def convert(self, logical_plan: LogicalPlan) -> None:
nodes_to_skip = set()
while True: # repeat until the logical_graph converges
input_nodes = logical_plan.logical_graph.get_nodes_by_type("_inputs")
# _PseudoOperation(type_name="_inputs"))
root_node = None
for node in input_nodes:
if node in nodes_to_skip:
continue
root_node = node
break
if root_node == None:
break # end of convert
else:
nodes_to_dedup = []
for node in input_nodes:
if node in nodes_to_skip:
continue
if self._check_deduplicate_by_node(root_node, node):
nodes_to_dedup.append(node)
assert(len(nodes_to_dedup) >= 1)
if len(nodes_to_dedup) == 1:
assert(nodes_to_dedup[0] == root_node)
nodes_to_skip.add(root_node)
else:
dedup_node = DedupInputNode(logical_plan.logical_graph, uid(), nodes_to_dedup)._register()
for edge in logical_plan.logical_graph.edges:
if edge.head in nodes_to_dedup:
edge.head = dedup_node
if edge.tail in nodes_to_dedup:
edge.tail = dedup_node
for node in nodes_to_dedup:
node.remove()
def extract_mutation_from_pt_module(
pytorch_model: nn.Module) -> Tuple[Model, Optional[List[Mutator]]]:
model = Model(_internal=True)
graph = Graph(model, uid(), '_model', _internal=True)._register()
model.python_class = pytorch_model.__class__
if len(inspect.signature(model.python_class.__init__).parameters) > 1:
if not is_model_wrapped(pytorch_model):
raise ValueError(
'Please annotate the model with @model_wrapper decorator in python execution mode '
'if your model has init parameters.')
model.python_init_params = cast(dict, pytorch_model.trace_kwargs)
else:
model.python_init_params = {}
# hyper-parameter choice
namespace: ModelNamespace = cast(ModelNamespace,
pytorch_model._model_namespace)
for param_spec in namespace.parameter_specs:
assert param_spec.categorical and param_spec.type == 'choice'
node = graph.add_node(f'param_spec_{param_spec.name}',
'ModelParameterChoice',
{'candidates': param_spec.values})
node.label = param_spec.name
for name, module in pytorch_model.named_modules():
# tricky case: value choice that serves as parameters are stored in traced arguments
if is_basic_unit(module):
trace_kwargs = cast(Dict[str, Any], module.trace_kwargs)
for key, value in trace_kwargs.items():
if isinstance(value, ValueChoiceX):
for i, choice in enumerate(value.inner_choices()):
node = graph.add_node(
f'{name}.init.{key}.{i}', 'ValueChoice',
{'candidates': choice.candidates})
node.label = choice.label
if isinstance(module, (LayerChoice, InputChoice, ValueChoice)):
# TODO: check the label of module and warn if it's auto-generated
pass
if isinstance(module, LayerChoice):
node = graph.add_node(name, 'LayerChoice',
{'candidates': module.names})
node.label = module.label
if isinstance(module, InputChoice):
node = graph.add_node(name, 'InputChoice', {
'n_candidates': module.n_candidates,
'n_chosen': module.n_chosen
})
node.label = module.label
if isinstance(module, ValueChoiceX):
for i, choice in enumerate(module.inner_choices()):
node = graph.add_node(f'{name}.{i}', 'ValueChoice',
{'candidates': choice.candidates})
node.label = choice.label
if isinstance(module, NasBench101Cell):
node = graph.add_node(name, 'NasBench101Cell',
{'max_num_edges': module.max_num_edges})
node.label = module.label
if isinstance(module, Placeholder):
raise NotImplementedError(
'Placeholder is not supported in python execution mode.')
model.status = ModelStatus.Frozen
if not graph.hidden_nodes:
return model, None
mutators = []
mutators_final = []
for nodes in _group_by_label_and_type(graph.hidden_nodes):
label = nodes[0].label
assert label is not None, f'label of {nodes[0]} can not be None.'
assert _is_all_equal(map(lambda n: n.operation.type, nodes)), \
f'Node with label "{label}" does not all have the same type.'
assert _is_all_equal(map(lambda n: n.operation.parameters, nodes)), \
f'Node with label "{label}" does not agree on parameters.'
if nodes[0].operation.type == 'NasBench101Cell':
# The mutation of Nas-bench-101 is special, and has to be done lastly.
mutators_final.append(NasBench101Mutator(label))
else:
mutators.append(ManyChooseManyMutator(label))
return model, mutators + mutators_final
def assemble(self, multi_model_placement: Dict[Model, Device]) \
-> Tuple[Model, Dict[Node, Device]]:
"""
Given a set of models to be formed in a physical model and their device placement,
this function replaces all the logical node in this LogicalPlan with executable physical nodes
for the physical model.
