array_ops.reshape(self.kernel, [
self.kernel_size[0], self.kernel_size[1], input_dim, self.filters
])
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
self.input_spec = InputSpec(ndim=4, axes={channel_axis: input_dim})
if self.padding == 'causal':
op_padding = 'valid'
else:
op_padding = self.padding
if not isinstance(op_padding, (list, tuple)):
op_padding = op_padding.upper()
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=op_padding,
data_format=conv_utils.convert_data_format(self.data_format, 4))
self.built = True
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if input_shape.dims[channel_axis].value is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
input_dim = int(input_shape[channel_axis])
kernel_shape = self.kernel_size + (input_dim, self.filters)
self.wn_g = self.add_weight(
name='wn_g',
shape=(self.filters, ),
initializer=tf.keras.initializers.RandomUniform(1, 1),
trainable=True,
dtype=self.dtype)
self.kernel = self.add_weight(name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_dim})
if self.padding == 'causal':
op_padding = 'valid'
else:
op_padding = self.padding
if not isinstance(op_padding, (list, tuple)):
op_padding = op_padding.upper()
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=op_padding,
data_format=conv_utils.convert_data_format(self.data_format,
self.rank + 2))
self.built = True
def call(self, inputs, training=True): # 抄自系统的源码
inputs_shape = array_ops.shape(inputs)
batch_size = inputs_shape[0]
if self.data_format == 'channels_first':
h_axis, w_axis = 2, 3
else:
h_axis, w_axis = 1, 2
height, width = inputs_shape[h_axis], inputs_shape[w_axis]
kernel_h, kernel_w = self.kernel_size
stride_h, stride_w = self.strides
if self.output_padding is None:
out_pad_h = out_pad_w = None
else:
out_pad_h, out_pad_w = self.output_padding
# Infer the dynamic output shape:
out_height = conv_utils.deconv_output_length(height,
kernel_h,
padding=self.padding,
output_padding=out_pad_h,
stride=stride_h,
dilation=self.dilation_rate[0])
out_width = conv_utils.deconv_output_length(width,
kernel_w,
padding=self.padding,
output_padding=out_pad_w,
stride=stride_w,
dilation=self.dilation_rate[1])
if self.data_format == 'channels_first':
output_shape = (batch_size, self.filters, out_height, out_width)
else:
output_shape = (batch_size, out_height, out_width, self.filters)
output_shape_tensor = array_ops.stack(output_shape)
outputs = backend.conv2d_transpose(
inputs,
self.compute_spectral_normal(training=training), # self.kernel,
output_shape_tensor,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if not context.executing_eagerly():
# Infer the static output shape:
out_shape = self.compute_output_shape(inputs.shape)
outputs.set_shape(out_shape)
if self.use_bias:
outputs = nn.bias_add(
outputs,
self.bias,
data_format=conv_utils.convert_data_format(self.data_format, ndim=4))
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, inputs, **kwargs):
outputs = self.pool_function(
inputs,
self.pool_size,
strides=self.strides,
padding=self.padding.upper(),
data_format=conv_utils.convert_data_format(self.data_format, 3))
return outputs
def __init__(self,
rank,
filters,
kernel_size,
strides=1,
padding='valid',
data_format=None,
dilation_rate=1,
groups=1,
activation=None,
use_bias=True,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
bias_constraint=None,
trainable=True,
name=None,
conv_op=None,
**kwargs):
super(Conv_ada, self).__init__(
trainable=trainable,
name=name,
activity_regularizer=regularizers.get(activity_regularizer),
**kwargs)
self.rank = rank
if isinstance(filters, float):
filters = int(filters)
self.filters = filters
self.groups = groups or 1
self.kernel_size = conv_utils.normalize_tuple(kernel_size, rank,
'kernel_size')
self.strides = conv_utils.normalize_tuple(strides, rank, 'strides')
self.padding = conv_utils.normalize_padding(padding)
self.data_format = conv_utils.normalize_data_format(data_format)
self.dilation_rate = conv_utils.normalize_tuple(
