def attention(query, key, value):
dk = C.reduce_sum(C.ones_like(query)) # cannot use sequence.last, will conflict with recurrence
# dk: [#, *] [1, ] and value = int(dim_of_query)
unpacked_key = C.sequence.unpack(key, padding_value=0, no_mask_output=True) # [#] [-3, key_dim]
unpacked_value = C.sequence.unpack(value, padding_value=0, no_mask_output=True) # [#] [-3, value_dim]
broadcasted_key = C.sequence.broadcast_as(unpacked_key, query) # [#, *] [-3, key_dim]
scaled = C.times_transpose(query, broadcasted_key) / dk
# [#, *] [q_dim] @ [#, *] [key_dim, -3], assert q_dim == key_dim
# scaled: [#, *] [-3, ] => for every key seq element, there is a corresponding score
# masked out invalid temporal connections to obey_sequence_order
if obey_sequence_order and max_seq_len:
unpacked_scaled, scaled_mask = C.sequence.unpack(scaled, padding_value=0).outputs
# unpacked_scaled: [#] [-3, -3] <== matrix will be top right diagonally zero-ed
# scaled_mask: [#] [-3,]
minus_inf = C.constant(-1e+30)
valid_connections = C.Constant(np.tril(np.ones((max_seq_len, max_seq_len)), k=0)) # [] [max_seq, max_seq]
valid_connections = C.reconcile_dynamic_axes(valid_connections, unpacked_scaled) # [#] [max_seq, max_seq]
valid_connections = C.crop_manual(valid_connections, unpacked_scaled, 0, 0) # [#] [-3, -3]
unpacked_scaled = C.element_select(valid_connections, unpacked_scaled, minus_inf) # [#] [-3, -3]
scaled = C.to_sequence_like(unpacked_scaled, query) # [#, *] [-3]
elif obey_sequence_order and not max_seq_len:
raise ValueError("max_seq_len must be defined when obey_sequence_order is True")
attended = C.times(C.softmax(scaled, axis=-1), C.sequence.broadcast_as(unpacked_value, query)) # [#, *] [value_dim,]
return attended
def centre_crop(larger_image, smaller_image, name: str = ''):
""" Centre crop spatial dimensions only.
Arguments:
larger_image: class:`~cntk.ops.functions.Function` that outputs the tensor to be centre cropped
smaller_image: class:`~cntk.ops.functions.Function` that outputs the reference tensor
name (str, optional): the name of the Function instance in the network
Returns:
:class:`~cntk.ops.functions.Function`
"""
input_shape = larger_image.shape # larger
referent_shape = smaller_image.shape # smaller
row_offset = int((input_shape[1] - referent_shape[1]) / 2)
col_offset = int((input_shape[2] - referent_shape[2]) / 2)
if row_offset == 0 and col_offset == 0:
return larger_image
elif row_offset < 0 or col_offset < 0:
raise ValueError(
f"offset became negative, check if image was passed correctly. "
f"larger image {larger_image.shape}, smaller image {smaller_image.shape}"
)
return C.crop_manual(larger_image,
smaller_image,
row_offset,
col_offset,
name=name)
def test_crop():
# Small network.
node_input = C.input_variable((1, 5, 5))
node_referent = C.input_variable((1, 5, 5))
node_output = C.layers.Sequential([
C.layers.Convolution2D(filter_shape=(3, 3),
num_filters=1,
init=1,
strides=(2, 2),
pad=True,
bias=False),
C.layers.MaxPooling(filter_shape=(3, 3), strides=(2, 2), pad=True),
C.layers.ConvolutionTranspose(filter_shape=(4, 4),
num_filters=1,
strides=(4, 4),
init=1,
bias=False)
])(node_input)
# Input data.
input_map = {
node_input: -np.arange(25).reshape(1, 1, 5, 5).astype(np.float32),
node_referent: np.zeros([1, 1, 5, 5]).astype(np.float32)
}
# Expected cropped output.
expected = [-12, -12, -12, -24, -24] * 3 + [-63, -63, -63, -81, -81] * 2
expected = np.asarray(expected, dtype=np.float32).reshape(1, 1, 5, 5)
# Test crop with explicitly specified offsets.
cropped = C.crop_manual(node_output, node_referent, 1, 1).eval(input_map)
assert np.array_equal(cropped, expected)
# Test crop with automatically computed offsets where inputs
# have common ancestor.
cropped = C.crop_automatic(node_output, node_input).eval(input_map)
assert np.array_equal(cropped, expected)
# Test crop with automatically computed offsets where inputs do not
# have common ancestor.
cropped = C.crop_automatic_with_ancestors(node_output, node_referent,
node_input,
node_referent).eval(input_map)
assert np.array_equal(cropped, expected)
def test_crop():
# Small network.
node_input = C.input_variable((1, 5, 5))
node_referent = C.input_variable((1, 5, 5))
node_output = C.layers.Sequential([
C.layers.Convolution2D(filter_shape = (3, 3),
num_filters = 1,
init = 1,
strides = (2, 2),
pad = True,
bias = False),
C.layers.MaxPooling(filter_shape = (3, 3),
strides = (2, 2),
pad = True),
C.layers.ConvolutionTranspose(filter_shape = (4, 4),
num_filters = 1,
strides = (4, 4),
init = 1,
bias = False)])(node_input)
# Input data.
input_map = {
node_input: -np.arange(25).reshape(1, 1, 5, 5).astype(np.float32),
node_referent: np.zeros([1, 1, 5, 5]).astype(np.float32)
}
# Expected cropped output.
expected = [-12, -12, -12, -24, -24] * 3 + [-63, -63, -63, -81, -81] * 2
expected = np.asarray(expected, dtype = np.float32).reshape(1, 1, 5, 5)
# Test crop with explicitly specified offsets.
cropped = C.crop_manual(node_output, node_referent, 1, 1).eval(input_map)
assert np.array_equal(cropped, expected)
# Test crop with automatically computed offsets where inputs
# have common ancestor.
cropped = C.crop_automatic(node_output, node_input).eval(input_map)
assert np.array_equal(cropped, expected)
# Test crop with automatically computed offsets where inputs do not
# have common ancestor.
cropped = C.crop_automatic_with_ancestors(
node_output, node_referent, node_input, node_referent).eval(input_map)
assert np.array_equal(cropped, expected)