Parameters
----------
multi_model_placement : dict
a dict of models and device placement.
These models will be assembled into the same physical model to run.
Returns
-------
phy_model : Model
the physical model formed by models in `multi_model_placement`
all logical node are replaced by physical nodes
node_placements : dict
the device placement of the nodes in `phy_model`
"""
phy_model = Model(_internal=True)
phy_graph = self.lp_model.root_graph._fork_to(phy_model)
phy_graph._rename_graph(phy_graph.name, "_model")
# merge sub-graphs
for model in multi_model_placement:
if phy_model.evaluator is None and model.evaluator is not None:
phy_model.evaluator = model.evaluator
for graph_name in model.graphs:
if graph_name != model._root_graph_name:
new_graph = model.graphs[graph_name]._fork_to(
phy_model, name_prefix=f'M_{model.model_id}_')
# prefix of M_ of hidden_nodes name in non-root graphs is added here
for new_node in new_graph.hidden_nodes:
if isinstance(new_node.operation, Cell):
old_cell_name = new_node.operation.cell_name
new_node.operation = copy.deepcopy(
new_node.operation)
new_node.operation.cell_name = f'M_{model.model_id}_{old_cell_name}'
assert (phy_model.evaluator is not None)
# When replace logical nodes, merge the training configs when
# input/output nodes are replaced.
evaluator_slot = {} # Model ID -> Slot ID
input_slot_mapping = {}
output_slot_mapping = {}
# Replace all logical nodes to executable physical nodes
hidden_nodes = phy_graph.hidden_nodes.copy()
node_placements = {}
added_models = []
for node in hidden_nodes:
if isinstance(node, OriginNode):
model_id = node.original_graph.model.model_id
if node.original_graph.model not in multi_model_placement:
for edge in node.incoming_edges:
edge.remove()
for edge in node.outgoing_edges:
edge.remove()
node.remove()
continue
if isinstance(node, AbstractLogicalNode):
new_node, placement = node.assemble(multi_model_placement)
if isinstance(new_node.operation, _IOPseudoOperation):
model_id = new_node.graph.model.model_id
if model_id not in evaluator_slot:
added_models.append(model_id)
evaluator_slot[model_id] = len(added_models) - 1
slot = evaluator_slot[model_id]
else:
slot = evaluator_slot[model_id]
# If a model's inputs/outputs are not used in the multi-model
# the codegen and trainer should not generate and use them
# "use_input" and "use_output" are used to mark whether
# an input/output of a model is used in a multi-model
if new_node.operation.type == '_inputs':
input_slot_mapping[new_node] = slot
if new_node.operation.type == '_outputs':
output_slot_mapping[new_node] = slot
self.node_replace(node, new_node)
# name prefix of M_ of cells in hidden_nodes of root graphs is added here
# FIXME: merge this rename with non-root graph, only do once.
if isinstance(new_node.operation, Cell):
old_cell_name = new_node.operation.cell_name
new_node.operation = copy.deepcopy(new_node.operation)
new_node.operation.cell_name = f'M_{model_id}_{old_cell_name}'
# input should be at CPU, move it to GPU first if necessary
if isinstance(new_node.operation, _IOPseudoOperation
) and new_node.operation.type == '_inputs':
# hack: only support single_server
node_placements[new_node] = CPUDevice(
node_id=placement.node_id)
else:
node_placements[new_node] = placement
node.remove()
# If two nodes are placed on different devices, use ToDevice op to copy the node
# TODO: when copying one node to multiple devices, broadcast is more efficient than P2P communication
existing_edges = phy_graph.edges.copy()
# Avoid a node is copied multiple times on the same device
copied_op: Dict[Tuple(Node, Device), Node] = {}
for edge in existing_edges:
head_placement = node_placements[edge.head]
tail_placement = node_placements[edge.tail]
if head_placement != tail_placement:
if head_placement.node_id != tail_placement.node_id:
raise ValueError(
'Cross-server placement is not supported.')
