fastmachinelearning
`Resize` node empty `RoI` input cause issue for shape inference
Prerequisites
- [x] Test that the bug appears on the current version of the main branch. Make sure to include the commit hash of the commit you checked out.
- [x] Check that the issue hasn't already been reported, by checking the currently open issues.
- [x] If there are steps to reproduce the problem, make sure to write them down below.
- [x] If relevant, please include the ONNX files, which were created directly before and/or after the bug.
Quick summary
For models containing Resize node of version > 10, if RoI input is not used and left empty check_all_tensor_shapes_specified(...) (https://github.com/fastmachinelearning/qonnx/blob/279f9c3c915063f1d1a3b5cdd8f22d6578dbf0d0/src/qonnx/core/modelwrapper.py#L493) return False.
Steps to Reproduce
- Clone the qonnx repository
- Checkout the main branch, with commit hash: 279f9c3c915063f1d1a3b5cdd8f22d6578dbf0d0
- Run the following code:
import onnx
import numpy as np
from onnx import helper, TensorProto
from qonnx.core.modelwrapper import ModelWrapper
import qonnx.core.onnx_exec as oxe
# Define the input tensor (e.g., a 4D tensor with shape [N, C, H, W])
input_tensor = helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 3, 64, 64])
# Define the output tensor shape (e.g., resizing to [1, 3, 128, 128])
output_tensor = helper.make_tensor_value_info("output", TensorProto.FLOAT, [1, 3, 128, 128])
# Define the scale tensor (for upscaling by a factor of 2 along height and width)
scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32)
scales_initializer = helper.make_tensor("scales", TensorProto.FLOAT, [4], scales)
# Create the Resize node
resize_node = helper.make_node(
"Resize",
inputs=["input", "", "scales"], # Empty string for "roi" input as it's not used here
outputs=["output"],
mode="nearest"
)
# Create the graph with input, scales, and output
graph = helper.make_graph(
nodes=[resize_node],
name="ResizeGraph",
inputs=[input_tensor],
outputs=[output_tensor],
initializer=[scales_initializer]
)
# Define the model
model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 13)])
# Save the model
onnx.save(model, "resize_model.onnx")
qonnx_model = ModelWrapper('resize_model.onnx')
# Calling InferShapes() does not solve the issue
qonnx_model = qonnx_model.transform(InferShapes())
ishape = tuple(qonnx_model.get_tensor_shape(qonnx_model.graph.input[0].name))
X = np.random.uniform(low=0, high=1, size=np.prod(ishape)).reshape(ishape).astype(np.float32)
idict = {qonnx_model.graph.input[0].name: X}
y_qonnx = oxe.execute_onnx(qonnx_model, idict)[qonnx_model.graph.output[0].name]
Expected behavior
The empty input should not be considered since when RoI is empty, scales or other inputs should be considered to computed the output shape of the node.
Actual behavior
raise Exception("Found unspecified tensor shapes, try infer_shapes") Exception: Found unspecified tensor shapes, try infer_shapes
from https://github.com/fastmachinelearning/qonnx/blob/279f9c3c915063f1d1a3b5cdd8f22d6578dbf0d0/src/qonnx/core/modelwrapper.py#L493
Possible fix
I fixed the issue by creating the following transformation (note, this is just to show how I fixed the issue, I don't think code should be added in the repository)
class FillEmptyRoI(Transformation):
"Fill empty RoI input tensor of Resize node if is empty to avoid issues during shape inference"
def apply(self, model):
graph_modified = False
for i, node in enumerate(model.graph.node):
if node.op_type == 'Resize':
# Assuming 'roi' is the second input
if len(node.input) > 2 and node.input[1] == '':
roi = onnx.numpy_helper.from_array(np.empty([0], dtype=np.float32), node.name + "_roi")
model.graph.initializer.append(roi)
roi_value_info = helper.make_tensor_value_info(node.name + "_roi", onnx.TensorProto.FLOAT, [0])
model.graph.value_info.append(roi_value_info)
inputs = [node.input[0], node.name + "_roi", node.input[2]]
mode_string = ''
for attr in model.graph.node[i].attribute:
if attr.name == 'mode':
mode_string = attr.s
new_node = onnx.helper.make_node(
"Resize",
coordinate_transformation_mode="asymmetric",
cubic_coeff_a=-0.75,
mode=mode_string,
nearest_mode="floor",
inputs=inputs,
outputs=node.output
)
model.graph.node.remove(node)
model.graph.node.insert(i, new_node)
graph_modified = True
return (model, graph_modified)
The following code works as expected
import onnx
import numpy as np
from onnx import helper, TensorProto
from qonnx.core.modelwrapper import ModelWrapper
import qonnx.core.onnx_exec as oxe
