finn
finn copied to clipboard
Xilinx
Exception: Found multiple get_by_name matches, undefined behavior
Greetings,
I'm currently trying to compile a convolutional model generated via Brevitas and structured as follows:
When coming to the streamline step, I get the following Exception:
---------------------------------------------------------------------------
Exception Traceback (most recent call last)
<ipython-input-7-c3ce74697ba8> in <module>
8 # model = model.transform(MoveScalarLinearPastInvariants())
9 # streamline
---> 10 model = model.transform(Streamline())
11
12 chkpt_name = build_dir + model_name + streamline + ".onnx"
/home/jacopo/git/finn/deps/qonnx/src/qonnx/core/modelwrapper.py in transform(self, transformation, make_deepcopy, cleanup)
138 model_was_changed = True
139 while model_was_changed:
--> 140 (transformed_model, model_was_changed) = transformation.apply(transformed_model)
141 if cleanup:
142 transformed_model.cleanup()
/home/jacopo/git/finn/src/finn/transformation/streamline/__init__.py in apply(self, model)
99 model = model.transform(RemoveIdentityOps())
100 model = model.transform(GiveUniqueNodeNames())
--> 101 model = model.transform(GiveReadableTensorNames())
102 model = model.transform(InferDataTypes())
103 return (model, False)
/home/jacopo/git/finn/deps/qonnx/src/qonnx/core/modelwrapper.py in transform(self, transformation, make_deepcopy, cleanup)
138 model_was_changed = True
139 while model_was_changed:
--> 140 (transformed_model, model_was_changed) = transformation.apply(transformed_model)
141 if cleanup:
142 transformed_model.cleanup()
/home/jacopo/git/finn/deps/qonnx/src/qonnx/transformation/general.py in apply(self, model)
128 out_num = 0
129 for o in n.output:
--> 130 model.rename_tensor(o, "%s_out%d" % (n.name, out_num))
131 out_num += 1
132 init_in_num = 0
/home/jacopo/git/finn/deps/qonnx/src/qonnx/core/modelwrapper.py in rename_tensor(self, old_name, new_name)
302 util.get_by_name(graph.initializer, old_name).name = new_name
303 # sweep over quantization annotations
--> 304 if util.get_by_name(graph.quantization_annotation, old_name, "tensor_name") is not None:
305 util.get_by_name(graph.quantization_annotation, old_name, "tensor_name").tensor_name = new_name
306 # sweep over node i/o
/home/jacopo/git/finn/deps/qonnx/src/qonnx/util/basic.py in get_by_name(container, name, name_field)
80 inds = [i for i, e in enumerate(names) if e == name]
81 if len(inds) > 1:
---> 82 raise Exception("Found multiple get_by_name matches, undefined behavior")
83 elif len(inds) == 0:
84 return None
Exception: Found multiple get_by_name matches, undefined behavior
FINN branch: main (v0.8.1) OS: Windows 10 (WSL2 Ubuntu 18.04) Python: 3.8
The model is a custom-made model made with Brevitas and exported in the ONNX format.
EDIT: I'm also copying the original Brevitas model I used.