# Same server different devices
if (edge.head, tail_placement) in copied_op:
to_node = copied_op[(edge.head, tail_placement)]
else:
dst_name = edge.head.name + "_to_" + edge.tail.name
to_operation = Operation.new(
'ToDevice', {
"device": tail_placement,
"src": (edge.head.name, edge.head_slot),
"dst": dst_name
})
to_node = Node(phy_graph, uid(), dst_name,
to_operation)._register()
Edge((edge.head, edge.head_slot), (to_node, None),
_internal=True)._register()
copied_op[(edge.head, tail_placement)] = to_node
node_placements[to_node] = head_placement
edge.head = to_node
edge.head_slot = None
# merge all input nodes into one with multiple slots
input_nodes = []
for node in phy_graph.hidden_nodes:
if isinstance(
node.operation,
_IOPseudoOperation) and node.operation.type == '_inputs':
input_nodes.append(node)
for edge in phy_graph.edges:
if edge.head in input_nodes:
edge.head_slot = input_slot_mapping[edge.head]
edge.head = phy_graph.input_node
# merge all output nodes into one with multiple slots
output_nodes = []
for node in phy_graph.hidden_nodes:
if isinstance(
node.operation,
_IOPseudoOperation) and node.operation.type == '_outputs':
output_nodes.append(node)
for edge in phy_graph.edges:
if edge.tail in output_nodes:
edge.tail_slot = output_slot_mapping[edge.tail]
edge.tail = phy_graph.output_node
for node in input_nodes:
node.remove()
for node in output_nodes:
node.remove()
return phy_model, node_placements
def assemble(self, multi_model_placement: Dict[Model, PhysicalDevice]) \
-> Tuple[Model, Dict[Node, PhysicalDevice], List[Model]]:
phy_model = Model(_internal=True) # self.lp_model.fork()
phy_graph = self.lp_model.root_graph._fork_to(phy_model)
# Add a flag to mark multi-model in graph json.
# Multi-model has a list of training configs in kwargs['model_kwargs']
if len(multi_model_placement) > 1:
phy_model.training_config.kwargs['is_multi_model'] = True
phy_model.training_config.kwargs['model_cls'] = phy_graph.name
phy_model.training_config.kwargs['model_kwargs'] = []
# FIXME: allow user to specify
phy_model.training_config.module = 'nni.retiarii.trainer.PyTorchMultiModelTrainer'
# merge sub-graphs
for model in multi_model_placement:
for graph_name in model.graphs:
if graph_name != model._root_graph_name:
model.graphs[graph_name]._fork_to(
phy_model, name_prefix=f'M_{model.model_id}_')
# When replace logical nodes, merge the training configs when
# input/output nodes are replaced.
training_config_slot = {} # Model ID -> Slot ID
input_slot_mapping = {}
output_slot_mapping = {}
# Replace all logical nodes to executable physical nodes
hidden_nodes = phy_graph.hidden_nodes.copy()
node_placements = {}
for node in hidden_nodes:
if isinstance(node, OriginNode):
model_id = node.original_graph.model.model_id
if node.original_graph.model not in multi_model_placement:
for edge in node.incoming_edges:
edge.remove()
for edge in node.outgoing_edges:
edge.remove()
node.remove()
continue
if isinstance(node, AbstractLogicalNode):
new_node, placement = node.assemble(multi_model_placement)
if isinstance(new_node.operation, _IOPseudoOperation):
model_id = new_node.graph.model.model_id
if model_id not in training_config_slot:
phy_model.training_config.kwargs['model_kwargs'].append(
new_node.graph.model.training_config.kwargs.copy())
training_config_slot[model_id] = len(
phy_model.training_config.kwargs['model_kwargs']
) - 1
slot = training_config_slot[model_id]
phy_model.training_config.kwargs['model_kwargs'][slot][
'model_id'] = model_id
phy_model.training_config.kwargs['model_kwargs'][slot][
'use_input'] = False
phy_model.training_config.kwargs['model_kwargs'][slot][
'use_output'] = False
else:
slot = training_config_slot[model_id]
# If a model's inputs/outputs are not used in the multi-model
# the codegen and trainer should not generate and use them
# "use_input" and "use_output" are used to mark whether
# an input/output of a model is used in a multi-model
if new_node.operation.type == '_inputs':
input_slot_mapping[new_node] = slot
phy_model.training_config.kwargs['model_kwargs'][slot][
'use_input'] = True
if new_node.operation.type == '_outputs':
output_slot_mapping[new_node] = slot
phy_model.training_config.kwargs['model_kwargs'][slot][
'use_output'] = True
self.node_replace(node, new_node)
if isinstance(new_node.operation, Cell):
old_cell_name = new_node.operation.cell_name
new_node.operation = copy.deepcopy(new_node.operation)
new_node.operation.cell_name = f'M_{model_id}_{old_cell_name}'
node_placements[new_node] = placement
node.remove()
# If two nodes are placed on different devices, use ToDevice op to copy the node
existing_edges = phy_graph.edges.copy()
# Avoid a node is copied multiple times on the same device
copied_op: Dict[Tuple(Node, PhysicalDevice), Node] = {}
for edge in existing_edges:
head_placement = node_placements[edge.head]
tail_placement = node_placements[edge.tail]
if head_placement != tail_placement:
if head_placement.server != tail_placement.server:
raise ValueError(
'Cross-server placement is not supported.')
