def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
modified_graph = False
candidates = [node for node in G.nodes()
if len(G.indexed_out_edges(node.name)) == 1 and len(G.out_edges(node.name)) > 1]
while candidates:
node = candidates.pop(0)
strings = self.explore(G, [node])
if not strings:
continue
modified_graph = True
primary = strings.pop(0)
for pnode in primary:
if pnode in candidates:
candidates.remove(pnode)
out_edges = []
for other in strings:
out_edges.extend(G.out_edges(other[-1].name))
for other_node in other:
if other_node in candidates:
candidates.remove(other_node)
G.remove(other_node)
nid = NodeId(other_node)
if G.quantization and nid in G.quantization:
del G.quantization[nid]
LOG.info(
f'removed duplicates from {primary[0].name} {",".join(node.name for node in other)}')
pend = primary[-1]
for edge in out_edges:
G.add_edge(
NNEdge(from_node=pend, to_node=edge.to_node, to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return modified_graph
def match(self, G: GraphView, set_identity: bool = True):
if not G.quantization:
return
for nid in [nid for nid, qrec in G.quantization.sorted_iterator(G) if qrec is None or not (qrec.in_qs and qrec.out_qs)]:
if nid.fnode_name:
LOG.warning("can't add quantization to fused node %s", nid.fnode_name)
continue
if nid.node_name not in G:
# previous fusions may have removed nodes from the graph
continue
node = nid.get_node(G)
predecessors = [NodeId(pred) for pred in G.predecessors(node.name)]
successors = [NodeId(succ) for succs in G.successors(node.name) for succ in succs]
go_back = not successors or (predecessors and all(pred in G.quantization for pred in predecessors))
go_forward = not predecessors or (successors and all(succ in G.quantization for succ in successors))
if not (go_back or go_forward):
LOG.warning("node %s is not connected to anything and has no quantization", node.name)
continue
if go_forward:
out_qrecs = set(G.quantization[nid] for nid in successors)
if not all(isinstance(out_qrec, MultQuantizationRecord) for out_qrec in out_qrecs):
continue
out_qtypes = reduce_qtypes([(edge.from_idx, G.quantization[NodeId(edge.to_node)].in_qs[edge.to_idx])
for edge in G.out_edges(node.name)])
else:
out_qtypes = None
if go_back:
in_qrecs = set(G.quantization[nid] for nid in predecessors)
if not all(isinstance(in_qrec, MultQuantizationRecord) for in_qrec in in_qrecs):
continue
in_qtypes = reduce_qtypes([(edge.to_idx, G.quantization[NodeId(edge.from_node)].out_qs[edge.from_idx])
for edge in G.in_edges(node.name)])
else:
in_qtypes = None
if not in_qtypes:
if not predecessors:
LOG.info("setting quantization on input node %s", node.name)
qrec = MultQuantizationRecord(in_qs=deepcopy(out_qtypes), out_qs=deepcopy(out_qtypes))
else:
raise NotImplementedError("propagating qrecs not implemented")
elif not out_qtypes:
if not successors:
LOG.info("setting quantization on output node %s", node.name)
qrec = MultQuantizationRecord(in_qs=deepcopy(in_qtypes), out_qs=deepcopy(in_qtypes))
else:
raise NotImplementedError("propagating qrecs not implemented")
else:
LOG.info("setting quantization on node %s", node.name)
qrec = MultQuantizationRecord(in_qs=deepcopy(in_qtypes), out_qs=deepcopy(out_qtypes))
G.quantization[nid] = qrec
if set_identity:
self.set_identity(G)
return False
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
const_ops = [node for node in G.nodes()
if isinstance(node, MatrixMulParameters)
and any([isinstance(edge.from_node, ConstantInputParameters)
and check_equals(G, edge.from_node, 1.0/6.0)
for edge in G.in_edges(node.name)])]
oprecs = [oprec for oprec in (look_back(G, op)
for op in const_ops)
if oprec is not None]
has_modified_graph = False
for oprec in oprecs:
mul_edge = G.out_edges(oprec['mul'][0].name)
if len(mul_edge) == 1:
mul_edge = mul_edge[0]
if isinstance(mul_edge.to_node, ReluActivationParameters):
oprec['relu3'] = (mul_edge.to_node,
G.quantization[NodeId(mul_edge.to_node)])
has_modified_graph = True
process_rec(G, oprec)
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
for node in G.nodes(node_classes=tuple(VALID_FUSIONS.keys())):
node_list = self.get_node_list(G, node,
FusionMatch(self._default_ktype))
if node_list is None or len(node_list.order) < 2:
continue
LOG.info("fusing nodes %s", ",".join(
(node.name for node in node_list.order)))
has_modified_graph = True
subgraph = GraphView()
last_node = None
for snode in node_list.order:
if last_node is not None:
subgraph.add_edge(
NNEdge(from_node=last_node, to_node=snode))
last_node = snode
# assumption here is that the first node could have multiple inputs but definitely has only
# one output
input_mapping = [[
(node_list.node, idx)
] for idx in range(G.num_in_edges(node_list.node.name))]
output_mapping = [(last_node, 0)]
pnode = node_list.fusions_class(node_list.node.name + '_fusion',
fusion_type=node_list.fusion_type,
subgraph=subgraph,
input_mapping=input_mapping,
output_mapping=output_mapping)
if G.quantization:
# TODO - stats
qrecs = G.quantization.get_all(pnode.contained_nodes())
if qrecs:
prec = QRec.copy_ktype(qrecs[0],
in_qs=qrecs[0].in_qs,
out_qs=qrecs[-1].out_qs)
for fnode in pnode.contained_nodes():
G.quantization.move_to_fusion(fnode, pnode)
G.quantization[NodeId(pnode)] = prec
in_edges = G.in_edges(node_list.node.name)
out_edges = G.out_edges(last_node.name)
for snode in node_list.order:
G.remove(snode)
for edge in in_edges:
G.add_edge(
NNEdge(edge.from_node,
pnode,
from_idx=edge.from_idx,
to_idx=edge.to_idx))
for edge in out_edges:
G.add_edge(
NNEdge(pnode,
edge.to_node,
from_idx=edge.from_idx,
to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def find_forward(G: GraphView,
edge,
find_node_classes,
skip_node_classes=None,