from qonnx.transformation.base import Transformation
from qonnx.transformation.infer_shapes import InferShapes
class FillEmptyRoI(Transformation):
"Fill empty RoI input tensor of Resize node if is empty to avoid issues during shape inference"
def apply(self, model):
graph_modified = False
for i, node in enumerate(model.graph.node):
if node.op_type == 'Resize':
# Assuming 'roi' is the second input
if len(node.input) > 2 and node.input[1] == '':
roi = onnx.numpy_helper.from_array(np.empty([0], dtype=np.float32), node.name + "_roi")
model.graph.initializer.append(roi)
roi_value_info = helper.make_tensor_value_info(node.name + "_roi", onnx.TensorProto.FLOAT, [0])
model.graph.value_info.append(roi_value_info)
inputs = [node.input[0], node.name + "_roi", node.input[2]]
mode_string = ''
for attr in model.graph.node[i].attribute:
if attr.name == 'mode':
mode_string = attr.s
new_node = onnx.helper.make_node(
"Resize",
coordinate_transformation_mode="asymmetric",
cubic_coeff_a=-0.75,
mode=mode_string,
nearest_mode="floor",
inputs=inputs,
outputs=node.output
)
model.graph.node.remove(node)
model.graph.node.insert(i, new_node)
graph_modified = True
return (model, graph_modified)
# Define the input tensor (e.g., a 4D tensor with shape [N, C, H, W])
input_tensor = helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 3, 64, 64])
# Define the output tensor shape (e.g., resizing to [1, 3, 128, 128])
output_tensor = helper.make_tensor_value_info("output", TensorProto.FLOAT, [1, 3, 128, 128])
# Define the scale tensor (for upscaling by a factor of 2 along height and width)
scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32)
scales_initializer = helper.make_tensor("scales", TensorProto.FLOAT, [4], scales)
# Create the Resize node
resize_node = helper.make_node(
"Resize",
inputs=["input", "", "scales"], # Empty string for "roi" input as it's not used here
outputs=["output"],
mode="nearest"
)
# Create the graph with input, scales, and output
graph = helper.make_graph(
nodes=[resize_node],
name="ResizeGraph",
inputs=[input_tensor],
outputs=[output_tensor],
initializer=[scales_initializer]
)
# Define the model
model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 13)])
# Save the model
onnx.save(model, "resize_model.onnx")
qonnx_model = ModelWrapper('resize_model.onnx')
# Calling InferShapes() does not solve the issue
# qonnx_model = qonnx_model.transform(InferShapes())
qonnx_model = qonnx_model.transform(FillEmptyRoI())
ishape = tuple(qonnx_model.get_tensor_shape(qonnx_model.graph.input[0].name))
X = np.random.uniform(low=0, high=1, size=np.prod(ishape)).reshape(ishape).astype(np.float32)
idict = {qonnx_model.graph.input[0].name: X}
y_qonnx = oxe.execute_onnx(qonnx_model, idict)[qonnx_model.graph.output[0].name]
Note: if you want to ingest in hls4ml, I found useful to revert the model to the previous state, i.e. with RoI input set to ''.
I used this optimiser to implement this behaviour:
class EmptyFilledRoI(Transformation):
"Remove RoI tensor of Resize node added for shape inference"
def apply(self, model):
graph_modified = False
for node in model.graph.node:
if node.op_type == 'Resize':
# Assuming 'roi' is the second input
if len(node.input) > 2 and node.input[1] != '':
init_names = [x.name for x in model.graph.initializer]
i = init_names.index(node.input[1])
init_to_remove = model.graph.initializer[i]
model.graph.initializer.remove(init_to_remove)
node.input[1] = ''
graph_modified = True
return (model, graph_modified)
Thanks for flagging this @nghielme - I think we can actually generalize this. It comes from a mix of 1) ONNX node inputs are an ordered list, 2) some nodes can have multiple optional inputs. When working with operators with multiple optional inputs, say optional inputs A, B and C, if one wants to specify a non-default value for B but use the default for A, the ONNX standard uses an empty string as the input name for A.
In QONNX we want to have all tensor shapes specified when working with cleaned-up graphs, which raises an error for this case because there is no tensor with empty string as the name - that empty string is a special placeholder indicating the default input value should be used for that particular input.
I suspect there may be a quick solution - simply adding an exception to check_all_tensor_shapes_specified to ignore any tensors with name = empty string. Once this exception is added, we can bring in your testcase above to check that this no longer fails.
I merged something that might fix for this as part of #125 , please give it a try and let me know.