# Same server different devices
if (edge.head, tail_placement) in copied_op:
to_node = copied_op[(edge.head, tail_placement)]
else:
to_operation = Operation.new(
'ToDevice', {"device": tail_placement.device})
to_node = Node(phy_graph, uid(),
edge.head.name + "_to_" + edge.tail.name,
to_operation)._register()
Edge((edge.head, edge.head_slot), (to_node, None),
_internal=True)._register()
copied_op[(edge.head, tail_placement)] = to_node
edge.head = to_node
edge.head_slot = None
# merge all input nodes into one with multiple slots
input_nodes = []
for node in phy_graph.hidden_nodes:
if isinstance(
node.operation,
_IOPseudoOperation) and node.operation.type == '_inputs':
input_nodes.append(node)
for edge in phy_graph.edges:
if edge.head in input_nodes:
edge.head_slot = input_slot_mapping[edge.head]
edge.head = phy_graph.input_node
# merge all output nodes into one with multiple slots
output_nodes = []
for node in phy_graph.hidden_nodes:
if isinstance(
node.operation,
_IOPseudoOperation) and node.operation.type == '_outputs':
output_nodes.append(node)
for edge in phy_graph.edges:
if edge.tail in output_nodes:
edge.tail_slot = output_slot_mapping[edge.tail]
edge.tail = phy_graph.output_node
for node in input_nodes:
node.remove()
for node in output_nodes:
node.remove()
return phy_model, node_placements
def extract_mutation_from_pt_module(
pytorch_model: nn.Module) -> Tuple[Model, Optional[List[Mutator]]]:
model = Model(_internal=True)
graph = Graph(model, uid(), '_model', _internal=True)._register()
model.python_class = pytorch_model.__class__
if len(inspect.signature(model.python_class.__init__).parameters) > 1:
if not getattr(pytorch_model, '_nni_model_wrapper', False):
raise ValueError(
'Please annotate the model with @model_wrapper decorator in python execution mode '
'if your model has init parameters.')
model.python_init_params = pytorch_model.trace_kwargs
else:
model.python_init_params = {}
for name, module in pytorch_model.named_modules():
# tricky case: value choice that serves as parameters are stored in traced arguments
if is_basic_unit(module):
for key, value in module.trace_kwargs.items():
if isinstance(value, ValueChoice):
node = graph.add_node(name + '.init.' + key, 'ValueChoice',
{'candidates': value.candidates})
node.label = value.label
if isinstance(module, (LayerChoice, InputChoice, ValueChoice)):
# TODO: check the label of module and warn if it's auto-generated
pass
if isinstance(module, LayerChoice):
node = graph.add_node(name, 'LayerChoice',
{'candidates': module.names})
node.label = module.label
if isinstance(module, InputChoice):
node = graph.add_node(name, 'InputChoice', {
'n_candidates': module.n_candidates,
'n_chosen': module.n_chosen
})
node.label = module.label
if isinstance(module, ValueChoice):
node = graph.add_node(name, 'ValueChoice',
{'candidates': module.candidates})
node.label = module.label
if isinstance(module, Repeat) and module.min_depth <= module.max_depth:
node = graph.add_node(name, 'Repeat', {
'candidates':
list(range(module.min_depth, module.max_depth + 1))
})
node.label = module.label
if isinstance(module, NasBench101Cell):
node = graph.add_node(name, 'NasBench101Cell',
{'max_num_edges': module.max_num_edges})
node.label = module.label
if isinstance(module, Placeholder):
raise NotImplementedError(
'Placeholder is not supported in python execution mode.')
model.status = ModelStatus.Frozen
if not graph.hidden_nodes:
return model, None
mutators = []
mutators_final = []
for nodes in _group_by_label_and_type(graph.hidden_nodes):
assert _is_all_equal(map(lambda n: n.operation.type, nodes)), \
f'Node with label "{nodes[0].label}" does not all have the same type.'
assert _is_all_equal(map(lambda n: n.operation.parameters, nodes)), \
f'Node with label "{nodes[0].label}" does not agree on parameters.'
if nodes[0].operation.type == 'NasBench101Cell':
mutators_final.append(NasBench101Mutator(nodes[0].label))
else:
mutators.append(ManyChooseManyMutator(nodes[0].label))
return model, mutators + mutators_final