find_skip=None):
if find_skip is None:
find_skip = [find_node_classes, skip_node_classes]
for idx, elem in enumerate(find_skip):
if elem is not None and not isinstance(elem, tuple):
if isinstance(elem, list):
find_skip[idx] = tuple(elem)
else:
find_skip[idx] = tuple([elem])
if isinstance(edge.to_node, find_skip[0]):
return [[edge]]
if skip_node_classes and isinstance(edge.to_node, find_skip[0]):
res = []
for out_edge in G.out_edges(edge.to_node.name):
edge_lists = find_forward(G,
out_edge,
find_node_classes,
find_skip=find_skip)
if not edge_lists:
continue
res.extend([[edge] + edge_list for edge_list in edge_lists])
return res
return []
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
fragment = GraphMatcher(
match_function=lambda state, frag: (frag, state['match']))
fragment.add_node(MatScaleNodeMatch())
has_modified_graph = False
for frag, match in fragment.match_graph(G):
match_edges = [
G.indexed_in_edges(node.name)[idx]
for node, idx in match['inputs']
]
matched_node = list(frag.nodes())[0]
out_edges = G.out_edges(matched_node.name)
has_modified_graph = True
G.remove(matched_node)
fnode = MatScaleFusionParameters(
"{}_fusion".format(matched_node.name),
fusion_type=match['type'],
subgraph=frag,
input_mapping=[[(matched_node, 0)], [(matched_node, 1)]])
G.add_node(fnode)
for idx, edge in enumerate(match_edges):
edge.to_node = fnode
edge.to_idx = idx
G.add_edge(edge)
for edge in out_edges:
edge.from_node = fnode
G.add_edge(edge)
if set_identity:
self.set_identity(G)
return has_modified_graph
def match(self, G: GraphView, set_identity: bool = True):
if not G.quantization:
return
softmaxes = [
node for node in G.nodes() if isinstance(node, SoftMaxParameters)
]
qrecs = [G.quantization[NodeId(node)] for node in softmaxes]
if not all(isinstance(qrec, MultQuantizationRecord) for qrec in qrecs):
return
for softmax, qrec in zip(softmaxes, qrecs):
in_q = qrec.in_qs[0]
in_q.scale_to_pow2()
for edge in G.in_edges(softmax.name):
propagate_qtype_up(G, in_q, edge)
for edge in G.out_edges(softmax.name):
assert isinstance(
edge.to_node,
(OutputParameters, QuantizeParameters
)), "Softmax is supported only at the end of the graph"
out_qrec = G.quantization[NodeId(edge.to_node)]
out_qrec.in_qs[0] = qrec.out_qs[0]
out_qrec.out_qs[0] = qrec.out_qs[0]
if set_identity:
self.set_identity(G)
return False
def match(self, G: GraphView, set_identity: bool = True):
# get a list of all the nodes that are transposable but not transposes
# Need to do this first to avoid mutating it when doing the modifications
tnodes = list(filter(lambda n: isinstance(n, Transposable) and\
not isinstance(n, TransposeParameters),
G.nodes()))
for node in tnodes:
if node.transpose_in:
for idx, edge in enumerate(G.in_edges(node.name)):
in_params = TransposeParameters("%s_TIN_%s" % (node.name, idx),
transpose=node.transpose_in)
if node.in_dims_hint:
in_hint = node.in_dims_hint[edge.to_idx]
out_hint = apply_reverse_transpose_to_hint(in_hint, node.transpose_in)
in_params.in_dims_hint = [in_hint.copy()]
in_params.out_dims_hint = [out_hint.copy()]
node.in_dims_hint[edge.to_idx] = out_hint
G.insert_node(in_params, edge.from_node.name, edge.to_node.name,
from_idx=edge.from_idx, to_idx=edge.to_idx)
node.transpose_in = None
if node.transpose_out:
for idx, edge in enumerate(G.out_edges(node.name)):
out_params = TransposeParameters("%s_TOUT_%s" % (node.name, idx),
transpose=node.transpose_out)
if node.out_dims_hint:
out_hint = node.out_dims_hint[edge.from_idx]
in_hint = apply_reverse_transpose_to_hint(out_hint, node.transpose_out)
out_params.in_dims_hint = [in_hint.copy()]
out_params.out_dims_hint = [out_hint.copy()]
node.out_dims_hint[edge.from_idx] = in_hint
G.insert_node(out_params, edge.from_node.name, edge.to_node.name,
from_idx=edge.from_idx, to_idx=edge.to_idx)
node.transpose_out = None
if set_identity:
self.set_identity(G)
def _match(self,
G: GraphView,
set_identity: bool = True,
**kwargs) -> bool:
has_modified_graph = False
for node in [
node for node in G.nodes() if self.node_does_nothing(G, node)
]:
has_modified_graph = True
in_edge = G.in_edges(node.name)[0]
G.remove_edge(in_edge)
for out_edge in G.out_edges(node.name):
G.remove_edge(out_edge)
G.add_edge(
NNEdge(in_edge.from_node,
out_edge.to_node,
from_idx=in_edge.from_idx,
to_idx=out_edge.to_idx))
LOG.info(f'removing {node.name} that does nothing')
G.remove(node)
if set_identity:
self.set_identity(G)
return has_modified_graph
def match(self, G: GraphView, set_identity: bool = True):
has_modified = False
for node in G.nodes(node_classes=ConstantInputParameters):
out_edges = G.out_edges(node.name)
if len(out_edges) <= 1:
continue
has_modified = True
LOG.info(
'node %s has more than one out edge and will be duplicated',
node.name)
idx = 1
for out_edge in out_edges[1::]:
new_constant = ConstantInputParameters(f'{node.name}_{idx}',
dims=Dim.unnamed(
node.dims.shape),
value=node.value.copy())
G.remove_edge(out_edge)
G.add_edge(
NNEdge(from_node=new_constant,
to_node=out_edge.to_node,
to_idx=out_edge.to_idx))
idx += 1
if set_identity:
self.set_identity(G)
return has_modified
def match(self, G: GraphView, set_identity: bool = True):
# Only works for reverses connected to one RNN node
reverse_nodes = set([
node for node in G.nodes()
if (isinstance(node, ReverseParameters)
and len(G.out_edges(node.name)) == 1 and isinstance(
G.out_edges(node.name)[0].to_node, RNNBaseParameters))
])
has_modified_graph = False
for reverse_node in reverse_nodes:
in_edges = G.in_edges(reverse_node.name)
rnn_edge = G.out_edges(reverse_node.name)[0]
if rnn_edge.to_idx != 0:
LOG.warning("reverse on rnn input %s", rnn_edge.to_idx)
continue
assert not rnn_edge.to_node.revert, "RNN node is already reversed!"
