Module furiosa.quantizer.frontend.onnx.transformer.deprecated.fuse_reshape_into_initializer
Expand source code
import onnx
from onnx.helper import make_tensor
from onnx import numpy_helper
from furiosa_sdk_quantizer.interfaces.transformer import Transformer
from furiosa_sdk_quantizer.frontend.onnx.transformer import utils
from furiosa_sdk_quantizer.frontend.onnx.utils.check_model import check_model
class FuseReshapeIntoInitializer(Transformer):
def transform(self, model: onnx.ModelProto) -> onnx.ModelProto:
nodes_by_output_name = {node.output[0]: node for node in model.graph.node}
initializer = {init.name: init for init in model.graph.initializer}
initializer_key = initializer.keys()
# assume Conv is followed by Mul(Conv --> Mul) & Mul takes one data input and one init input
# a * (x * w + b) = (x * aw + ab)
optimized_nodes = []
removed_nodes = []
for node in model.graph.node:
for node_input in node.input:
try:
prev_node = nodes_by_output_name[node_input]
except KeyError:
optimized_nodes.append(node)
continue
if prev_node.op_type != 'Reshape':
optimized_nodes.append(node)
continue
if prev_node.input[0] not in initializer_key:
optimized_nodes.append(node)
continue
init = initializer[prev_node.input[0]]
shape_init = initializer[prev_node.input[1]]
init_arr = numpy_helper.to_array(initializer[prev_node.input[0]])
shape_init_arr = numpy_helper.to_array(shape_init)
reshaped_init_arr = init_arr.reshape(shape_init_arr)
model.graph.initializer.append(
make_tensor(name=node_input, data_type=init.data_type,
dims=shape_init_arr, vals=reshaped_init_arr.flatten())
)
model.graph.initializer.remove(init)
model.graph.initializer.remove(shape_init)
removed_nodes.append(prev_node)
seen = []
for op_node in optimized_nodes:
if op_node in seen:
continue
seen.append(op_node)
optimized_nodes = seen
new_nodes = list(filter(lambda node: node not in removed_nodes, optimized_nodes))
model = utils.rebuild_model(model, new_nodes)
check_model(model)
return modelClasses
class FuseReshapeIntoInitializer (*args, **kwds)-
Abstract base class for generic types.
A generic type is typically declared by inheriting from this class parameterized with one or more type variables. For example, a generic mapping type might be defined as::
class Mapping(Generic[KT, VT]): def getitem(self, key: KT) -> VT: … # Etc.
This class can then be used as follows::
def lookup_name(mapping: Mapping[KT, VT], key: KT, default: VT) -> VT: try: return mapping[key] except KeyError: return default
Expand source code
class FuseReshapeIntoInitializer(Transformer): def transform(self, model: onnx.ModelProto) -> onnx.ModelProto: nodes_by_output_name = {node.output[0]: node for node in model.graph.node} initializer = {init.name: init for init in model.graph.initializer} initializer_key = initializer.keys() # assume Conv is followed by Mul(Conv --> Mul) & Mul takes one data input and one init input # a * (x * w + b) = (x * aw + ab) optimized_nodes = [] removed_nodes = [] for node in model.graph.node: for node_input in node.input: try: prev_node = nodes_by_output_name[node_input] except KeyError: optimized_nodes.append(node) continue if prev_node.op_type != 'Reshape': optimized_nodes.append(node) continue if prev_node.input[0] not in initializer_key: optimized_nodes.append(node) continue init = initializer[prev_node.input[0]] shape_init = initializer[prev_node.input[1]] init_arr = numpy_helper.to_array(initializer[prev_node.input[0]]) shape_init_arr = numpy_helper.to_array(shape_init) reshaped_init_arr = init_arr.reshape(shape_init_arr) model.graph.initializer.append( make_tensor(name=node_input, data_type=init.data_type, dims=shape_init_arr, vals=reshaped_init_arr.flatten()) ) model.graph.initializer.remove(init) model.graph.initializer.remove(shape_init) removed_nodes.append(prev_node) seen = [] for op_node in optimized_nodes: if op_node in seen: continue seen.append(op_node) optimized_nodes = seen new_nodes = list(filter(lambda node: node not in removed_nodes, optimized_nodes)) model = utils.rebuild_model(model, new_nodes) check_model(model) return modelAncestors
- furiosa_sdk_quantizer.interfaces.transformer.Transformer
- typing.Generic
Methods
def transform(self, model: onnx.onnx_ml_pb2.ModelProto) ‑> onnx.onnx_ml_pb2.ModelProto-
Expand source code
def transform(self, model: onnx.ModelProto) -> onnx.ModelProto: nodes_by_output_name = {node.output[0]: node for node in model.graph.node} initializer = {init.name: init for init in model.graph.initializer} initializer_key = initializer.keys() # assume Conv is followed by Mul(Conv --> Mul) & Mul takes one data input and one init input # a * (x * w + b) = (x * aw + ab) optimized_nodes = [] removed_nodes = [] for node in model.graph.node: for node_input in node.input: try: prev_node = nodes_by_output_name[node_input] except KeyError: optimized_nodes.append(node) continue if prev_node.op_type != 'Reshape': optimized_nodes.append(node) continue if prev_node.input[0] not in initializer_key: optimized_nodes.append(node) continue init = initializer[prev_node.input[0]] shape_init = initializer[prev_node.input[1]] init_arr = numpy_helper.to_array(initializer[prev_node.input[0]]) shape_init_arr = numpy_helper.to_array(shape_init) reshaped_init_arr = init_arr.reshape(shape_init_arr) model.graph.initializer.append( make_tensor(name=node_input, data_type=init.data_type, dims=shape_init_arr, vals=reshaped_init_arr.flatten()) ) model.graph.initializer.remove(init) model.graph.initializer.remove(shape_init) removed_nodes.append(prev_node) seen = [] for op_node in optimized_nodes: if op_node in seen: continue seen.append(op_node) optimized_nodes = seen new_nodes = list(filter(lambda node: node not in removed_nodes, optimized_nodes)) model = utils.rebuild_model(model, new_nodes) check_model(model) return model