def shp_apply(self, inputs):
info = self.info
batch_size, channels, height, width = cgt.shape(inputs[0])
pooled_height = cgt.ceil_divide(height + 2*info.pad_h - info.kernel_h, info.stride_h)
pooled_width = cgt.ceil_divide(width + 2*info.pad_w - info.kernel_w, info.stride_w)
outshape = [batch_size , channels, pooled_height, pooled_width]
return outshape
def _subtensor2(x, slis, y):
dims2drop = []
for (ax,sli) in enumerate(slis):
if _is_int_scalar(sli):
if y is None:
dims2drop.append(ax)
else:
yshape = cgt.shape(y)
yshape.insert(ax, 1)
y = y.reshape(yshape)
sli = slice(sli, sli + 1, 1)
assert isinstance(sli.step, int) or sli.step is None
step = 1 if sli.step is None else sli.step
if step < 0:
start = size(x, ax)-1 if sli.start is None else sli.start
stop = -1 if sli.stop is None else sli.stop
else:
start = 0 if sli.start is None else sli.start
stop = size(x, ax) if sli.stop is None else sli.stop
assert isinstance(step, (int, core.Node)), "step argument of a slice should be an integer or a symbolic variable"
if y is None:
x = core.Result(core.GetSli(ax), [x, start, stop, step])
else:
# note: we only support incrementing slice along one axis
return core.Result(core.IncSli(ax), [x, start, stop, step, y])
x = _dropdims(x, dims2drop)
return x
def _subtensor2(x, slis, y):
dims2drop = []
for (ax,sli) in enumerate(slis):
if _is_int_scalar(sli):
if y is None:
dims2drop.append(ax)
else:
yshape = cgt.shape(y)
yshape.insert(ax, 1)
y = y.reshape(yshape)
sli = slice(sli, sli + 1, 1)
assert isinstance(sli.step, int) or sli.step is None
step = 1 if sli.step is None else sli.step
if step < 0:
start = size(x, ax)-1 if sli.start is None else sli.start
stop = -1 if sli.stop is None else sli.stop
else:
start = 0 if sli.start is None else sli.start
stop = size(x, ax) if sli.stop is None else sli.stop
assert isinstance(step, (int, core.Node)), "step argument of a slice should be an integer or a symbolic variable"
if y is None:
x = core.Result(core.GetSli(ax), [x, start, stop, step])
else:
# note: we only support incrementing slice along one axis
return core.Result(core.IncSli(ax), [x, start, stop, step, y])
x = _dropdims(x, dims2drop)
return x
def dropout(x, p=0):
if p == 0:
return x
else:
mask = cgt.greater(cgt.rand(*cgt.shape(x)), p)
x = x * mask
x = x / (1.0 - p)
return x
def dropout(x, p=0):
if p==0:
return x
else:
mask = cgt.greater(cgt.rand(*cgt.shape(x)), p)
x = x * mask
x = x /(1.0-p)
return x
def shp_apply(self, inputs):
info = self.info
batch_size, channels, height, width = cgt.shape(inputs[0])
pooled_height = cgt.ceil_divide(
height + 2 * info.pad_h - info.kernel_h, info.stride_h)
pooled_width = cgt.ceil_divide(width + 2 * info.pad_w - info.kernel_w,
info.stride_w)
outshape = [batch_size, channels, pooled_height, pooled_width]
return outshape
def shp_apply(self, inputs):
# pooled_height_ = static_cast<int>(ceil(static_cast<float>(height_ + 2 * pad_h_ - kernel_h_) / stride_h_)) + 1;
# pooled_width_ = static_cast<int>(ceil(static_cast<float>(width_ + 2 * pad_w_ - kernel_w_) / stride_w_)) + 1;
info = self.info
batch_size, channels, height, width = cgt.shape(inputs[0])
pooled_height = cgt.ceil_divide(height + 2*info.pad_h - info.kernel_h, info.stride_h)
pooled_width = cgt.ceil_divide(width + 2*info.pad_w - info.kernel_w, info.stride_w)
outshape = [batch_size , channels, pooled_height, pooled_width]
return (outshape, outshape)
def pullback(self, inputs, output, goutput):
if self.pullback_impl is None:
raise core.MethodNotDefined
pb_input_types = self.input_types + [self.output_type]*2
pb_output_type = core.TupleType(*self.input_types)
pbop = EasyCustomOp(pb_input_types, pb_output_type,
forward_impl=self.pullback_impl, pullback_impl=None,
shapefun = lambda *args : tuple(cgt.shape(x) for x in inputs) )
