def compress_weights(model: Graph):
"""Apply transformations to save model weights to INT8."""
add_removed_converts(model)
CompressQuantizeWeights().find_and_replace_pattern(model)
model.clean_up()
ForceStrictPrecision().find_and_replace_pattern(model)
model.clean_up()
def build_graph(graph_attrs, meta_data, nodes, edges):
""" Build the Graph with specific nodes and edges.
:param graph_attrs: dictionary with graph attributes
:param nodes: list of nodes where each node is tuple (node_name, type, attrs)
nodes=[
('input', 'Parameter', {}),
('weights', 'Const', {}),
('conv', 'Convolution', {}),
('output', 'Result', {})
]
:param edges: list of edges where each edge is tuple (node_out, node_in, attrs)
edges=[
('input', 'conv', {'out': 0, 'in': 0}),
('weights', 'conv', {'out': 0, 'in': 1}),
('conv', 'output', {'out': 0, 'in': 0})
]
:return: generated graph.
"""
graph = Graph()
graph.graph = graph_attrs
graph.meta_data = meta_data
for node in nodes:
create_node(graph, node[0], node[1], node[2])
for edge in edges:
out_port = edge[2].get('out', 0)
in_port = edge[2].get('in', 0)
connect_nodes_by_name(graph, edge[0], out_port, edge[1], in_port)
graph.clean_up()
return graph
def grouped_convolutions_fusing(graph: Graph):
while True:
is_fused = False
graph.clean_up()
for node in graph.pseudo_topological_sort():
if node.kind == 'op' and len(node.out_nodes()) > 1:
if node.soft_get('can_be_fused') == False:
continue
is_valid_convolutions = True
last_layer = None
next_nodes = get_next_operation(node)
# Check that all operation after this one are Convolutions
# and all convolutions has same output
if len(next_nodes) > 1 and all(_node.soft_get('type') in ['Convolution', 'Deconvolution'] for _node in next_nodes):
for conv in next_nodes:
conv_outputs = get_next_operation(conv)
if conv.soft_get('can_be_fused') == False:
is_valid_convolutions = False
if len(conv_outputs) != 1:
is_valid_convolutions = False
if last_layer is None:
last_layer = conv_outputs[0].id
# TODO: this check is not working for V10 where Biases appears as separate operations
elif conv_outputs[0].id != last_layer:
is_valid_convolutions = False
if is_valid_convolutions:
is_fused = concat_convolutions(graph, node, Node(graph, last_layer))
if is_fused:
break
if not is_fused:
break
def find_and_replace_pattern(self, graph: Graph):
for permute_node in graph.get_op_nodes(type='Transpose'):
if permute_node.id not in graph.nodes():
continue
list_of_permutes = [permute_node]
# Get sequence of permutations
node = permute_node
while True:
next_ops = get_next_operation(node)
if len(next_ops) != 1:
break
next_op = next_ops[0]
if next_op.soft_get('type') == 'Transpose':
list_of_permutes.append(next_op)
node = next_op
else:
break
final_permutation = int64_array([
x for x in range(
len(list_of_permutes[0].in_port(1).data.get_value()))
])
for permute in list_of_permutes:
order = permute.in_port(1).data.get_value()
if order is None:
raise Error(
"Transpose node {} has wrong order for permute = None".
format(permute.name))
final_permutation = final_permutation[int64_array(order)]
if np.array_equal(final_permutation, [
x for x in range(
len(list_of_permutes[0].in_port(1).data.get_value()))
]):
first_data_node, last_data_node = list_of_permutes[0].in_node(
), list_of_permutes[-1].out_node()
graph.remove_edge(first_data_node.id, list_of_permutes[0].id)
else:
if len(list_of_permutes) < 2:
continue
first_data_node, last_data_node = list_of_permutes[0].out_node(
), list_of_permutes[-1].out_node()
list_of_permutes[0].in_port(1).data.set_value(
final_permutation)
graph.remove_edge(first_data_node.id,
first_data_node.out_node().id)
graph.remove_edge(last_data_node.in_node().id, last_data_node.id)
merge_data_nodes(graph, first_data_node, last_data_node)
graph.remove_node(last_data_node.id)
graph.clean_up()
def remove_converts(graph: Graph):
for op in graph.get_op_nodes(type='Convert'):
source_op = op.in_port(0).get_source().node
if source_op.type == 'Const' and source_op.data_type == np.float16:
# Get access to data node after Convert operation and set Insert_Convert_operation_after
# to restore Convert operation later
op.out_node(0)['Insert_Convert_operation_after'] = True
# Mark Const and Convert operation to fold them
source_op['need_shape_inference'] = True
op['stop_value_propagation'] = False
op['need_shape_inference'] = True
graph.clean_up()
def find_and_replace_pattern(self, graph: Graph):
cleanup_called_once = False
# walk through all Loop nodes and find Const inputs
for loop_node in graph.get_op_nodes(op='Loop'):
# call clean-up only once that performs constant folding
if not cleanup_called_once:
graph.clean_up()
cleanup_called_once = True
# move constant node into the body graph and removes body parameter nodes corresponding to them
Loop.pull_constant_inputs_into_body(loop_node)
# since some input ports can be removed after the pulling constants, normalization of Loop node is required
Loop.normalize_input_output_ports(loop_node)
# perform shape inference for the Loop node again since new constant can be appeared
# and constant folding can be helpful for weights path to Convolution node inside the body graph
loop_node['need_shape_inference'] = True
def copy_graph_with_ops(graph: Graph) -> Graph:
"""
Function to copy graph and apply extenders to appropriate nodes
:param graph: Graph to copy
:return:Copied graph with applied extenders
"""
new_graph = Graph()
new_graph.stage = 'back'
new_graph.graph = graph.graph
node_connections = dict()
mapping_of_old_idx_into_new = dict()
restore_correct_ports(graph)