import brevitas.nn as qnn
import torch.nn as nn
from brevitas.quant import (
Int8ActPerTensorFixedPoint as ActQuant,
Int8WeightPerTensorFixedPoint as WeightQuant,
)
class ExtendedActQuant(ActQuant):
bit_width = 16
class ExtendedWeightQuant(WeightQuant):
bit_width = 16
class QSPTModel(nn.Module):
def __init__(self, in_channels: int, image_size: int, convolution_sizes: list, dense_sizes: list, output_size: int) -> None:
super(QSPTModel, self).__init__()
self.identity = qnn.QuantIdentity(
act_quant=ExtendedActQuant,
return_quant_tensor=True
)
self.layers = nn.Sequential()
in_size = in_channels
for idx, size in enumerate(convolution_sizes, 1):
self.layers.add_module(
name=f"conv_{idx}",
module=qnn.QuantConv2d(
in_size,
size,
kernel_size=(3, 3),
padding_type="same",
weight_quant=ExtendedWeightQuant,
return_quant_tensor=True
)
)
self.layers.add_module(
name=f"conv_{idx}_relu",
module=qnn.QuantReLU(
act_quant=ExtendedActQuant,
return_quant_tensor=True
)
)
self.layers.add_module(
name=f"maxpool_{idx}",
module=qnn.QuantMaxPool2d(
(2, 2),
return_quant_tensor=True
)
)
in_size = size
self.layers.add_module(name="flatten", module=nn.Flatten())
# the convolution layers squeeze the image of a factor 2
# after len(convolution_sizes) layers, the final size will be
# convolution_sizes[-1]*((image_size / (2**len(convolution_sizes)))**2)
maxpool_final_size = image_size // (2**len(convolution_sizes))
in_size = convolution_sizes[-1]*(maxpool_final_size*maxpool_final_size)
for idx, size in enumerate(dense_sizes, 1):
self.layers.add_module(
name=f"linear_{idx}",
module=qnn.QuantLinear(
in_size,
size,
bias=True,
weight_quant=ExtendedWeightQuant,
return_quant_tensor=True
)
)
self.layers.add_module(
name=f"linear_{idx}_relu",
module=qnn.QuantReLU(
act_quant=ExtendedActQuant,
return_quant_tensor=True
)
)
in_size = size
self.layers.add_module(
name=f"linear_output",
module=qnn.QuantLinear(
in_size,
output_size,
bias=True,
weight_quant=ExtendedWeightQuant,
return_quant_tensor=False
)
)
def forward(self, x):
x = self.identity(x)
x = self.layers(x)
return x
Quick update: oddly enough, but I really cannot give an explanation to this, now the Streamline call does not cause Exception and works fine. I'll leave the issue open in case something new comes up.
I tried "hacking" the workflow a bit by copy-pasting the original Streamline class source code and removing the line of code which causes the exception as follows:
def local_streamline(model):
from qonnx.transformation.base import Transformation
from qonnx.transformation.batchnorm_to_affine import BatchNormToAffine
from qonnx.transformation.general import (
ConvertDivToMul,
ConvertSubToAdd,
GiveReadableTensorNames,
GiveUniqueNodeNames,
)
from qonnx.transformation.infer_datatypes import InferDataTypes
from qonnx.transformation.remove import RemoveIdentityOps
from finn.transformation.streamline.absorb import (
Absorb1BitMulIntoConv,
Absorb1BitMulIntoMatMul,
AbsorbAddIntoMultiThreshold,
AbsorbMulIntoMultiThreshold,
AbsorbSignBiasIntoMultiThreshold,
FactorOutMulSignMagnitude,
)
from finn.transformation.streamline.collapse_repeated import (
CollapseRepeatedAdd,
CollapseRepeatedMul,
)
from finn.transformation.streamline.reorder import (
MoveAddPastConv,
MoveAddPastMul,
MoveMulPastMaxPool,
MoveScalarAddPastMatMul,
MoveScalarLinearPastInvariants,
MoveScalarMulPastConv,
MoveScalarMulPastMatMul,
)
from finn.transformation.streamline.round_thresholds import RoundAndClipThresholds
from finn.transformation.streamline.sign_to_thres import ConvertSignToThres
streamline_transformations = [
ConvertSubToAdd(),
ConvertDivToMul(),
BatchNormToAffine(),
ConvertSignToThres(),
MoveMulPastMaxPool(),
MoveScalarLinearPastInvariants(),
AbsorbSignBiasIntoMultiThreshold(),
MoveAddPastMul(),
MoveScalarAddPastMatMul(),
MoveAddPastConv(),
MoveScalarMulPastMatMul(),
MoveScalarMulPastConv(),
MoveAddPastMul(),
CollapseRepeatedAdd(),
CollapseRepeatedMul(),
MoveMulPastMaxPool(),
AbsorbAddIntoMultiThreshold(),
FactorOutMulSignMagnitude(),
AbsorbMulIntoMultiThreshold(),
Absorb1BitMulIntoMatMul(),
Absorb1BitMulIntoConv(),
RoundAndClipThresholds(),
]
for trn in streamline_transformations:
model = model.transform(trn)
model = model.transform(RemoveIdentityOps())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(InferDataTypes())
model = model.transform(GiveReadableTensorNames())
return model
This fixes the problem, but still no clue why calling GiveReadableTensorNames() causes that exception.