dilation_rate, rank, 'dilation_rate')
self.activation = activations.get(activation)
self.use_bias = use_bias
self.kernel_initializer = initializers.get(kernel_initializer)
self.bias_initializer = initializers.get(bias_initializer)
self.kernel_regularizer = regularizers.get(kernel_regularizer)
self.bias_regularizer = regularizers.get(bias_regularizer)
self.kernel_constraint = constraints.get(kernel_constraint)
self.bias_constraint = constraints.get(bias_constraint)
self.input_spec = InputSpec(min_ndim=self.rank + 2)
self._validate_init()
self._is_causal = self.padding == 'causal'
self._channels_first = self.data_format == 'channels_first'
self._tf_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
def build(
self,
input_shape,
):
input_shape = tf.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
kernel_shape = self.kernel_size + (input_channel, self.filters)
self.kernel = self.add_weight(
name="kernel",
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype,
)
# Add mask to remove weights on half of the kernel to the left
# (only keep future
# context)
left_kernel_dims = (self.future_context, input_channel, self.filters)
left_kernel = tf.fill(dims=left_kernel_dims, value=0)
right_kernel_dims = (self.future_context + 1, input_channel, self.filters)
right_kernel = tf.fill(dims=right_kernel_dims, value=1)
mask_kernel = tf.cast(tf.concat([left_kernel, right_kernel], axis=0), dtype=self.dtype)
self.kernel = tf.multiply(self.kernel, mask_kernel)
if self.use_bias:
self.bias = self.add_weight(
name="bias",
shape=(self.filters,),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype,
)
else:
self.bias = None
channel_axis = self._get_channel_axis()
self.input_spec = tf.keras.layers.InputSpec(ndim=self.rank + 2, axes={channel_axis: input_channel})
self.make_conv_op_input_shape = input_shape
self.make_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(self.data_format, self.rank + 2)
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._padding_op,
data_format=self._conv_op_data_format,
)
self.built = True
def call(self, inputs, **kwargs):
M = self.kernel_size[0]
N = self.kernel_size[1]
# (D_mul, input_channel // groups, filters)
W_shape = self.W.get_shape().as_list()
# (M, N, input_channel // groups, filters)
DoW_shape = (M, N, W_shape[1], W_shape[2])
if M * N > 1:
input_channel = W_shape[1] * self.groups
D_diag = tf.tile(tf.reshape(tf.eye(M * N), (M * N, M * N, 1)),
(1, self.D_mul // (M * N), input_channel))
if self.D_mul % (M * N) != 0: # the cases when D_mul > MxN
zeros = tf.zeros((M * N, self.D_mul % (M * N), input_channel))
D_diag = tf.concat([D_diag, zeros], axis=1)
######################### Compute DoW #################
# (M * N, D_mul, input_channel)
D = self.D + D_diag
# (D_mul, input_channel, filters // groups)
W = tf.reshape(self.W, (self.D_mul, input_channel, -1))
# einsum outputs (M * N, input_channel, filters // groups),
# which is reshaped to
# (M, N, input_channel // groups, filters)
DoW = tf.reshape(tf.einsum('msi,sio->mio', D, W), DoW_shape)
#######################################################
else:
# in this case D_mul == M*N
# reshape from
# (D_mul, input_channel // groups, filters)
# to
# (M, N, input_channel // groups, filters)
DoW = tf.reshape(self.W, DoW_shape)
data_format = conv_utils.convert_data_format(self.data_format, ndim=4)
outputs = tf.nn.conv2d(inputs, DoW, strides=self.strides,
padding=self.padding.upper(),
data_format=data_format,
dilations=self.dilation_rate)
if self.use_bias:
if self.data_format == 'channels_first':
if self.rank == 1:
# nn.bias_add does not accept a 1D input tensor.
bias = array_ops.reshape(self.bias, (1, self.filters, 1))
outputs += bias
else:
outputs = nn.bias_add(outputs, self.bias, data_format='NCHW')
else:
outputs = nn.bias_add(outputs, self.bias, data_format='NHWC')
if self.activation is not None:
return self.activation(outputs)
return outputs
def convolutional_similarity_single(input_i: tf.Tensor, bias_i: tf.Tensor):