rnn_edge.to_node.revert = True
LOG.info("fusing reverses into node %s", rnn_edge.to_node.name)
has_modified_graph = True
G.remove(reverse_node)
for edge in in_edges:
G.add_edge(
NNEdge(edge.from_node,
rnn_edge.to_node,
from_idx=edge.from_idx,
to_idx=rnn_edge.to_idx))
for edge in G.out_edges(rnn_edge.to_node.name):
if not isinstance(edge.to_node, ReverseParameters):
continue
if edge.from_idx != 0:
LOG.warning("reverse on rnn output %s", edge.from_idx)
continue
rev_edges = G.out_edges(edge.to_node.name)
G.remove(edge.to_node)
for rev_edge in rev_edges:
G.add_edge(
NNEdge(edge.from_node,
rev_edge.to_node,
from_idx=edge.from_idx,
to_idx=rev_edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
rnn_nodes = [
self.find_unpack(G, node) for node in G.nodes()
if isinstance(node, RNNBaseParameters) and node.n_output_cells > 1
]
rnn_nodes_by_slice = self.validate_slices(G, rnn_nodes)
rnn_nodes_by_slice = self.validate_multi_branch(G, rnn_nodes_by_slice)
if not rnn_nodes_by_slice:
return False
for unpack_node, rnn_unpacks in rnn_nodes_by_slice.items():
modified_nodes = set()
for rnn_unpack in rnn_unpacks:
self.process_path(G, rnn_unpack, modified_nodes)
# since process path will have removed all unnecessary nodes the edges will be correct here
out_edges = G.out_edges(unpack_node.name)
in_edges = G.in_edges(unpack_node.name)
assert len(in_edges
) == 1, "expecting unpack node to have only one in edge"
in_edge = in_edges[0]
changes_shape = unpack_node.changes_shape if isinstance(
unpack_node, StridedSliceParameters) else False
LOG.info("Eliminating last cell unpack: %s", unpack_node.name)
G.remove(unpack_node)
# Here the strided slice can change the output shape of the RNN
# so insert a reshape to do the shape change
if changes_shape:
reshape = ReshapeParameters(
unpack_node.name + '_reshape',
old_shape=Dim.unnamed(unpack_node.post_slice_shape),
shape=Dim.unnamed(unpack_node.out_shape))
G.add_edge(
NNEdge(from_node=in_edge.from_node,
to_node=reshape,
from_idx=in_edge.from_idx))
for out_edge in out_edges:
G.add_edge(
NNEdge(from_node=reshape,
to_node=out_edge.to_node,
to_idx=out_edge.to_idx))
if G.quantization:
G.quantization[NodeId(reshape)] = G.quantization[NodeId(
unpack)]
else:
for out_edge in out_edges:
G.add_edge(
NNEdge(from_node=in_edge.from_node,
to_node=out_edge.to_node,
from_idx=in_edge.from_idx,
to_idx=out_edge.to_idx))
if G.quantization:
del G.quantization[NodeId(unpack_node)]
if set_identity:
self.set_identity(G)
return True
def match(self, G: GraphView, set_identity: bool = True):
has_modified_graph = False
for conv_node in [params for params in G.nodes() if isinstance(params, Conv2DParameters)]:
node_list = self.get_node_list(G, conv_node)
if node_list is None or len(node_list.order) < 2:
continue
if node_list.fusion_type == 'conv_active_pool':
if node_list.pool.pool_type == "average":
node_list.order = node_list.order[:2:]
node_list.pool = None
elif node_list.fusion_type == 'conv_pool_active':
if node_list.pool.pool_type == "average" and node_list.active.activation != "relu":
continue
LOG.info("fusing nodes %s", ",".join((node.name for node in node_list.order)))
has_modified_graph = True
subgraph = GraphView()
last_node = None
for node in node_list.order:
if last_node is not None:
subgraph.add_edge(NNEdge(from_node=last_node, to_node=node))
last_node = node
input_mapping = [[(node_list.conv, idx)] for idx in range(3)]
output_mapping = [(last_node, 0)]
pnode = ConvFusionParameters(
node_list.conv.name + '_fusion',
fusion_type=node_list.fusion_type,
subgraph=subgraph,
in_dims_hint=node_list.conv.in_dims_hint,
out_dims_hint=node_list.conv.out_dims_hint,
input_mapping=input_mapping,
output_mapping=output_mapping)
if G.quantization:
qrecs = G.quantization.get_all(pnode.contained_nodes())
if qrecs:
prec = None
if isinstance(qrecs[0], (SymmetricQuantizationRecord, SymmetricScalableFilterQuantizationRecord)):
prec = SymmetricQuantizationRecord(
in_qs=qrecs[0].in_qs, out_qs=qrecs[-1].out_qs)
elif isinstance(qrecs[0], (MultQuantizationRecord, MultScalableFilterQuantizationRecord)):
prec = MultQuantizationRecord(in_qs=qrecs[0].in_qs, out_qs=qrecs[-1].out_qs)
elif isinstance(qrecs[0], (Float32QuantizationRecord, Float32ScalableFilterQuantizationRecord)):
prec = Float32QuantizationRecord(
in_qs=qrecs[0].in_qs, out_qs=qrecs[-1].out_qs)
for node in pnode.contained_nodes():
G.quantization.move_to_fusion(node, pnode)
G.quantization[NodeId(pnode)] = prec
in_edges = G.in_edges(node_list.conv.name)
out_edges = G.out_edges(last_node.name)
for node in node_list.order:
G.remove(node)
for edge in in_edges:
G.add_edge(NNEdge(edge.from_node, pnode, from_idx=edge.from_idx, to_idx=edge.to_idx))
for edge in out_edges:
G.add_edge(NNEdge(pnode, edge.to_node, from_idx=edge.from_idx, to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
something_changed = False
for relu_node in [node for node in G.nodes(node_classes=ReluActivationParameters) if node.upper_bound == 6]:
out_edges = G.out_edges(relu_node)
if len(out_edges) != 1 or not isinstance(out_edges[0].to_node, MatrixMulParameters):
continue
mul_node = out_edges[0].to_node
in_edges = G.in_edges(mul_node)
if len(in_edges) != 2:
continue
other_edge = (set(in_edges) - {out_edges[0]}).pop()