return cgt.core.unpack(core.Result(pbop, inputs + [output, goutput]))
def sample(self, p, shape=None, numeric=False):
""" Element-wise sampling for each component of p """
# TODO_TZ maybe cgt has mechanism to eval an expr
if not numeric:
p = core.as_node(p)
shape = shape or cgt.shape(p)
return cgt.rand(*shape) <= p
else:
assert isinstance(p, np.ndarray)
return np.array(nr.rand(*p.shape) <= p, dtype="i2")
def shp_apply(self, inputs):
info = self.info
batch_size, channels, height, width = cgt.shape(inputs[0])
height_out = (height + 2 * info.pad_h -
info.kernel_h) // info.stride_h + 1
width_out = (width + 2 * info.pad_w -
info.kernel_w) // info.stride_w + 1
return [
batch_size, height_out, width_out,
channels * info.kernel_w * info.kernel_h
]
def shp_apply(self, inputs):
# pooled_height_ = static_cast<int>(ceil(static_cast<float>(height_ + 2 * pad_h_ - kernel_h_) / stride_h_)) + 1;
# pooled_width_ = static_cast<int>(ceil(static_cast<float>(width_ + 2 * pad_w_ - kernel_w_) / stride_w_)) + 1;
info = self.info
batch_size, channels, height, width = cgt.shape(inputs[0])
pooled_height = cgt.ceil_divide(
height + 2 * info.pad_h - info.kernel_h, info.stride_h)
pooled_width = cgt.ceil_divide(width + 2 * info.pad_w - info.kernel_w,
info.stride_w)
outshape = [batch_size, channels, pooled_height, pooled_width]
return (outshape, outshape)
class Im2Col(core.Op):
available_impls = ("native_cpu", )
def __init__(self, info):
assert info.stride_h > 0 and info.stride_w > 0
self.info = info
def get_diff(self, _):
return [True]
def get_py_impl(self):
raise core.MethodNotDefined
def pullback(self, (x, ), _y, gy):
return [core.Result(Col2Im(self.info), [gy] + cgt.shape(x))]
def determine_memowner(nodes_sorted, updates, node2dev, outputs):
# First determine how many "child" nodes each node has
node2child = defaultdict(list)
for node in nodes_sorted:
for parent in node.parents:
node2child[parent].append(node)
# Now traverse graph again and see where we can use the same memory
node2memowner = {} # mapping node x -> the node that owns its memory
# For updates, memlocation(RHS) = memlocation(LHS)
after2before = {after:before for (before,after) in updates}
enable_inplace_opt = core.get_config()["enable_inplace_opt"]
for node in nodes_sorted:
base = node # by default,
if node.is_argument():
pass
elif node.op.writes_to_input >= 0:
base = node2memowner[node.parents[node.op.writes_to_input]]
elif node in after2before:
base = after2before[node]
elif enable_inplace_opt and node.op.return_type == "byref": # TODO think about if we need any other conditions
nodeshape = node.op.shp_apply(node.parents)
for parent in node.parents:
parentowner = node2memowner[parent]
if (len(node2child[parent])==1
and nodeshape==cgt.shape(parent) # XXX not a very robust way to check
and node.dtype == parent.dtype
and _is_data_mutable(parentowner)
and parent not in outputs
):
base = parentowner
break
# TODO: add optimization for in-place incrementing
node2memowner[node] = base
return node2memowner
def shp_apply(self, inputs):
return cgt.shape(inputs[0])
def sample(self, p, shape=None):
p = core.as_node(p)
shape = shape or cgt.shape(p)
return cgt.rand(*shape) <= p
def shp_apply(self, inputs):
info = self.info
batch_size, channels, height, width = cgt.shape(inputs[0])
height_out = (height + 2 * info.pad_h - info.kernel_h) // info.stride_h + 1
width_out = (width + 2 * info.pad_w - info.kernel_w) // info.stride_w + 1
return [batch_size , height_out, width_out, channels * info.kernel_w * info.kernel_h]
def shp_apply(self, parents):
if self.shapefun:
return self.shapefun(parents)
else:
return cgt.shape(self)
def shp_apply(self, inputs):
return cgt.shape(inputs[0])
def sample(self, p, shape=None):
p = core.as_node(p)
shape = shape or cgt.shape(p)
return cgt.rand(*shape) <= p