# tf.nn.conv* like its filter shaped `spatial_dims + (channels_in, channels_out)`
# EG: conv2d filters are shaped `(H, W, C, filters)`.
# Since I only want 1 filter, I insert an axis at -1
bias_i = tf.expand_dims(bias_i, axis=-1)
conv_norm = tf.nn.convolution(input=input_i,
filters=bias_i,
strides=strides,
padding=padding,
data_format=convert_data_format('channels_last', ndim=ndim))
return tf.abs(tf.squeeze(conv_norm, axis=-1))
def test_convert_data_format(self):
self.assertEqual('NCDHW', conv_utils.convert_data_format(
'channels_first', 5))
self.assertEqual('NCHW', conv_utils.convert_data_format(
'channels_first', 4))
self.assertEqual('NCW', conv_utils.convert_data_format('channels_first', 3))
self.assertEqual('NHWC', conv_utils.convert_data_format('channels_last', 4))
self.assertEqual('NWC', conv_utils.convert_data_format('channels_last', 3))
self.assertEqual('NDHWC', conv_utils.convert_data_format(
'channels_last', 5))
with self.assertRaises(ValueError):
conv_utils.convert_data_format('invalid', 2)
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
if input_channel % self.groups != 0:
raise ValueError(
'The number of input channels must be evenly divisible by the number '
'of groups. Received groups={}, but the input has {} channels '
'(full input shape is {}).'.format(self.groups, input_channel,
input_shape))
kernel_shape = self.kernel_size + (
input_channel // self.groups,
self.filters,
)
self.kernel = self.add_weight(
name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype,
)
if self.use_bias:
self.bias = self.add_weight(
name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype,
)
else:
self.bias = None
channel_axis = self._get_channel_axis()
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_channel})
self._build_conv_op_input_shape = input_shape
self._build_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
self._convolution_op = Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._padding_op,
data_format=self._conv_op_data_format,
)
self.built = True
def build(self, input_shape):
stack_size, input_dim = input_shape[-2:]
if stack_size is None or input_dim is None:
raise ValueError(
"The two last dimensions of the inputs should be defined. Found `None`."
)
kernel_shape = (stack_size, *self.kernel_size, stack_size, input_dim,
self.filters)
self.kernel = self.add_weight(name="kernel",
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
if self.use_bias:
bias_shape = (stack_size, self.filters)
self.bias = self.add_weight(name="bias",
shape=bias_shape,
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
kernel_mask = get_stack_kernel_mask(self.kernel_size, stack_size,
self.mask_center)
self.kernel_mask = tf.constant(kernel_mask,
dtype=tf.float32,
name="kernel_mask")
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={
-1: input_dim,
-2: stack_size
})
strides = conv_utils.normalize_tuple(1, self.rank, "strides")
dilation_rate = conv_utils.normalize_tuple(1, self.rank,
"dilation_rate")
self._convolution_op = nn_ops.Convolution(
input_shape=input_shape,
filter_shape=tf.TensorShape(kernel_shape[1:]),
dilation_rate=dilation_rate,
strides=strides,
padding=get_stack_padding(self.kernel_size),
data_format=conv_utils.convert_data_format("channels_last",
self.rank + 2))
def call(self, inputs):
if self.data_format == 'channels_last':
pool_shape = (1, ) + self.pool_size + (1, )
strides = (1, ) + self.strides + (1, )
else:
pool_shape = (1, 1) + self.pool_size
strides = (1, 1) + self.strides
outputs = self.pool_function(
inputs,
ksize=pool_shape,
strides=strides,
padding=self.padding.upper(),
data_format=conv_utils.convert_data_format(self.data_format, 4))
return outputs
def call(self, inputs):
if self.data_format == 'channels_last':
pool_shape = (1,) + self.pool_size + (1,)
strides = (1,) + self.strides + (1,)
else:
pool_shape = (1, 1) + self.pool_size
strides = (1, 1) + self.strides
outputs = self.pool_function(
inputs,
ksize=pool_shape,
strides=strides,
padding=self.padding.upper(),
data_format=conv_utils.convert_data_format(self.data_format, 4))
return outputs
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
if self.data_format == 'channels_first':
channel_axis = 1
else:
channel_axis = -1
if input_shape[channel_axis].value is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