constant_node = other_edge.from_node
if len(G.out_edges(constant_node)) != 1:
continue
if (not isinstance(constant_node, ConstantInputParameters) or
not check_equals(G, constant_node, 1.0/6.0)):
continue
something_changed = True
activation = HSigmoidActivationParameters(
G.unique_name(f'{mul_node.name}_hsigmoid'), offset=0)
in_edges = G.in_edges(relu_node)
out_edges = G.out_edges(mul_node)
nodes_to_replace = [relu_node, mul_node, constant_node]
LOG.info(f'fusing {", ".join(node.name for node in nodes_to_replace)} into HSIGMOID {activation.name}')
G.remove_all(nodes_to_replace)
for in_edge in in_edges:
G.add_edge(NNEdge.clone(in_edge, to_node=activation, to_idx=0))
for out_edge in out_edges:
G.add_edge(NNEdge.clone(
out_edge, from_node=activation, from_idx=0))
if G.quantization:
reluqrec = G.quantization[NodeId(relu_node)]
mulqrec = G.quantization[NodeId(mul_node)]
del G.quantization[NodeId(constant_node)]
pqrec = QRec.copy_ktype(
reluqrec, in_qs=reluqrec.in_qs, out_qs=mulqrec.out_qs)
G.quantization[NodeId(activation)] = pqrec
return something_changed
def match(self, G: GraphView, set_identity: bool = True):
visited_edges = {}
nodes_to_remove = []
has_modified_graph = False
for node in G.inputs():
# check if constantinput. if is then check if positive and check max value
if isinstance(node, ConstantInputParameters):
if node.value is not None:
if G.has_quantized_parameters:
qrec = G.quantization[NodeId(node)]
qtype = qrec.out_qs[0]
if hasattr(qtype, 'wrapped'):
qtype = qtype.wrapped
val = qtype.dequantize(node.value)
else:
val = node.value
if val.min() >= 0:
status = (True, val.max())
else:
status = (False, False)
else:
status = (False, False)
for edge in G.out_edges(node.name):
visited_edges[edge] = status
nodes_to_remove += find_redundant_relus(
G, edge.to_node, visited_edges)
for node in nodes_to_remove:
has_modified_graph = True
# Only relus so only one in edge
in_edge = G.in_edges(node.name)[0]
for edge in G.out_edges(node.name):
G.add_edge(
NNEdge(from_node=in_edge.from_node,
from_idx=in_edge.from_idx,
to_node=edge.to_node,
to_idx=edge.to_idx))
G.remove(node)
if set_identity:
self.set_identity(G)
return has_modified_graph
def match(self, G: GraphView, set_identity: bool = True):
# get a list of all the nodes that are transposable but not transposes
# Need to do this first to avoid mutating it when doing the modifications
tnodes = list(filter(lambda n: isinstance(n, Transposable) and
not isinstance(n, TransposeParameters),
G.nodes()))
has_modified_graph = False
for node in tnodes:
if node.transpose_in:
for idx, edge in enumerate(G.in_edges(node.name)):
if edge.to_idx >= len(node.transpose_in):
continue
trans = node.transpose_in[edge.to_idx]
if trans is None:
continue
has_modified_graph = True
in_params = TransposeParameters("%s_TIN_%s" % (node.name, idx),
transpose=trans)
if node.in_dims_hint and node.in_dims_hint[edge.to_idx]:
in_hint = node.in_dims_hint[edge.to_idx]
out_hint = apply_reverse_transpose_to_hint(in_hint, trans)
in_params.in_dims_hint = [in_hint.copy()]
in_params.out_dims_hint = [out_hint.copy()]
node.in_dims_hint[edge.to_idx] = out_hint
if G.quantization:
G.quantization.copy_to_node(node, in_params)
G.insert_node(in_params, edge.from_node.name, edge.to_node.name,
from_idx=edge.from_idx, to_idx=edge.to_idx,
edge_class=NNEdge)
node.transpose_in = None
if node.transpose_out:
for idx, edge in enumerate(G.out_edges(node.name)):
if edge.from_idx >= len(node.transpose_out):
continue
trans = node.transpose_out[edge.from_idx]
if trans is None:
continue
has_modified_graph = True
out_params = TransposeParameters("%s_TOUT_%s" % (node.name, idx),
transpose=trans)
if node.out_dims_hint:
out_hint = node.out_dims_hint[edge.from_idx]
in_hint = apply_reverse_transpose_to_hint(out_hint, trans)
out_params.in_dims_hint = [in_hint.copy()]
out_params.out_dims_hint = [out_hint.copy()]
node.out_dims_hint[edge.from_idx] = in_hint
if G.quantization:
G.quantization.copy_to_node(node, out_params)
G.insert_node(out_params, edge.from_node.name, edge.to_node.name,
from_idx=edge.from_idx, to_idx=edge.to_idx,
edge_class=NNEdge)
node.transpose_out = None
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
group_identity = kwargs.get('group_identity')
if group_identity == 'pow2_match_group':
valid_activations = VALID_ACTIVATIONS_POW2
else:
valid_activations = VALID_ACTIVATIONS_SQ8
for fc_node in [params for params in G.nodes() if isinstance(params, FcParameters)]:
node_list = self.get_node_list(G, fc_node, valid_activations)
if node_list is None or len(node_list.order) < 2:
continue
LOG.info("fusing nodes %s", ",".join(
(node.name for node in node_list.order)))
has_modified_graph = True
subgraph = GraphView()
last_node = None
for node in node_list.order:
if last_node is not None:
subgraph.add_edge(
NNEdge(from_node=last_node, to_node=node))
last_node = node
input_mapping = [[(node_list.linear, idx)] for idx in range(3)]
output_mapping = [(last_node, 0)]
pnode = LinearFusionParameters(
node_list.linear.name + '_fusion',
fusion_type=node_list.fusion_type,
subgraph=subgraph,
input_mapping=input_mapping,
output_mapping=output_mapping)
if G.quantization:
# TODO - stats
qrecs = G.quantization.get_all(pnode.contained_nodes())
if qrecs:
prec = QRec.copy_ktype(