# Nodes preprocessing stage in source graph
# Firstly propagate values only for Const nodes, because other preprocessings
# assumes Const nodes are already preprocessed.
for op in graph.get_op_nodes(type='Const'):
preprocessing_op_nodes[op.type](op)
for op in graph.get_op_nodes():
if op.soft_get('type') != 'Const' and op.soft_get(
'type') in preprocessing_op_nodes:
preprocessing_op_nodes[op.type](op)
# Create a new copy of graph with correct attributes (shape & type infer, backend attrs etc.)
for op in graph.get_op_nodes():
# Save input shapes restored from IR
op['old_input_shapes'] = list()
for n in op.in_nodes():
op.old_input_shapes.append(int64_array(op.in_node(n).shape))
# Apply extenders to nodes in source graph
if op.type in Extender.registered_ops:
Extender.get_extender_class_by_name(op.type).extend(op)
else:
log.debug(
'Extender for node {} with type={} not found, please note.'.
format(op.name, op.type))
# Add node with necessary type and extended attrs in new graph
op_type = op.soft_get('type_to_create', op.type)
if op_type in custom_ops:
node = custom_ops[op_type](new_graph, op.attrs()).create_node()
else:
if op_type not in Op.registered_ops:
log.warning(
'Operation {} is not found in MO operations, please check it! '
'Simple shape infer function is used'.format(op_type))
node = Op(new_graph, op.attrs()).create_node()
assert 'type' in node, 'Operation {} have no `type` attribute.'.format(
node.soft_get('name'))
node['op'] = node.type
node['infer'] = Extender.use_shapes_from_ir
if 'ir_data_attrs' in op:
node['IE'] = [('layer', [
('id', lambda node: node.node), 'name', 'type',
'version'
], [('data', list(op.ir_data_attrs.keys()), []), '@ports',
'@consts'])]
else:
node = Op.get_op_class_by_name(op_type)(
new_graph, op.attrs()).create_node()
# Fill out_ports_count attribute
if 'out_ports_count' not in node and node.soft_get(
'type') != 'Result':
node['out_ports_count'] = len(op.out_edges())
# This attribute is no longer needed and we can delete it
if 'ir_data_attrs' in node:
del node['ir_data_attrs']
if op.has_and_set('need_copy_input_blobs'):
copy_input_blobs(op, node)
# Collect node connections
mapping_of_old_idx_into_new[op.id] = node.id
node_connections[op.id] = collect_node_outputs(op)
# Restore connections in new graph
for input_node_idx, its_outputs in list(node_connections.items()):
for out_port_idx, out_port_dest in its_outputs.items():
for dest_in_port_idx, dest_node_idx in out_port_dest:
src = Node(new_graph,
mapping_of_old_idx_into_new[input_node_idx])
dst = Node(new_graph,
mapping_of_old_idx_into_new[dest_node_idx])
src.out_port(out_port_idx).connect(
dst.in_port(dest_in_port_idx))
# Nodes postprocessing stage in new graph
for op in new_graph.get_op_nodes():
# Call normalize node outputs for restored operations to connect temporary Result operations for disconnected
# output ports. We need to do that for correct shape inference. These Result operations will be removed during
# IR emitting. For TopK operation outputs normalizing we should use specific
# function TopKNormalizer.normalize_outputs.
if op.soft_get('type') != 'TopK':
Op.normalize_outputs(op)
# Set correct_data_type attribute to Const data nodes to correct processing of restored values
if op.soft_get('type') == 'Const':
assert len(op.out_nodes()) == 1 and op.out_node(0).soft_get('kind') == 'data',\
'Const node {} not properly corrected to appropriate data node'.format(op.soft_get('name'))
op.out_node(0)['correct_data_type'] = True
if op.has_and_set('rt_info'):
op.out_node(0)['rt_info'] = op.rt_info
# operations postprocessing with some special types
if op.soft_get('type') in postprocessing_op_nodes:
postprocessing_op_nodes[op.type](op)
restore_tensor_names(op)
# clean up graph to shape inference
new_graph.clean_up()
return new_graph