input_dim = int(input_shape[channel_axis])
kernel_shape = self.kernel_size + (input_dim, self.filters)
self.kernel = self.add_weight(name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
self.loc = self.add_variable(
'loc',
shape=kernel_shape,
initializer=tf.keras.initializers.RandomNormal(
mean=self.loc_mean, stddev=self.loc_stddev, seed=None),
regularizer=None,
constraint=None,
dtype=self.dtype,
trainable=True)
self.loc2 = self.loc.numpy()
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_dim})
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.get_shape(),
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self.padding.upper(),
data_format=conv_utils.convert_data_format(self.data_format,
self.rank + 2))
self.built = True
def test_convert_data_format(self):
self.assertEqual('NCDHW',
conv_utils.convert_data_format('channels_first', 5))
self.assertEqual('NCHW',
conv_utils.convert_data_format('channels_first', 4))
self.assertEqual('NCW',
conv_utils.convert_data_format('channels_first', 3))
self.assertEqual('NHWC',
conv_utils.convert_data_format('channels_last', 4))
self.assertEqual('NWC',
conv_utils.convert_data_format('channels_last', 3))
self.assertEqual('NDHWC',
conv_utils.convert_data_format('channels_last', 5))
with self.assertRaises(ValueError):
conv_utils.convert_data_format('invalid', 2)
def call(self, inputs):
# Apply the actual ops
if self.data_format is not 'channels_last':
raise ValueError("mpusim_separable_conv2d "
"requires NHWC data format")
strides = (1,) + self.strides + (1,)
outputs = mpusim_separable_conv2d_op_impl(inputs,
self.depthwise_kernel,
self.pointwise_kernel,
strides,
self.padding.upper(),
self.dilation_rate,
None,
conv_utils.convert_data_format(self.data_format, ndim=4),
self.activations_datatype_size_byte,
self.weights_datatype_size_byte,
self.results_datatype_size_byte,
self.systolic_array_height,
self.systolic_array_width,
self.activation_fifo_depth,
self.accumulator_array_height,
self.log_file_output_dir,
self.model_name)
if self.use_bias:
outputs = nn.bias_add(outputs,
self.bias,
data_format= \
conv_utils.convert_data_format(self.data_format, ndim=4))
if self.activation is not None:
return self.activation(outputs)
return outputs
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
kernel_shape = self.kernel_size + (input_channel, self.filters)
self.kernel_weights = self.add_weight(
name='kernel_weights',
shape=kernel_shape,
initializer=tf.random_normal_initializer(mean=0.0, stddev=0.05),
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
self.kernel_log_scale = self.add_weight(name='kernel_log_scale',
shape=(1, 1, 1, self.filters),
initializer=tf.constant_initializer(value=0.),
regularizer=None,
constraint=None,
trainable=True,
dtype=self.dtype)
if self.use_bias:
self.bias = self.add_weight(
name='bias',
shape=(self.filters,),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
channel_axis = self._get_channel_axis()
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_channel})
self._build_conv_op_input_shape = input_shape
self._build_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._padding_op,
data_format=self._conv_op_data_format)
self.built = True
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
kernel_shape = self.kernel_size + (input_channel, self.filters)
if self.use_bias:
pass
else:
self.bias = None
channel_axis = self._get_channel_axis()
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_channel})
self._build_conv_op_input_shape = input_shape
self._build_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
self.built = True
def call(self, x, mask=None):
"""Layer functionality."""
self.mem_input.assign_add(x)
_, max_idxs = tf.nn.max_pool_with_argmax(
self.mem_input,
self.pool_size,
self.strides,
self.padding.upper(),
conv_utils.convert_data_format(self.data_format, 4),
include_batch_in_index=True)
x_max = tf.scatter_nd(tf.reshape(max_idxs, (-1, 1)),
tf.ones(tf.size(max_idxs)),
[tf.reduce_prod(x.shape)])
x_max = tf.reshape(x_max, x.shape)
x_masked = x * x_max
return MaxPooling2D.call(self, x_masked)
def call(self, inputs):
if self.padding == 'causal':
inputs = array_ops.pad(inputs, self._compute_causal_padding())
if self.data_format == 'channels_last':
strides = (1, ) + self.strides * 2 + (1, )
spatial_start_dim = 1
else:
strides = (1, 1) + self.strides * 2
spatial_start_dim = 2
# Explicitly broadcast inputs and kernels to 4D.
inputs = array_ops.expand_dims(inputs, spatial_start_dim)
if self.common_kernel == True:
#Need to replicate kernel {channels} times over axis 1
dw_kernel = tf.tile(self.depthwise_kernel, (1, self.channels, 1))
bias_kernel = tf.tile(self.bias, (self.channels, ))
else:
dw_kernel = self.depthwise_kernel
bias_kernel = self.bias
dw_kernel = array_ops.expand_dims(dw_kernel, 0)
if self.padding == 'causal':
op_padding = 'valid'
else:
op_padding = self.padding
outputs = nn.depthwise_conv2d(
inputs,
dw_kernel,
strides=strides,
padding=op_padding.upper(),
data_format=conv_utils.convert_data_format(self.data_format,
ndim=4))
outputs = array_ops.squeeze(outputs, [spatial_start_dim])
if self.use_bias:
outputs = backend.bias_add(outputs,
bias_kernel,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
input_channel = self._get_input_channel(input_shape)
kernel_shape = self.kernel_size + (input_channel, self.filters)
self.kernelA = self.add_weight(name='kernelA',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
self.kernelB = K.constant(self.kernelB_init_weight)
self.kernel = K.transpose(
K.dot(K.transpose(self.kernelA), self.kernelB))
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
channel_axis = self._get_channel_axis()
self.input_spec = InputSpec(ndim=self.rank + 2,
axes={channel_axis: input_channel})
self._build_conv_op_input_shape = input_shape
self._build_input_channel = input_channel
self._padding_op = self._get_padding_op()
self._conv_op_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
self._convolution_op = nn_ops.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._padding_op,
data_format=self._conv_op_data_format)
self.built = True
def __init__(self, rank, filters, kernel_size, dtype, strides=1, padding='valid', data_format=None, dilation_rate=1,
groups=1, activation=None, use_bias=True,
kernel_initializer=ComplexGlorotUniform(), bias_initializer=Zeros(),
kernel_regularizer=None, bias_regularizer=None, # TODO: Not yet working
activity_regularizer=None, kernel_constraint=None, bias_constraint=None,
trainable=True, name=None, conv_op=None, **kwargs):
if kernel_regularizer is not None or bias_regularizer is not None:
logger.warning(f"Sorry, regularizers are not implemented yet, this parameter will take no effect")
super(ComplexConv, self).__init__(
trainable=trainable,
name=name,
activity_regularizer=regularizers.get(activity_regularizer),
**kwargs)
self.rank = rank
self.my_dtype = tf.dtypes.as_dtype(dtype)
# I use no default dtype to make sure I don't forget to give it to my ComplexConv layers
if isinstance(filters, float):
filters = int(filters)
self.filters = filters
self.groups = groups or 1
self.kernel_size = conv_utils.normalize_tuple(
kernel_size, rank, 'kernel_size')
self.strides = conv_utils.normalize_tuple(strides, rank, 'strides')
self.padding = conv_utils.normalize_padding(padding)
self.data_format = conv_utils.normalize_data_format(data_format)
self.dilation_rate = conv_utils.normalize_tuple(
dilation_rate, rank, 'dilation_rate')
self.activation = activations.get(activation)
self.use_bias = use_bias
self.kernel_initializer = initializers.get(kernel_initializer)
self.bias_initializer = initializers.get(bias_initializer)
self.kernel_regularizer = regularizers.get(kernel_regularizer)
self.bias_regularizer = regularizers.get(bias_regularizer)
self.kernel_constraint = constraints.get(kernel_constraint)
self.bias_constraint = constraints.get(bias_constraint)
self.input_spec = InputSpec(min_ndim=self.rank + 2)
self._validate_init()
self._is_causal = self.padding == 'causal'
self._channels_first = self.data_format == 'channels_first'
self._tf_data_format = conv_utils.convert_data_format(
self.data_format, self.rank + 2)
def build(self, input_shape):
input_shape = tf.TensorShape(input_shape)
channel_axis = self._get_channel_axis()
if input_shape.dims[channel_axis].value is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
input_dim = int(input_shape[channel_axis])
kernel_shape = self.kernel_size + (input_dim, self.filters)
self.kernel = self.add_weight(name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
constraint=self.kernel_constraint,
trainable=True,
dtype=self.dtype)
if self.use_bias:
self.bias = self.add_weight(name='bias',
shape=(self.filters, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
constraint=self.bias_constraint,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
mean, variance = tf.nn.moments(self.kernel.value(), [0, 1, 2],
keepdims=True)
self.kernel.assign_sub(mean)
self.kernel.assign(self.kernel.value() / (tf.sqrt(variance) + 1e-5))
self.input_spec = tf.keras.backend.InputSpec(
ndim=self.rank + 2, axes={channel_axis: input_dim})
self._convolution_op = tf.nn.Convolution(
input_shape,
filter_shape=self.kernel.shape,
dilation_rate=self.dilation_rate,
strides=self.strides,
padding=self._get_padding_op(),
data_format=conv_utils.convert_data_format(self.data_format,
self.rank + 2))
self.built = True