def _normalize_sequence(length, inputs, layout, merge, in_layout=None):
assert inputs is not None, \
"unroll(inputs=None) has been deprecated. " \
"Please create input variables outside unroll."
axis = layout.find('T')
in_axis = in_layout.find('T') if in_layout is not None else axis
if isinstance(inputs, symbol.Symbol):
if merge is False:
assert len(inputs.list_outputs()) == 1, \
"unroll doesn't allow grouped symbol as input. Please convert " \
"to list with list(inputs) first or let unroll handle splitting."
inputs = list(
symbol.split(inputs,
axis=in_axis,
num_outputs=length,
squeeze_axis=1))
else:
assert length is None or len(inputs) == length
if merge is True:
inputs = [symbol.expand_dims(i, axis=axis) for i in inputs]
inputs = symbol.Concat(*inputs, dim=axis)
in_axis = axis
if isinstance(inputs, symbol.Symbol) and axis != in_axis:
inputs = symbol.swapaxes(inputs, dim0=axis, dim1=in_axis)
return inputs, axis
def add_loss(self, splits: sym.Variable):
"""Add loss functions.
Below, we splice the network output accordingly to compute losses for
the following:
1. Bounding box attributes
2. Class probabilities
3. IOUS as "confidence scores"
Below, the ugly splice functions are replacements for reshaping.
Instead, split along a dimension into multiple chunks, and then
restack the arrays in a consistent way.
Due to a quirk in MXNet, we create a placeholder label_score. However,
we actually use pred_box and label_box to compute IOU (true labels),
which are then compared with pred_score.
"""
num_splits = int(NUM_OUT_CHANNELS / ANCHORS_PER_GRID)
splits = list(sym.split(splits, num_outputs=num_splits))
# Compute loss for bounding box
pred_box = sym.concat(*splits[:NUM_BBOX_ATTRS])
loss_box = mx.sym.Custom(
data=pred_box,
label=self.label_box,
op_type='LinearRegressionOutputWithMask')
# Compute loss for class probabilities
cidx = NUM_BBOX_ATTRS + NUM_CLASSES
pred_class = reformat(sym.concat(*splits[NUM_BBOX_ATTRS:cidx]), pkg=sym)
label_class = reformat(self.label_class, pkg=sym)
loss_class = sym.SoftmaxOutput(data=pred_class, label=label_class)
# Compute loss for confidence scores - see doc above for explanation
pred_score = splits[cidx]
loss_iou = mx.symbol.Custom(
data=pred_score,
label=sym.concat(self.label_score, pred_box, self.label_box),
op_type='IOURegressionOutputWithMask')
return mx.sym.Group([loss_box, loss_class, loss_iou])
def format_sequence(length, inputs, layout, merge, in_layout=None):
"""
`Original Code <https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/gluon/rnn/rnn_cell.py#L52>`_
Parameters
----------
length
inputs
layout
merge
in_layout
Returns
-------
Examples
--------
>>> import mxnet.ndarray as nd
>>> seq = [[[0] * 4, [2] * 4, [4] * 4], [[1] * 4, [3] * 4, [5] * 4]]
>>> seq1, axis, _, batch_size = format_sequence(3, nd.array(seq), "NTC", False)
>>> seq1 # doctest: +NORMALIZE_WHITESPACE
[
[[0. 0. 0. 0.]
[1. 1. 1. 1.]]
<NDArray 2x4 @cpu(0)>,
[[2. 2. 2. 2.]
[3. 3. 3. 3.]]
<NDArray 2x4 @cpu(0)>,
[[4. 4. 4. 4.]
[5. 5. 5. 5.]]
<NDArray 2x4 @cpu(0)>]
>>> axis
1
>>> batch_size
2
>>> seq2, _, _, _ = format_sequence(3, nd.array(seq), "NTC", True)
>>> seq2 # doctest: +NORMALIZE_WHITESPACE
<BLANKLINE>
[[[0. 0. 0. 0.]
[2. 2. 2. 2.]
[4. 4. 4. 4.]]
<BLANKLINE>
[[1. 1. 1. 1.]
[3. 3. 3. 3.]
[5. 5. 5. 5.]]]
<NDArray 2x3x4 @cpu(0)>
>>> import mxnet.symbol as sym
>>> seq3, _, _, _ = format_sequence(3, sym.Variable("s", shape=(2, 3, 4)), "NTC", False)
>>> seq3
[<Symbol split0>, <Symbol split0>, <Symbol split0>]
>>> seq4 = [nd.array([[0] * 4, [1] * 4]), nd.array([[2] * 4, [3] * 4]), nd.array([[4] * 4, [5] * 4])]
>>> seq5, _, _, _ = format_sequence(3, seq4, "NTC", True)
>>> seq5 # doctest: +NORMALIZE_WHITESPACE
<BLANKLINE>
[[[0. 0. 0. 0.]
[2. 2. 2. 2.]
[4. 4. 4. 4.]]
<BLANKLINE>
[[1. 1. 1. 1.]
[3. 3. 3. 3.]
[5. 5. 5. 5.]]]
<NDArray 2x3x4 @cpu(0)>
>>> seq6 = [sym.Variable("1", shape=(2, 4)), sym.Variable("2", shape=(2, 4)), sym.Variable("3", shape=(2, 4))]
>>> seq7, _, _, _ = format_sequence(3, seq6, "NTC", True)
>>> seq7
<Symbol stack0>
"""
assert inputs is not None, \
"unroll(inputs=None) has been deprecated. " \
"Please create input variables outside unroll."
axis = layout.find('T')
batch_axis = layout.find('N')
batch_size = 0
in_axis = in_layout.find('T') if in_layout is not None else axis
if isinstance(inputs, symbol.Symbol):
F = symbol
if merge is False:
assert len(inputs.list_outputs()) == 1, \
"unroll doesn't allow grouped symbol as input. " \
"Please convert " \
"to list with list(inputs) first or " \
"let unroll handle splitting."
inputs = list(
symbol.split(inputs,
axis=in_axis,
num_outputs=length,
squeeze_axis=1))
elif isinstance(inputs, ndarray.NDArray):
F = ndarray
batch_size = inputs.shape[batch_axis]
if merge is False:
assert length is None or length == inputs.shape[in_axis]
inputs = as_list(
ndarray.split(inputs,
axis=in_axis,
num_outputs=inputs.shape[in_axis],
squeeze_axis=1))
else:
assert length is None or len(inputs) == length
if isinstance(inputs[0], symbol.Symbol):
F = symbol
else:
F = ndarray
batch_size = inputs[0].shape[batch_axis]
if merge is True:
inputs = F.stack(*inputs, axis=axis)
in_axis = axis
if isinstance(inputs,
tensor_types) and axis != in_axis: # pragma: no cover
# todo: find the test case
inputs = F.swapaxes(inputs, dim1=axis, dim2=in_axis)
return inputs, axis, F, batch_size