qrecs[0], in_qs=qrecs[0].in_qs, out_qs=qrecs[-1].out_qs)
for node in pnode.contained_nodes():
G.quantization.move_to_fusion(node, pnode)
G.quantization[NodeId(pnode)] = prec
in_edges = G.in_edges(node_list.linear.name)
out_edges = G.out_edges(last_node.name)
for node in node_list.order:
G.remove(node)
for edge in in_edges:
G.add_edge(NNEdge(edge.from_node, pnode,
from_idx=edge.from_idx, to_idx=edge.to_idx))
for edge in out_edges:
G.add_edge(NNEdge(pnode, edge.to_node,
from_idx=edge.from_idx, to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
for node in G.nodes(node_classes=SplitParameters):
same_op_edges = self.moveable_same_operation_edges(G, node)
if not same_op_edges:
continue
has_modified_graph = True
in_edges = G.in_edges(node.name)
assert len(in_edges) == 1
# sort by name to ensure that operation is repeatable
same_op_edges.sort(key=lambda x: x.to_node.name)
keep_node = same_op_edges[0].to_node
LOG.info('split node %s has duplicate operations on its out edges',
node.name)
LOG.info('moving %s before split node %s', keep_node.name,
node.name)
for edge in G.out_edges(node.name):
node_out_edges = G.out_edges(edge.to_node.name)
G.remove(edge.to_node)
if edge.to_node != keep_node:
LOG.info('deleting duplicate node %s', edge.to_node.name)
if G.quantization:
nid = NodeId(edge.to_node)
if nid in G.quantization:
del G.quantization[nid]
for out_edge in node_out_edges:
G.add_edge(
NNEdge(from_node=node,
from_idx=edge.from_idx,
to_node=out_edge.to_node,
to_idx=out_edge.to_idx))
G.insert_node_at_edge(keep_node, in_edges[0], edge_class=NNEdge)
if G.quantization:
quantizer = NewQuantizer.from_quantized_graph(G)
quantizer.quantize()
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
visited_edges = {}
nodes_to_remove = []
has_modified_graph = False
for node in G.inputs():
# check if constantinput. if is then check if positive and check max value
if isinstance(node, ConstantInputParameters):
if node.value is not None:
val = node.dqvalue
if np.min(val) >= 0:
status = (True, np.max(val))
else:
status = (False, False)
else:
status = (False, False)
else:
status = (False, False)
for edge in G.out_edges(node.name):
visited_edges[edge] = status
nodes_to_remove += find_redundant_relus(
G, edge.to_node, visited_edges)
for node in nodes_to_remove:
has_modified_graph = True
# Only relus so only one in edge
LOG.info("removing redundant relu %s", node.name)
in_edge = G.in_edges(node.name)[0]
out_edges = G.out_edges(node.name)
G.remove(node)
for edge in out_edges:
G.add_edge(NNEdge(from_node=in_edge.from_node,
from_idx=in_edge.from_idx,
to_node=edge.to_node,
to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self,
G: GraphView,
set_identity: bool = True,
**kwargs) -> bool:
has_modified_graph = False
for split_node in set(
[node for node in G.nodes() if isinstance(node, SplitParameters)]):
in_edges = G.in_edges(split_node.name)
if len(in_edges) > 1:
continue
in_edge = in_edges[0]
if not isinstance(in_edge.from_node, ConcatParameters):
continue
concat_node = in_edge.from_node
if len(G.out_edges(concat_node.name)) > 1:
continue
if concat_node.transpose_out or split_node.transpose_in:
continue
if concat_node.axis != split_node.axis:
continue
axis = concat_node.axis
split_out_sizes = [
out_shape[axis] for out_shape in split_node.out_shapes
]
if len(split_out_sizes) != len(concat_node.in_dims):
continue
if not all(split_out_sizes[idx] == in_dim.shape[axis]
for idx, in_dim in enumerate(concat_node.in_dims)):
continue
has_modified_graph = True
LOG.info("removing unnecessary concat/split pair %s/%s",
concat_node.name, split_node.name)
concat_in_edges = G.indexed_in_edges(concat_node.name)
split_out_edges = G.indexed_out_edges(split_node.name)
G.remove(split_node)
G.remove(concat_node)
for idx, in_edge in enumerate(concat_in_edges):
for out_edge in split_out_edges[idx]:
G.add_edge(
NNEdge(from_node=in_edge.from_node,
from_idx=in_edge.from_idx,
to_node=out_edge.to_node,
to_idx=out_edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self,
G: GraphView,
set_identity: bool = True,
**kwargs) -> bool:
replaced = True
has_modified_graph = False
while replaced:
replaced = False
for subgraph in self.match_function(G):
# TODO - Save in and out edges here since the replace function may modify the
# subgraph
in_edges = [
in_edge for input_node in subgraph.inputs()
for in_edge in G.in_edges(input_node.name)
]
out_edges = [
out_edge for output_node in subgraph.outputs()
for out_edge in G.out_edges(output_node.name)
]
try:
replacement, edge_in_mapping, edge_out_mapping = self.replace_function(
G, subgraph)
if replacement is None:
G.remove_fragment(subgraph)
has_modified_graph = True
elif isinstance(replacement, Node):
# use saved in and out edges
G.replace_fragment(subgraph,
replacement,
frag_in_edges=in_edges,
frag_out_edges=out_edges,
edge_in_mapping=edge_in_mapping,
edge_out_mapping=edge_out_mapping)
has_modified_graph = True
else:
raise TypeError(
"unexcepted return value from replace_function")
replaced = True
break
except DontReplaceError:
pass
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
fac = MatScalePairMatchFactory()
has_modified_graph = False
for frag, match in fac.get_matcher().match_graph(G):
match_edges = [
G.indexed_in_edges(node.name)[idx] for node, idx in match
]
first_node = frag.inputs()[0]
last_node = frag.outputs()[0]
out_edges = G.out_edges(last_node.name)
for node in frag.nodes():
G.remove(node)
input_mapping = MatScaleFusionParameters.get_mapping_from_edges(
match_edges)
fnode = MatScaleFusionParameters(
"{}_{}_fusion".format(first_node.name, last_node.name),
fusion_type="vec_scalar",
subgraph=frag,
input_mapping=MatScaleFusionParameters.convert_input_mapping(
input_mapping))
has_modified_graph = True
G.add_node(fnode)
fnode.in_dims_hint = [None] * 3
for idx, edge in enumerate(match_edges):
new_edge = edge.clone(
to_node=fnode,
to_idx=list(input_mapping[edge.to_node].keys())[0])
if new_edge.from_node.out_dims_hint:
fnode.in_dims_hint[idx] = new_edge.from_node.out_dims_hint[
edge.from_idx]
G.add_edge(new_edge)
for edge in out_edges:
new_edge = edge.clone(from_node=fnode)
G.add_edge(new_edge)
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
modified_graph = True
while modified_graph:
modified_graph = False
for reshape in G.nodes(node_classes=(ReshapeParameters, )):
if not reshape.has_transpose and reshape.shape.shape == reshape.old_shape.shape:
modified_graph = True
LOG.info('removing reshape that does nothing %s',
reshape.name)
G.remove_and_reconnect(reshape, edge_class=NNEdge)
nid = NodeId(reshape)
if G.quantization and nid in G.quantization:
del G.quantization[nid]
res = None
for reshape in G.nodes(node_classes=(ReshapeParameters, )):
res = self.validate_reshape(G, reshape)
if res:
LOG.info('unnecessary reshape found after %s',
reshape.name)
modified_graph = True
(reshape, candidates, out_shape) = res
for candidate in candidates:
LOG.info(
'removing unnecessary reshape or transpose %s',
candidate.name)
edges = G.out_edges(candidate.name)
G.remove(candidate)
nid = NodeId(candidate)
if G.quantization and nid in G.quantization:
del G.quantization[nid]
for edge in edges:
G.add_edge(
NNEdge(from_node=reshape,
to_node=edge.to_node,
to_idx=edge.to_idx))
reshape.shape = Dim.unnamed(out_shape)
break
if set_identity:
self.set_identity(G)
return modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
nodes = list(G.nodes(node_classes=GlobalPoolingParameters))
modified_graph = False
while nodes:
node = nodes.pop()
node_group = self.reductions(G, node)
if len(node_group) <= 1:
continue
modified_graph = True
reduction_axes, new_shape, has_keepdims, _ = reduce(
reduce_reduction, node_group, None)
new_node = node_group[0]
new_node.axis = sorted(list(reduction_axes))
new_node.keep_dims = has_keepdims
out_edges = G.out_edges(node_group[-1].name)
if G.quantization:
last_qrec = G.quantization[NodeId(node_group[-1])]
G.quantization[NodeId(new_node)].out_qs = last_qrec.out_qs
for node in node_group[1::]:
G.remove(node.name)
nid = NodeId(node)
if G.quantization and nid in G.quantization:
del G.quantization[nid]
if has_keepdims and len(new_shape) != len(
new_node.in_dims[0].shape):
rparams = ReshapeParameters(
G.unique_name(f'{new_node.name}_reshape'),
shape=Dim.unnamed(new_shape))
if G.quantization:
G.quantization.copy_qrec(last_qrec, 'out', 0, rparams)
G.add_edge(NNEdge(new_node, rparams))
new_node = rparams
for edge in out_edges:
G.add_edge(NNEdge(new_node, edge.to_node, to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
nodes_to_remove = []
for node in G.nodes(node_classes=CopyParameters):
out_edges = G.out_edges(node)
if len(out_edges) > 1:
continue
if (search_down(
G,
out_edges[0],
(OutputParameters, InputParameters, ConstantInputParameters,
SplitParameters, ConcatParameters),
can_pass=(ReshapeParameters, NoOPParameters),
can_pass_fn=lambda G, node: isinstance(
node, TransposeParameters) and node.does_nothing,
follow_multi=True) and search_up(
G,
G.in_edges(node)[0],
(InputParameters, OutputParameters,
ConstantInputParameters, SplitParameters,
ConcatParameters),
can_pass=(ReshapeParameters, NoOPParameters),
can_pass_fn=lambda G, node: isinstance(
node, TransposeParameters) and node.does_nothing,
follow_multi=True)):
continue
nodes_to_remove.append(node)
for node in nodes_to_remove:
LOG.info("remove redundant copy %s", node.name)
has_modified_graph = True
G.remove_and_reconnect(node, edge_class=NNEdge)
if G.quantization:
nid = NodeId(node)
if nid in G.quantization:
del G.quantization[nid]
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
modified_graph = False
candidates = set(G.nodes(node_classes=(ReshapeParameters, )))
while candidates:
node = candidates.pop()
out_edges = G.out_edges(node.name)
if len(out_edges) != 1 or not isinstance(
out_edges[0].to_node,
FcParameters) or out_edges[0].to_node.batch_size > 1:
continue
LOG.info('removing unnecessary reshape before linear %s',
node.name)
G.remove_and_reconnect(node, edge_class=NNEdge)
modified_graph = True
nid = NodeId(node)
if G.quantization and G.quantization.get(nid):
del G.quantization[nid]
modified_graph = True
if set_identity:
self.set_identity(G)
return modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
group_identity = kwargs.get('group_identity')
if group_identity == 'pow2_match_group':
valid_activations = VALID_ACTIVATIONS_POW2
valid_activations_wo_pool = VALID_ACTIVATIONS_POW2_WO_POOL
else:
valid_activations = VALID_ACTIVATIONS_SQ8
valid_activations_wo_pool = VALID_ACTIVATIONS_SQ8_WO_POOL
for pool_node in G.nodes(node_classes=(PoolingParameters,
GlobalPoolingParameters)):
node_list = self.get_node_list(G, pool_node, valid_activations,
valid_activations_wo_pool)
if node_list is None or len(node_list.order) < 2:
continue
LOG.info("fusing nodes %s", ",".join(
(node.name for node in node_list.order)))
has_modified_graph = True
subgraph = GraphView()
last_node = None
for node in node_list.order:
if last_node is not None:
subgraph.add_edge(NNEdge(from_node=last_node,
to_node=node))
last_node = node
input_mapping = [[(node_list.pool, 0)]]
output_mapping = [(last_node, 0)]
pnode = ActivationFusion(node_list.pool.name + '_fusion',
fusion_type=node_list.fusion_type,
subgraph=subgraph,
input_mapping=input_mapping,
output_mapping=output_mapping)
if G.quantization:
# TODO - stats
qrecs = G.quantization.get_all(pnode.contained_nodes())
if qrecs:
prec = QRec.copy_ktype(qrecs[0],
in_qs=qrecs[0].in_qs,
out_qs=qrecs[-1].out_qs)
for node in pnode.contained_nodes():
G.quantization.move_to_fusion(node, pnode)
if isinstance(node, GlobalPoolingParameters):
# Global pooling fused with activations need to have only the activation scale
G.quantization[NodeId(pnode,
node)].out_qs[0] = deepcopy(
G.quantization[NodeId(
pnode,
node)].in_qs[0])
G.quantization[NodeId(
pnode, node)].out_qs[0].dtype = np.int32
G.quantization[NodeId(pnode)] = prec
in_edges = G.in_edges(node_list.pool.name)
out_edges = G.out_edges(last_node.name)
for node in node_list.order:
G.remove(node)
for edge in in_edges:
G.add_edge(
NNEdge(edge.from_node,
pnode,
from_idx=edge.from_idx,
to_idx=edge.to_idx))
for edge in out_edges:
G.add_edge(
NNEdge(pnode,
edge.to_node,
from_idx=edge.from_idx,
to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def match(self, G: GraphView, set_identity: bool = True):
has_modified_graph = False
for pad_node in [
params for params in G.nodes()
if isinstance(params, PadParameters)
]:
node_list = self.get_node_list(G, pad_node)
if node_list is None or len(node_list.order) < 2:
continue
LOG.info("fusing nodes %s", ",".join(
(node.name for node in node_list.order)))
has_modified_graph = True
subgraph = GraphView()
padded_input_idx = G.out_edges(node_list.pad.name)[0].to_idx
subgraph.add_edge(
NNEdge(from_node=node_list.pad,
to_node=node_list.add,
to_idx=padded_input_idx))
last_node = node_list.add
node_list.add.force_quantized_index = 0
if node_list.active:
subgraph.add_edge(
NNEdge(from_node=node_list.add, to_node=node_list.active))
last_node = node_list.active
if padded_input_idx == 0:
input_mapping = [[(node_list.pad, 0)], [(node_list.add, 1)]]
else:
input_mapping = [[(node_list.add, 0)], [(node_list.pad, 1)]]
output_mapping = [(last_node, 0)]
pnode = PaddedAddFusionParameters(
"PADDED_" + node_list.add.name,
fusion_type=node_list.fusion_type,
subgraph=subgraph,
input_mapping=input_mapping,
output_mapping=output_mapping)
if G.quantization:
qrecs = G.quantization.get_all(pnode.contained_nodes())
if qrecs:
prec = None
if isinstance(qrecs[0],
(SymmetricQuantizationRecord,
SymmetricScalableFilterQuantizationRecord)):
prec = SymmetricQuantizationRecord(
in_qs=qrecs[0].in_qs, out_qs=qrecs[-1].out_qs)
elif isinstance(qrecs[0],
(MultQuantizationRecord,
MultScalableFilterQuantizationRecord)):
prec = MultQuantizationRecord(in_qs=qrecs[0].in_qs,
out_qs=qrecs[-1].out_qs)
elif isinstance(qrecs[0],
(Float32QuantizationRecord,
Float32ScalableFilterQuantizationRecord)):
prec = Float32QuantizationRecord(
in_qs=qrecs[0].in_qs, out_qs=qrecs[-1].out_qs)
for node in pnode.contained_nodes():
G.quantization.move_to_fusion(node, pnode)
G.quantization[NodeId(pnode)] = prec
if padded_input_idx == 0:
in_edges = G.in_edges(node_list.pad.name) + G.indexed_in_edges(
node_list.add.name)[1::]
else:
in_edges = G.indexed_in_edges(
node_list.add.name)[0:1:] + G.in_edges(node_list.pad.name)
out_edges = G.out_edges(last_node.name)
for node in node_list.order:
G.remove(node)
for edge in in_edges:
G.add_edge(
NNEdge(edge.from_node,
pnode,
from_idx=edge.from_idx,
to_idx=edge.to_idx))
for edge in out_edges:
G.add_edge(
NNEdge(pnode,
edge.to_node,
from_idx=edge.from_idx,
to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
filter_nodes = [
node for node in G.nodes() if isinstance(node, FilterParameters)
]
for filter_node in filter_nodes:
while True:
out_edges = G.out_edges(filter_node.name)
# can't fuse if there is a branch
if len(out_edges) > 1:
break
out_edge = out_edges[0]
op_node = out_edge.to_node
# must be a valid matrix op
if not isinstance(op_node, tuple(OPS.keys())):
break
# other edge to the op must be a constant
other_idx = 1 if out_edge.to_idx == 0 else 0
other_in_edge = G.indexed_in_edges(op_node.name)[other_idx]
if not isinstance(other_in_edge.from_node,
ConstantInputParameters):
break
const_node = other_in_edge.from_node
remove_constant = len(G.out_edges(const_node.name))
flat_value = const_node.dqvalue.flatten()
out_c = filter_node.filter.out_c
op, weights_and_biases = OPS[op_node.__class__]
# it would be possible to support mult bias addition by out channel but only supporting a
# scalar at present
if len(flat_value) != 1 and (weights_and_biases
or len(flat_value) != out_c):
LOG.warning('could not absorb %s into %s', const_node.name,
filter_node.name)
break
# If there is quantization then essentially the output of the filter
# takes the quantization of the output of the operation.
# The biases will not change since their quantization depends on the weights
# and input
fnid = NodeId(filter_node)
opnid = NodeId(op_node)
if G.quantization and (fnid in G.quantization
or opnid in G.quantization):
if not (fnid in G.quantization
and opnid in G.quantization):
LOG.warning(
'could not absorb %s into %s - graph is partially quantized',
const_node.name, filter_node.name)
break
fqrec = G.quantization[fnid]
opqrec = G.quantization[opnid]
fqrec.out_qs[0] = opqrec.out_qs[0]
has_modified_graph = True
LOG.info("fusing bias in %s into %s", const_node.name,
filter_node.name)
self.fuse_bias(G, filter_node, other_idx, op, flat_value, 2)
if weights_and_biases:
# TODO - need to adjust weights quantization here
LOG.info("fusing multiplicative bias in %s into %s",
const_node.name, filter_node.name)
self.fuse_bias(G, filter_node, other_idx, op, flat_value,
1)
out_edges = G.out_edges(op_node.name)
G.remove(op_node)
if remove_constant:
G.remove(const_node)
for edge in out_edges:
G.add_edge(
NNEdge(from_node=filter_node,
to_node=edge.to_node,
to_idx=edge.to_idx))
if set_identity:
self.set_identity(G)
return has_modified_graph
def _match(self, G: GraphView, set_identity: bool = True, **kwargs):
has_modified_graph = False
has_transposed = False
for params in G.nodes(node_classes=MatMulOpParameters):
while True:
out_edges = G.out_edges(params.name)
# can't fuse if there is a branch
if len(out_edges) > 1:
break
out_edge = out_edges[0]
op_node = out_edge.to_node
# must be a valid matrix op
if not isinstance(op_node,
(MatrixAddParameters, MatrixMulParameters)):
break
# other edge to the op must be a constant
other_idx = 1 if out_edge.to_idx == 0 else 0
other_in_edge = G.indexed_in_edges(op_node.name)[other_idx]
if not isinstance(other_in_edge.from_node,
ConstantInputParameters):
break
const_node = other_in_edge.from_node
remove_constant = len(G.out_edges(const_node.name))
flat_value = const_node.dqvalue.flatten()
out_shape = params.out_dims[0].shape
if len(out_shape) != 2:
raise ValueError(
f'strange outputs shape of {out_shape} for matmul {params.name}'
)
if len(flat_value) != out_shape[0] and len(
flat_value) != out_shape[1]:
LOG.info(
"can't fuse %s into %s - value shape is not correct for bias",
const_node.name, params.name)
break
has_bias = len(params.in_dims) == 3
if isinstance(op_node, MatrixAddParameters):
if has_bias:
if len(flat_value.shape) != len(params.in_dims[2]):
LOG.info(
"can't fuse %s into %s - bias shape is not the same",
const_node.name, params.name)
break
bias_node = G.indexed_in_edges(
params.name)[2].from_node
LOG.info(
"folding additive bias from %s into existing bias on %s",
op_node.name, params.name)
bias_node.value = bias_node.dq_value + flat_value
else:
if len(flat_value) == out_shape[1]:
# matmul needs to be transposed to fuse this
reverse_matmul(G, params)
has_transposed = True
bias_node = ConstantInputParameters(
G.unique_name(f'{params.name}_bias'),
value=flat_value,
dims=Dim.unnamed(flat_value.shape))
G.add_edge(
NNEdge(from_node=bias_node,
to_node=params,
to_idx=2))
# extend the inward transpose
if params.transpose_in:
params.transpose_in = params.transpose_in + [None]
LOG.info(
"folding additive bias from %s into new bias on %s",
op_node.name, params.name)
else:
params_in = G.indexed_in_edges(params.name)
consts = [
isinstance(edge.from_node, ConstantInputParameters)
for edge in params_in
]
if not any(consts):
break
mult_const_node = params_in[1].from_node if consts[
1] else params_in[0].from_node
mult_const_node.value = mult_const_node.dqvalue * const_node.dqvalue
if has_bias:
bias_node = params_in[2].from_node
bias_node.value = bias_node.dqvalue * const_node.dqvalue
LOG.info(
"folding multaplicative bias from %s into new bias on %s",
op_node.name, params.name)
out_edges = G.out_edges(op_node.name)
G.remove(op_node)
if remove_constant:
G.remove(const_node)
for edge in out_edges:
G.add_edge(
NNEdge(from_node=params,
to_node=edge.to_node,
to_idx=edge.to_idx))
G.add_dimensions()
if G.quantization:
quantizer = UnifiedQuantizer.from_quantized_graph(G)
quantizer.quantize(G, start_nodes=[params])
RemoveUnnecessaryQuantizeOperators().match(G)
if has_transposed:
G.adjust_order()
if set_identity:
self.set_identity(G)
return has_modified_graph
